Want signal? We need more noise (more examples to address the quiet bottleneck)

by Dana Lewis

Note: I’m assuming if you are reading this post that you’ve first read this post, summarizing the concept. The below is a companion piece with some expanded details about the concepts and more examples of how I think addressing this bottleneck will help make a difference.

—

You might hold the perspective that in the growing era of AI, there’s too much noise already. Slopocalypse. (Side note: this entire post, and my other post, was written by me, a human.) That’s not the biggest problem in healthcare, though: in both research and clinical care, so much critical data is simply not collected. We are missing sooooo much important data. Some of this is an artifact of past clinical trial design and how it was hard to collect or analyze or store the data; there were less established norms around data re-use; and some of it is a “collect the bare minimum to study the endpoints because that’s all we are supposed to do” phenomenon.

Nowadays, we should be thinking about how to make data and insights from clinical work and research available to AI, too, because AI will increasingly be used by humans to sort through what is known and where the opportunities are, plus make cross-domain connections that humans have been missing. And we need to be thinking about whether the incentives are set up correctly (spoiler: they’re not) to make sure all available data is able to be collected and shared or at least stored for humans and AI to have access to for future insight. What is studied is also ‘what is able to be funded’, which is disproportionately things that come to the commercial market eventually after regulatory approval.)

“But what’s the point?” you ask. “If no one is looking at the data from this study, why collect it?” Because the way we design studies – to answer a very limited-scope question, i.e. safety and efficacy for labeling claims and regulatory approval – is very different than the studies we need around optimizing and personalizing treatments. If you look at really large populations of people accessing a treatment –  think GLP-1 RA injectables, for example – you eventually start to see follow-on studies around optimization and different titration recommendations. But most diseases and most treatments aren’t for a population even 1/10th the size of people accessing those meds. So those studies don’t get done. Doctors don’t necessarily pay attention to this; patients may or may not report relevant data about this back to the doctor (and again even if they do: doctors don’t mentally or physically store this data, often); and the signal is missed because we don’t capture this ‘noise’.

This means people with rare diseases, undiagnosed diseases, atypical or seronegative diseases, unusual responses to treatments, multiple conditions (comorbidities), or any other scenario that results in them being part of a small population and where customizing and individualizing their care is VERY important…they have no evidence-based guidance. Clinicians do the best they can to interpret in the absence of evidence or guidelines, but no wonder patients often turn to LLMs (AI) for additional input and contextualization and discussion of the tradeoffs and pros and cons of different approaches.

We have to: the best available human (our clinicians) may not have the relevant expertise; or they may not be available at all; or they may be biased (even subconsciously) or forgetful of critical details or not up to date on the latest evidence; or they may be faced with a truly novel situation and they don’t have the skills to address it because they’re used to cookie-cutter standardized cases. This is not a dig at clinicians, but I recognize that we now have tools that can address some of the existing flaws in our human-based healthcare system. I keep talking about how we need to recognize that evaluations of AI in healthcare shouldn’t treat the status quo as the baseline to defend, because the status quo itself has problems, as I just described.

And AI has some of the same problems. (Except for the ‘not available’ part, unless you consider the lower-utility free access models to mean that the more advanced, thinking-based models are ‘not available’ because of cost access barriers). An AI may not have any training data on a rare disease… because nothing exists. It may drop information out of the context window, and we may not realize that this has happened (i.e., it ‘forgets’ something). Usually these are critiques of AI, juxtaposed against the implication that humans are better. But notice that these critiques are the same of humans, too! This happens all the time with human clinicians in healthcare. A human can’t make decisions on data that doesn’t exist in the world, either!

So…how do we “fix” AI? Or, how do we fix human healthcare? We should be asking BOTH questions. Maybe the answer is the same: increase the noise so we increase the signal. Argue about the ratio later (of signal:noise), but increase the amount of everything first. That is most important.

How might we do this? I have been thinking about this a lot, and Astera recently posted an essay contest asking for ideas that don’t fit the current infrastructure. (That’s why I’m finally writing this up, not because I think I’ll win the essay contest, but mostly because i’s an opportunity for people to consider whether/not this type of solution is a good idea as a separate analysis from “well, who is going to fund THAT?”. Let’s discuss and evaluate the idea, or riff on it, without being bogged down by the ‘how exactly it gets funded and managed over time’.)

I think we should create some kind of digitally-managed platform/ecosystem to do the following:

  1. Incentivize and facilitate written (AI-assisted allowed) output of everything. Case reports and scenarios of people and what they’re facing. All the background information that might have contributed. All the data we have access to that is related to the case, plus other passive, easy-to-collect data (that may already be collected, e.g., wearables and phone accelerometer data) even if it does not appear to be related. Patient-perspective narratives, clinical interpretations, clinical data, wearable data – all of this.

    A.) And, be willing to take variable types of data even from the same study population. For example, as a person with type 1 diabetes, I have 10+ years of CGM data. For a future study on exocrine pancreatic insufficiency, for example, not everyone will have CGM data, but because CGM data is increasingly common in people with T2D (a much larger population) and in the general population, a fraction of people will have CGM data and a fraction of people are willing to share it. We should enable this, even if it’s not for a primary endpoint analysis on the EPI study and even if only a fraction of people choose to share it – it might still be useful for subgroup analysis, determining power for a future study where it is part of the protocol, and identifying new research directions!

  2. Host storage somewhere. There can be some automated checking against identifiable information and sanity checking what’s submitted into the repository, plus consent (either self-consent or signed consent forms for this purpose.)
  3. Provide ‘rewards’ as incentives for inputting cases and data.

    A) Rewards might differ for patients self-submitting data and researchers and clinicians inputting data and cases. Maybe it’s one type of reward for patients and a batch reward of a free consult or service of some kind for researchers/clinicians (e.g. every 3 case reports submitted = 1 credit, credits can be used toward a new AI tool or token budget for LLMs or human expertise consult around study design, or all kinds of things.)

    B) Host data challenges where different types of funders incentivize different disease groups or families of conditions to be added, to round out what data is available in different areas.

  4. Enable AI and human access (including to citizen scientists/independent researchers who don’t have institutions) to these datasets, after self-credentialing and providing a documented use case for why the data is being accessed.

    A) Require anyone accessing the dataset to make their analysis code open source, or otherwise openly available for others to use, and to make the results available as well.

    B) Tag anytime an individual dataset is used by a project, that way individuals with self-donated data (or clinicians submitting a case) can revisit and see if there are any research insights or data analysis that applies to their case.

    C) Provide starter projects to show how the data can be used for novel insight generation. For example, develop sandboxes for different types of datasets with existing ‘lab notebooks’, so to speak, to onboard people to different datasets/groups of data, different types of analyses they might do, and to cut down on the environment setup barriers to getting started to analyzing this data.

    D) Facilitate outreach to institutions, disease-area patient nonprofits and advocacy groups both to solicit data inputs AND to use the data for generating outputs.

Why should we do this?

  • Clinical trials do not gather all of the possible useful data, because it’s not for their primary/secondary endpoints. Plus, they would have to clean it upon collection. Plus storage cost. Plus management decisions. Etc. There’s a lot of real barriers there, but the result is that clinical trials do not capture enough or the most possible relevant data. So we need a different path.
  • Clinicians/clinical pathways don’t have ways to review data, so they avoid doing it. There’s no path to “submit this data to my record but I don’t expect you to look at it” in medical charts (but there should be). So we need a different path.
  • Disease groups/organizations sometimes host and capture datasets, but like clinical trials, this is limited data; it often comes through clinical pathways (further limiting participation and diversity of the data); it doesn’t include all the additional data that might be useful. So we need a different path.
  • Some patients do publish or co-author in traditional journals, but it’s a limited, self-selected group. Then there are the usual journal hurdles that filter this down further. Not everyone knows about pre-prints. And, not everyone knows how useful their data can be – either as an n=1 alone or as an n=1*many for it to be worth sharing. A lot of people’s data may be in video/social media formats inaccessible to LLMs (right now) or locked in private Facebook groups or other proprietary platforms. So we need a different path.

Thus, the answer to ‘why we should do this’ is recognizing that: AI does not create magic out of nowhere. It relies on training data and extrapolating from that, plus web search. If it’s not searchable or findable, and it’s not in the training data, it’s not there. Yes, even the newer models and prompting it to extrapolate doesn’t solve all of these problems. What AI can do is shaped disproportionately by formal literature, institutional documents, and whatever scraps of publicly available content remain. If we want to shape what’s possible in the future, we need to start now by shaping and collecting the inputs to make it happen. We can’t fix the past trial designs but we can start to fill in the gaps in the data!

Here is an example of how this might work and why it matters to build and incentivize this type of data sharing.

  1. If we build and incentivize data sharing, the following individuals might self-donate their data and write-ups, or a clinician might submit them. Later, someone might analyze the data and put the pieces together.
  • Someone with type 1 diabetes (an autoimmune condition) and this new-onset muscle-related issue. Glucose levels are well-managed, and they don’t have neuropathy/sensory issues (which are common complications of decades of living with type 1 diabetes), and muscle damage and inflammation markers are normal. MRIs are normal. The person submits their data which includes their lab tests, clinical chart notes (scrubbed for anonymity), a patient write-up of what their symptoms are like, and exported data from their phone with years of motion/activity data.
  • A different person with Sjogren’s disease (also an autoimmune condition) and a similar new-onset muscle-related issue. There are known neurological manifestations or associations with Sjogren’s, but more typically those are small-fiber neuropathy or similar. The symptoms here are not sensory or neuropathy. MRIs don’t show inflammation or muscle atrophy. Their clinician is stumped, but knows they don’t see a lot of patients with Sjogren’s and wonders if there is a cohort of people with Sjogren’s facing this. The clinician asks the patient and gains consent; scrubs the chart note of identifying details, and submits the chart notes, labs, and a clinician summary describing the situation.
  • Later, a researcher (either a traditional institutional-affiliated researcher OR a citizen scientist, such as a third person with a new-onset muscle-related issue) decides to investigate a new-onset muscle-related issue. They register their hypothesis: that there may be a novel autoimmune condition that results in this unique muscle-related issue as a neuromuscular disease (that’s not myasthenia gravis or another common NMJ disease) that shows up in people with polyautoimmunity (multiple autoimmune conditions), but we don’t know which antibodies are likely correlated with it. The research question is to find people in the platform with similar muscle-related conditions and explore the available lab data to help find what might be the connecting situation and classify this disease or better understand the mechanism.
  • They are granted access to the platform, start analyzing, and come across an interesting correlation with antibody X, which is considered a standard autoimmune marker but one that doesn’t differentiate by disease, and which does highly correlate with this possibly novel muscle condition when it exists and is elevated in people with at least one existing autoimmune condition and these muscle-symptoms that are seronegative for every other autoimmune and neurologic condition. Further, by reading the clinical summary related to patient B and the self-written narrative from patient A, it becomes clear that this is likely neuromuscular – stemming from transmission failure along the nerves – and that the muscles are the ‘symptom’ but not the root cause of disease. This provides an avenue for a future research protocol to 1) allow these types of patients to be characterized into a cohort so this can be determined whether it is a novel disease or a subgroup of an adjacent condition (e.g., a seronegative subgroup);  2) track whether the antibody levels are treatment-sensitive or not or stay elevated always; and 3) cohorts of treatments that can be trialed off-label because they work in similar NMJ diseases even though the mechanism isn’t identical.
  • The mechanism for this novel disease isn’t proven, yet, but in the face of all the previous negative lab data and neurological testing patient A and patient B have experienced, it narrows it down from “muscle or neuromuscular” which is a significant improvement from their previous situations. Plus, this provides pathways for additional characterization; research; and eventual treatment options to explore versus the current dead ends both patients (and their clinical teams) are stuck in. And because there are no good clinical pathways for these types of undiagnosed cases, this type of insight development across multiple cases would not have occurred at all without this database of existing data.

2. The above is a small-n example, but consider a large dataset where there are hundreds or thousands of people with CGM data submitted. Plus meal tracking data, because people can export and provide that from whatever meal-logging apps some of them happen to use.

  • By analyzing this big dataset, an individual researcher could hypothesize that they could build a predictor to identity the onset of exocrine pancreatic insufficiency, which can occur in up to 10% of the general population (more frequently in older adults, people with any type of diabetes, as well as other pancreas-related conditions like pancreatitis, different types of cancer, etc), by comparing increases in glucose variability that correlate with a change in dietary consumption patterns, notably around decreasing meal size and eventually lowering the quantity of fat/protein consumed. (These are natural shifts people make when they notice they don’t feel good because their body is not effectively digesting what they are eating). They analyze and exclude the effect of GLP1-RA’s and other medications in this class: the effect size persists outside of medication usage patterns. This can later be validated and tested in a prospective clinical trial, but this dataset can be used to identify what level of correlation between meal consumption change and glucose variability change happen over what period of time in order to power a high-quality subsequent clinical trial. This may lead to an eventual non-invasive method to diagnose exocrine pancreatic insufficiency through wearable and meal-tracking data. (None exists today: only a messy stool test that no one wants to do and is hampered by other issues for accuracy.)

–

These are two examples and show how even small-n data or a large dataset where there are additional subgroups with additional datasets can be useful. In the past, it’s been a “we need X, Y, Z data from everyone. A, B, C would be nice to have, but it’s hard and not everyone will share it (or is willing to collect it due to burden), so we won’t enable it to be shared for the 30% of people who are willing”. Thus, we lose the gift and contributions from the people who are able and willing to share that data. Sometimes that 30% is a small n but that small n is >0 and may be the ONLY data that will eventually answer an important future research question.

signal-noise-more-examples-DanaMLewisWe are missing so much, because we don’t collect it. So, we should collect it. We need a platform to do this outside of any single disease group patient registry; we need to support clinician and patient entries into this platform; we need to support intake of a variety of types of data; and we need to have low (but sensible) barriers to access so individuals (citizen scientists, patients themselves) can leverage this data alongside traditional researchers. We need all hands on deck, and we need more data collected.

Want signal? We need more noise (looking at the quiet bottleneck)

by Dana Lewis

We need more signal, which means we want more noise. A lot of current scientific infrastructure is designed to minimize messiness: define a narrow question, collect the minimum data required to answer it, standardize the dataset, exclude complicating variables, finish the analysis, publish the result. That approach is understandable. It is also one reason we repeatedly fail the patients who most need evidence: people with unusual responses, multiple conditions, atypical phenotypes, rare diseases, or combinations of features that do not fit cleanly into existing bins. Our systems today are structurally designed to discard or never capture the kinds of heterogeneous, partial, contextual, or longitudinal data that could eventually make critical insights available to us.

My hypothesis is that one important scientific bottleneck is not a lack of intelligence, or even a lack of data in the abstract, but a lack of infrastructure for accepting, preserving, and reusing the kinds of data that fall outside formal trial endpoints and standard clinical workflows. I am proposing a solution: a shared platform that accepts optional, heterogeneous, participant- and clinician-contributed data, which will generate clinically useful subgroup hypotheses that standard trial and registry structures fail to generate.

Here are the quiet bottlenecks in the current system that I run into and that this platform would address:

  1. Atypical, comorbid, and small-population patients are systematically left without usable evidence because current research systems are optimized to suppress heterogeneous data rather than preserve it. (Clinical trials are usually designed around narrow endpoint collection, fixed inclusion and exclusion criteria, and standardized datasets that are tractable for a specific immediate question. Patients are often told there is “no evidence,” when what is really missing is a system that could have captured and preserved evidence that can help with future translation for atypical or edge-case presentations.)
  2. Useful data is routinely discarded because no existing structure is responsible for capturing it. (Clinical trials generally do not want to collect data outside primary and secondary endpoints because of cleaning, storage, analysis, and governance costs. Clinical care systems do not want to ingest large volumes of patient-generated data that clinicians are not expected to review. Disease registries are usually narrow and disease-specific. Journals are poor infrastructure for surfacing up data. There is no home for this data.)
  3. AI and humans are constrained by the same missingness problem: the knowledge base is biased and uneven because we do not capture enough of the right kinds of real-world data. (LLMs and other AI systems are often criticized for bias, holes, and nonrepresentative training data. But this overlooks that clinicians and researchers are also reasoning from the same incomplete literature, selective trial populations, under-collected real-world evidence, and publication-filtered case reports. The bottleneck is upstream of both.)AI now makes it possible to overcome these constraints, but only if the data is collected and made available in the first place.
  1. Current systems privilege uniform completeness over partial but valuable contribution, which causes preventable signal loss. (If not everyone has or is willing to provide the data, it’s not collected. Subsets of participants who are willing and able to contribute additional data, such as wearables, CGM, meal logs, phone sensor streams, or symptom data, are prevented from doing so.)

Why current structures produce these bottlenecks

No single existing institution is responsible for optional data intake, long-term stewardship, and broad downstream access. And different structures prioritize different data capture, with no harmonization between these two settings. Clinical trials are funded, regulated, and staffed to answer bounded questions. Clinical trial incentives reward endpoint completion, analyzable datasets, and publication-ready results. Anything beyond the core clinical trial protocol creates additional costs: more data cleaning, more storage, more governance, and more analytic work that may not directly serve the trial’s primary purpose. Under those conditions, narrowness is rational but it is also systematically lossy. Clinical care systems have a different but related design problem. Electronic health records are built for documentation, care coordination, and billing, not for capturing or allowing participant-generated, future-use data that a clinician does not need for real-time decision making. This illustrates the infrastructure gap: patients may have useful longitudinal data, but there is often no legitimate pathway to store it in a way that supports future analysis. Disease registries and advocacy-group datasets help in some cases, but they are typically narrow in scope, tied to a disease silo, and constrained by the same pressures toward standardization and limited datastreams. Publications are also poorly matched to this problem, because they are optimized for polished outputs, not for preserving heterogeneous data. Much of the infrastructure was shaped from prior underlying constraints (when data capture, storage, and analysis was more expensive).

Testable hypotheses that would address these challenges:

  1. Allowing partial, nonuniform data donation from participants will outperform “uniform minimum dataset only” models for early discovery in rare, atypical, and comorbid populations.
  2. Structured case narratives combined with quantitative data will enable more useful discovery than quantitative data alone for poorly characterized conditions, because the narrative context helps identify mechanistic or phenotypic patterns that are otherwise lost.
  3. Modern tools (LLMs etc) make it feasible to analyze and normalize real-world participant-contributed data at a scale and cost that was previously impractical.
  4. For some under-characterized conditions, a heterogeneous real-world repository can identify candidate biomarkers, phenotypic subgroups, or prospective-study designs and lead to a shorter timeline for protocol development and funding of eventual trials, compared to the status quo.

Here are a few example experiments we can use to validate these hypotheses:

  1. Use heterogenous real-world data to test whether a shared repository can generate an earlier detection hypothesis that existing structures are poorly positioned to generate. For example, exocrine pancreatic insufficiency. EPI is common but underdiagnosed and its current diagnostic pathway is unpleasant (a stool test) and imperfect. As digestion becomes less effective, people often change what and how they eat before anyone labels it as a problem: meal sizes may shrink, fat and protein consumption may drift downward, symptom patterns develop, and glucose variability may shift in parallel, especially in people with diabetes or others who happen to have CGM data. A conventional clinical trial would rarely collect all of these data together (GI symptoms, meal logs, and CGM data). We have shown that a patient-developed symptom survey can effectively distinguish between EPI and non-EPI gastrointestinal symptom patterns in the general population. But this relies on people to know they have a problem and fill out the symptom survey to assess their symptoms. By analyzing CGM and meal logging data, we may be able to create a diagnostic signal from CGM data that may provide an earlier detection of EPI or other digestive problems, noninvasively and much earlier. If successful, the output would be a concrete, testable hypothesis for a future prospective study: for example, that a specific combination of changing meal patterns and glucose variability could serve as an early noninvasive screening signal for EPI.
  2. A second, smaller experiment would test the same infrastructure in a very different setting: under-characterized autoimmune or neuromuscular overlap syndromes. Here the problem is not underdiagnosis at scale, but invisibility in small-n edge cases. Patients with unusual muscle-related symptoms, normal or ambiguous standard workups, and backgrounds that include autoimmune disease often remain isolated within separate clinics and disease silos. One person may carry a type 1 diabetes diagnosis, another Sjögren’s, another a different autoimmune history, and yet their new symptoms may share an overlooked mechanism. In current systems, these cases rarely become legible as a group because the relevant data are scattered across chart notes, patient stories, normal imaging, lab panels, and passive longitudinal data that no one is collecting or comparing systematically. This tests the same core hypothesis under tougher conditions: smaller numbers, less uniform data, less obvious endpoints, and a heavier dependence on narrative context. If the EPI case shows that optional heterogeneous data can support tractable hypothesis generation in an underdiagnosed condition, the neuromuscular case shows why the same infrastructure could matter even more in sparse-data, high-ambiguity situations where it is even harder to capture data to generate evidence in support of and funding for subsequent trials.

What we might learn if these fail

We will learn what the barrier is, whether it is still a problem of infrastructure; a lack of the right people (or AI) leveraging the data; whether partial nonuniform data donation is operationally feasible; and whether limiting factors are data availability, harmonization, governance, or analytic quality. Plus, we can determine whether certain diseases or use cases (e.g. developing novel diagnostics versus assessing medication responses) are better suited than others for this type of platform.

Why now?

Passive and participant-generated data collection is easier. Wearables, phone sensing, CGM data, meal logging apps, symptom trackers, and similar are now significantly more common. Technology makes it easier than ever to create custom apps to track n=1 data or study-specific data. Storage is cheaper. Technology improvements, most notably AI tools, have made it more tractable to collect data; and have made it more tractable for researchers to analyze this data. The remaining bottleneck is capturing, storing, and making the data available. It is less of a technological bottleneck and instead a bottleneck of funding/governance/etc. This is addressable now that the capture/analysis barriers have been lowered!

–

signal-noise-DanaMLewisI don’t think the question we should be asking is whether every piece of heterogenous data will be useful, but whether we can afford to keep doing what we are doing (throwing away data and still expecting discovery for the populations our current systems and infrastructure already quietly fail).  

–

Note: I am submitting this post to the Astera essay contest, which you can read about here. You should write up your ideas about the bottlenecks you see and submit as well! I also wrote an additional piece with more details and examples, which you can read here. 

The data we are leaving behind in clinical trials, what it costs us, and why we should talk about it

by Dana Lewis

I talk a lot about the data we leave behind. A lot of this time, this is in the context of clinical trial design. But more recently, it’s also about the data loss that drives bias in AI. And actually…that same bias exists in humans. And seeking to fix one may also fix the other!

To clarify, it’s not that no one cares (as a reason for why data loss or missingness happens). It is an artifact of our systems, funding priorities, and a whole bunch of historical decision making. Clinical trials cost money and are usually designed to answer a question around safety or efficacy, often in pursuit of regulatory approval and eventual commercial distribution. We need that (I’m not saying we don’t!), but we also need funding sources to answer additional questions related to titration, individual variance, medication impacts, more diseases, etc.

A lot of trials exclude people with type 1 diabetes, for example, so it’s always a question when new medications or treatments roll out whether they are appropriate or useful for people with T1D or if there’s a reason why they wouldn’t work. This is because sometimes T1D might be seen as a muddy variable in the study, but a lot of times it’s just a copy-paste decision from a previous study where it did matter that is propagated forward into the next study where it doesn’t matter. There is no resulting distinction between the two: there is simply a lack of representation of people with T1D in the study population and this means it’s hard to interpret and make decisions about this data in the future.

A lot of decisions around what data is collected, or who is enrolled in a population for a study, is an artifact of this copy-paste! Sometimes literal copy-paste, and sometimes a mental copy-paste for “we do things this way” because previously there was a reason to do so. Often that’s cost (financial or time or hardship or lack of expertise). But, in the era of increasing technological capabilities and lower cost (everything from the cost of tokens to LLM to the decreasing cost of storage for data long-term, plus the reduced cost on data collection from passive wearables and phone sensors or reduced hardship for participants to contribute data or reduced hardship for researchers to clean and study the data) this means that we should be revisiting A LOT of these decisions about ‘the way things are done’ because the WHY those things were done that way has likely changed.

The TLDR is that I want – and think we all need – more data collected in clinical trials. And, I don’t think we need to be as narrow-minded any more about having a single data protocol that all participants consent to. Hold up: before you get the torches and pitchforks out, I’m not saying not to have consent. I am saying we could have MULTIPLE consents. One consent for the study protocol. And a second consent where participants can evaluate and decide what additional permissions they want to provide around data re-use from the study, once the study is completed. (And possibly opt-in to additional data collection that they may want to passively provide in parallel).

The reason I think we should shift in this direction is because that, out of extreme respect to patient preferences around privacy and protecting patients (and participants in studies – I’ll use participants/patients interchangeably because it happens in clinical care as well as research), researchers sometimes err on the side of saying “data not available” after studies because they did not design the protocol to store the data and distribute it in the future. A lot of times this is because they believe patients did not consent (because researchers did not ask!) for data re-use. Other times it’s honestly – and I say this as a researcher who has collaborated with a bunch of people at a bunch of institutions globally – inadvertent laziness / copy-and-paste phenomenon where it’s hard to do the very first time so people don’t do it, and then it keeps getting passed on to the next project the same way, because admittedly figuring it out and doing it the first time takes work. (And, researchers may have not included in their budgets to cover data storage, etc.). Thus this propagates forward. Whereas you see researchers who do this on one project tend to continue doing it, because they’ve figured it out – and written it into their research budgets to do so for subsequent projects. And other times, the lack of data re-use is a protective instinct because researchers may have concern that the data may be used in a way that is harmful or negative. Depending on the type of data (genetic data versus glucose data, as two examples), that may be a very legitimate concern of study participants. Other times, it may be at odds with the wishes of the broad community perspective. And sometimes, individuals have different preferences!

We could – and should – be able to satisfy BOTH preferences, by having an explicit consent (either opt-in or opt-out, depending on the project) that is an ADDITIONAL consent specifically indicating preferences around data being re-used for additional studies.

To emphasize why this matters, I’ll circle back to AI/LLMs. A criticism of AI is that they are trained on data that is not representative; has holes; or is biased. This is legitimate and an artifact of the past design of trials. I have been thinking through the opportunities this provides now, in the current era of trial design, to be proactive and strategic about generating and collecting data that will fill these holes for the future! We can’t fix the way past trials were done; but we can adjust how we do trials moving forward, fill the holes, and update the knowledge base. This actually then fixes the HUMAN PROBLEM as well as the AI problem: because these criticisms of AI are the same criticism we should be making about humans (researchers and clinicians etc), who are also trained on the same missing, messy, possibly biased data!

More data; more opt-in/out; more strategery in thinking about plugging the holes in the knowledge base moving forward in part by not making the same mistakes in future research that we’ve made in the past. We should be designing trials not only for narrow endpoints (which we can also do) for safety and efficacy and regulatory approval and commercial availability, but also for answering the questions patients need answered when evaluating these treatments in the future: everything from titration to co-conditions to co-medications to the arbitrary research protocol decisions around the timing of treatment or the timing of the course of treatment. A lot of this could be better answered by data re-use beyond the additional trial as well as additional data collection in conjunction to the primary and secondary endpoints.

All of this is to say…you may not agree with me. I support that! I’d love to know (constructively; as I said, no pitchforks!) what you disagree about, or what you are violently nodding your head in agreement about, or what nuances or big picture things are being overlooked. And, I have an invitation for you: CTTI – the Clinical Trials Transformation Institute – is hosting a 2-hour summit on April 28 (2026) for patients, caregivers, and patient advocates designed to facilitate discussion around the range of perspectives on these topics as they apply to clinical trials. Specifically 1) the use of AI in clinical trials and 2) perspectives on data re-use beyond the initial trial that data is collected for.

If you have strong opinions, this is a great place for you to share them and hear from others who may have varying or similar perspectives. I am expecting to learn a lot! This event is not designed around creating consensus, because we know there are varying perspectives on these topics, and thus especially if you disagree or have a nuanced take your voice is needed! (I may be in the minority, for example, with my thoughts above.) The data we leave behind in clinical trials, what it costs us, and why we should talk about it - a blog by Dana M. Lewis on DIYPS.orgParticipation is focused (because these topics are scoped to focus on clinical trials) on anyone who has ever been part of a clinical trial, is currently navigating one, made the decision to leave a trial early, or even those who looked into participating but ultimately decided it wasn’t the right fit .

You can register for the summit here (April 28, 2026 10am-noon PT / 1-3pm ET), and if you can’t attend, feel free to drop a comment here with your thoughts and I’ll make sure it is shared and included in the notes for the event that will also be shared out afterward aggregating the perspectives we hear from.

Which LLMs and how I am using them as of February 2026 plus a remote LLM prompting tool

by Dana Lewis

I thought I would share some of how I am using AI. Previously I’ve talked a lot about broad concepts of prompting LLMs and the types of projects that can be done. I’ve provided particular examples that I’ve found useful for n=1 custom software (to address software-shaped feelings, often) or for improving n=1*many software like my apps PERT Pilot (iOS | Android), Carb Pilot, etc. But I haven’t talked a ton about which LLMs I use and why.

There’s a reason for that.

I noticed back in the day (of artisanal hand coding, going uphill in the snow both ways to build things) that a lot of programmers would be super dismissive and bullying about which language/program you talked about using. Let’s say you picked up a project and built it in C++: someone might come in and declare that it is the stupidest thing ever, and you should’ve used Python. But if it works in C++…who cares? The same thing about whether you use a visual “IDE” or code from the terminal, and whether you use tabs or spaces, or prefer vi or vim or nano…who cares? Use what works for you. I always noticed and hated and tried to ignore the “you should do it this way!” unless there was a “hey, heads up, doing it this way is better because very concrete reason X,Y,Z”.

So when you read my blog posts about LLMs, you might pick up on my generic use of talking about AI/LLMs without calling out a specific one (for example, see my last post where I talked about using LLMs to build solutions for software-shaped feelings). That’s because if the first one you picked up was Claude, and you liked it – great. Use that. All of my advice regarding prompting and other stuff applied. But if you used and liked ChatGPT? Also great, advice also applies. It’s easy to read (as a non-technical person) advice that’s for ChatGPT and be confused about whether you can extrapolate it to another tool and environment. That’s now exacerbated by the fact that we went from web-based AI (think ChatGPT and Claude websites) to desktop and phone-based apps to now also agentic “CLI” (command line interface) options. There’s oodles of ways to use an LLM of choice depending on where you are, what you like, and what you are trying to do. Try one and if it works for you, great. Be willing to try other things, too, but it’s fine if it doesn’t work as well for you and you keep using what does work. (If it doesn’t work, that is a good reason to try something else, as different models are better at different things and the different tools may have different capabilities).

Here’s what I have used over time, what I’ve tried and why I’ve switched over – or not switched at all.

First, I mainly used ChatGPT (web) because that’s what was available in the early days. I found it useful especially as the models progressed. I also had tried Claude when it became available (web), but didn’t find it as useful. I tried a few times with different types of projects and prompts. It was ok-ish, but I preferred ChatGPT especially for technical projects. I also tried Gemini (web) and had a similar lackluster response. I was thrilled when there became a (Mac) app for ChatGPT, and when there became more functionality like tool integration. Before, I was having to copy in code and get it to re-write or advise, then I would copy/paste code back over. (Hey, still easier than noodling over stack overflow entries and trying to figure out how it applied to my project). But, it was burdensome. Then it got to where it could see the context of my file, but only one file. That was better but still lacked the context of everything else. However, for one-off questions about random topics and other types of non-coding work, I still preferred ChatGPT’s models overall, especially increasingly the models with thinking. I now use 5.2-thinking as the default for all my work (this was written in early February 2026; assume that in future when you read this, I will probably be on the latest thinking/reasoning models).

Note: if someone says “I used GPT 5.2” when describing what they did, you have to ask what level of thinking they had it set to. Was it auto mode? Or thinking mode? Yes, sometimes auto mode will eventually trigger thinking mode but it drastically differs based on the prompt, and I find most users on free accounts using auto don’t see it go into full thinking the way my prompts get 5.2-thinking from the start, and that influences how useful the answer is. And again, it likely doesn’t have context for the project or if it does only one file’s worth.

Meanwhile, Scott has been a big fan of Claude Code, the agentic CLI for Claude’s models. He’s very, very productive with it for work on his laptop using the CLI tools. I had on my list to try CLI tools, but I put it off for a long while in part because I had finally experimented with Cursor.

I started using Cursor to test if it was better (for me) for coding-related projects. It’s basically a desktop app that I can open a project folder in and see all the files – whether they’re XCode (iOS app) files or Android Studio kotlin or javascript or python or markdown or whatever. I don’t have to open all the different ‘run’ apps to work on the code (although I still need to use XCode or Android Studio to ‘build’ and run the simulator or deploy to my physical phones for testing, these are no longer my primary interfaces for the coding work). And, this means all the files in the project are *in context* (or, able to be searched within a prompt to give it context), without me having to go tell it to look at file A for context – it can look itself and figure out which file is relevant or has the relevant code. (This is also because of the increasing capabilities of ‘agents’ which can do more sophisticated multi-step things in response to your prompts.) Cursor allows you to pick which models you want to use, so you could limit it to just Claude or ChatGPT-related models, but I don’t – I use all the latest models. I can’t really tell a difference with this setup what model is being used (!) which honestly shocked me because from the web version I *can* tell such a difference between the two, but the Cursor-style setup that has context for the full project takes away that difference. And Cursor has been very fast to iterate and instead of having it be slow to go prompt by prompt, they now have planning mode and agents where you can start by having it plan a set of work and build a checklist (even a big one with lots of tasks) and then it’ll go do some, most, or occasionally all of it. It’s also gotten better about running things in a sandbox to test its work, so I spend a lot less time on iOS projects: for example, going into XCode and having to come back to Cursor and complain about a build error that is preventing me from testing. Instead of that happening 3 of every 4 times, it now seems like it happens only 1 of 4 times and often this is a very minor issue that it fixes with a single follow up action and I can build. So that prompt>build>test timeline loop has gotten WAY faster, even just comparing the last 3 months or so. Because I have found such success with my iOS projects, Cursor has become my go-to interface of choice harness/setup for most projects, including Android apps, iOS apps, new data science projects, and anything else that involves code and/or data analysis.

However, Cursor is a desktop app. One of the benefits of ChatGPT is that I could do something on the desktop app and walk away and go for a walk and if I thought of something, pull out my phone and add another prompt to it via the phone app in the same thread that had the same context window. (Mostly – it can’t access a code file from my phone, but whatever else was loaded in the chat history it has). This means that ChatGPT was still my mobile/phone LLM of choice. I often found myself with my phone dreaming up a new project and essentially doing “planning” mode with ChatGPT (5.2-thinking, as of Feb 2026) for scoping out a project and thinking through constraints. When I was back at my laptop, I would have ChatGPT write a summary and/or prompt to give to another LLM, then go give it to Cursor. I’d have Cursor riff on that plan and add to it, then go implement it. This has worked wildly well, but I do wish I had a way to do at least planning mode with Cursor from my phone.

But, aside from that, I have felt as productive with Cursor (for laptop usage) and ChatGPT (for phone based and non-coding usage) as Scott does with Claude Code (on his laptop; he also uses ChatGPT as his phone LLM of choice).

Lately, you might have heard about “OpenClaw” (aka “MoltBot” aka “ClawdBot” – it had a few name changes). It’s a security nightmare, so don’t just use it if you’re not thinking it through, but the TL;DR is that it’s a way to set up a way to send prompts to an LLM and have it run them on your computer, with full access to all the different tools and accounts you’ve set up. However, it’s a security nightmare and very risky, so most people shouldn’t be using it (including you, probably). The only reason I’m bringing it up is because I *wish* I had a way for mobile commands to a coding LLM that had access to my code, so I could basically instruct Cursor to work on something while I’m away from my laptop. But, I am not willing to set something up with permissions to remote control my laptop fully (which is basically what OpenClaw does), so I haven’t been able to do so.

Until this weekend.

I kept thinking that what I wanted was the ability to remotely prompt Cursor, because it has access to the full project. There are a few ways to do this, either via something like OpenClaw (nope, see above) or if you host your code on Github and run ‘cloud’ based LLMs on it. I didn’t want to do that, the projects I have are local on my computer, and I want the LLM to have context and ability to write code and edit code on that project. (Otherwise, I’ll just use ChatGPT which is sufficient on my phone for other things). This could be because I’m away from my laptop – cross country skiing in the mountains and thinking through projects and wanting to start a prompt on the drive home so it can have something for me to review and test when I’m back, or just in the other room and don’t want to go get my laptop out. If only there was a way to have Cursor check a Google Sheet, for example, for a new prompt and then go run it.

Light bulb.

Cursor (desktop) can’t, but a scheduled job on my computer could be set up to check a Google Sheet every so often and find the prompt and send it to an LLM agent that can be run on my laptop. I could have the LLM log the prompt and what it did and the output, so even though I wouldn’t see it in the ‘chat history’ of the desktop app the way I do with ChatGPT or Cursor (desktop), I could still review a summary of what it thought/did and review the code when I was back on my laptop. (Bonus: because this project folder is synced to a backup cloud that I have access to on my phone, I can review this from my phone, too!).

How did I figure this out and make it work? I used ChatGPT (as my phone-based brainstorm LLM app of choice) to talk through the idea and talk through technical options, including my ‘requirements’/preferences of not using Github/cloud based actions for this. This discussion laid out a plan and I was surprised that the Cursor CLI (as opposed to desktop) would work for this, which means I could download Cursor CLI and log in/set that up (under the same account I used Cursor Desktop with), and with a simple script I added to my Mac via Terminal, have it check the Google Sheet at the designated interval, find the prompt, run it, record the status in the sheet plus record a status in the log of the project.

So after all, I am finally using a CLI for an LLM! (Although, I will say after I downloaded it and installed and logged in just to set it up for this project, I actually hate the UI, go figure, so my resistance to CLI-based LLMs was probably an intuitive strong preference for GUI-based LLMs.) But I’m not using it directly as a CLI, my script on my Mac is calling it to do the work based on the prompt I put into a Google Sheet. Google Sheets, which I’m very comfortable with, becomes the place I prompt my code-based LLM to do remote work from. (FYI, for setup, I’m also using a free Google Cloud Console Project which you give permissions to which is part of the way it can read/write from the Google Sheet remotely.)

End to end, this setup involved 1) terminal (on my Mac); 2) Cursor CLI; 3) Google Cloud Console; 4) Google Sheets.

I will have to continue to set this up for other projects that I want to do, but the bulk of the hard work is done – the additional projects will be a few minutes of configuration and testing each time, so I’ll do that every time I touch those projects and find myself wanting to do remote work.

But this is exactly what I love about AI-driven development. You can build amazing things that would not have been in your wheelhouse to even attempt before. You can troubleshoot and fix all kinds of issues along the way on the big ambitious project that becomes a realistic option to build and implement, whether it’s just for you or it’s something that a lot of other people would use. And, you can use it to customize your environment and tools and make up completely new tools (again, which you never would’ve thought was possible to do before!) that enable you to do work in ways that weren’t possible before. I keep saying this, but it’s worth keeping a log somewhere of projects that are still “too hard” and coming back to them every few months and trying them again with newer models or setups. (If you haven’t used anything with ‘agents’, you REALLY should try this on a project.) 

February 2026: which LLMs I use (and why they have changed), a blog post by Dana M. Lewis on DIYPS.orgIt’s exciting to keep adding to existing projects, building new projects you didn’t have time to do before, and building the projects that you only figured out how to dream up with the help of new technology.

Software-Shaped Feelings

by Dana Lewis

I have to ask myself these days, is this a software-shaped feeling?

This is a riff on something I see online more and more frequently, where people comment about recognizing if this is a “software-shaped problem”. Meaning, is this something that can be solved with software? More recently, this often means n=1 custom, boutique software that we can build ourselves. This is often easier and even faster than trying to search for something that maybe exists elsewhere and kinda works but not perfectly and requires a lot of work to tweak to the use case. Instead, now, we can spend the same (or often a lot less) time building something that perfectly fits.

I have been doing this for years. Juggling data in my head is hard – so I often turn to spreadsheets. For example, trying to figure out my enzyme dosing originally for exocrine pancreatic insufficiency – I had too many meals and too many variables to track, so I made a spreadsheet. When I then incorporated dosing enzymes for everything that I ate, that meant also everything I was eating while ultrarunning – which I did every 30 minutes to fuel. So I made a spreadsheet for that and then a template with dropdowns to use while running. I could quickly enter data that way, but Google sheets was slow to load on my phone and the extra seconds of delay was frustrating. Thus, I made a simple app I called “Macros On The Run” to allow me to more quickly open and tap the drop down and select my fuel item. Since I was building a custom app interface I could have all the summary data I wanted from my run, both hourly stats on calories and sodium consumed but also the entire per-run stats. I then was able to export my data to be able to analyze it later (particularly, also transposing it into the spreadsheet I use for food and enzyme intake). Does anyone else have this use case? No, probably not. But it was so empowering to be able to track and capture information in a format that worked well for me. And because I had learned how to build apps to build PERT Pilot to share with other people (because not everyone loves spreadsheet interfaces), I knew how to build apps, so I had the skillset that translated to figuring out how to build another app. This was before agentic AI, by the way, but this is now even easier.

Enter this week. I realized that I had expressed frustration to both my mom and Scott about a situation where I have a medicine/treatment that is done once a week, but it’s not a time-anchored or time-specific treatment. I don’t always take it Tuesday at 7pm. Depending on my schedule it could be 5pm, 9pm, whatever. But also, if I’m traveling, it can be a day earlier or later. It’s 7 days, plus or minus one day. Previously I was adding an event to the calendar for every week, and then moving it around. But I realized looking ahead at a busy summer full of travel that it was hard to move one without having to move another, because it’s not the week before that only matters, moving one also influences the chain of following events. If I usually do it on Tuesday but I’ll be traveling and not home until Thursday (and don’t want to do it while traveling), then I need to move the previous week from Tuesday to Wednesday so the following week’s Thursday is not out of the allowed ‘window’. So instead of 7-7-7, or ending up with 7-9-5 (which means two weeks are out of bounds), it becomes 7-8-6. But because these are single calendar events, there’s no alert when something goes out of bounds/out of the ‘allowed window’. Setting a recurring event (eg calendar event every 7 days) doesn’t resolve this, either. Juggling all of this in my head felt like a waste and non-optimal.

If only, I thought, I had a way to create an event type that was based on the “chain” relationship to the events before and after, that I could set to every 7 days but have an ‘allowed’ grace window of plus or minus a day, so that anything 6-8 days was allowed.

Google calendar doesn’t support this. But I know that you can subscribe to other calendars and show them, eg how I edit events in TripIt and subscribe to the calendar feed so I can see those events in Google calendar.

Thus, I found myself asking, “is this a software-shaped feeling?” of frustration? I can build apps, after all. Could I build an app that somehow publishes a calendar feed that I can see in Google calendar and see these events and get alerts when they are out of bounds? Ideally I’d be able to use Google calendar to edit these events, the way I do my other calendar events.

The mental model I had was app>feed>display in calendar. I was laying in bed, but I pulled open one of my LLMs that I have on my phone and sketched out the idea in a few sentences and asked what was possible to do and what technically it would take to do this. It brainstormed two key options, one with Google authentication and complicated multi-step back end (storage and databases and hosting and servers and all the stuff I *can* do but hate doing and doesn’t make sense for a quick n=1 prototype), and one simple one where the app would create its own iCloud calendar and then publish a feed I could see via the Google calendar. Aha, I thought, that’s it.

The next day I sat down at my laptop and used my coding LLM of choice. I asked it to build this app where it would allow me to create a ‘series’ of events with a set window (eg default to Tuesday at 7pm) and allow me to set a grace period (eg +/-1 day) and see the distance between events and flag a visual cue if they fall out of the window. It should then create events in an iCloud calendar. I would then subscribe and view it via Google calendar, so I could see it via my laptop or phone. Eventually I would see if I could edit it from those views, but for now, just being able to push and adjust from the app would be fine. (Note: my prompt was a little more specific than this, but the important part wasn’t telling it how to build the app or what I wanted it to look like. The important part is making sure it understands the end goal and user interactions so we design the backend and the front end to support what I actually want to do as the end user.)

It…did it. This isn’t a story of “one-shotting” and getting things perfectly from a single prompt, but I continue to be appreciative that the first hours of setting up a project and getting it to build on your device and get the basic functionality working that used to take 4-6 hours and lots of gnashing teeth is now a 15-20 minute endeavor, usually. And that’s what happened here, I had a working functional prototype working on my phone in about 15 minutes. I spent more time after that making it look better, testing it, adjusting events, etc, but the initial setup and functionality was a very quick process. That’s huge. For me, it doesn’t matter what the total amount of time it is to get a project ‘done’ or usable for real-life testing. You’ll see people talk about how you get 80% of the work done quicker but it ‘takes longer’ for the last 20%, but I honestly think I do way more (so it’s not just 20%) when I haven’t exhausted myself gnashing my teeth about the fundamental basic how-does-it-work and does-it-build and what-is-this-build-error.

This app probably doesn’t matter to anyone else, but if you have a use case for it and would like to try it, let me know. I called it SchedulerPilot. It creates a new iCloud calendar called SchedulerPilot that you can see on your phone, and then also subscribe to elsewhere to view (e.g. if you set it to be viewable via a link), such as in Google calendar. That means I can give Scott access to see it the way he sees my other calendars, plus view it from my browser-based calendar view on my laptop. I can’t edit events from Google calendar in the browser, but I was able to (much more easily than I expected) add functionality to edit directly from my phone’s calendar app the way I usually do on the go. I also then was able to add functionality in the app for it to flag when the in-app view and calendar view were out of sync. It shows a list of events and how they differ and I can choose whether I pull calendar changes into the app view, or if I did the tweaking in the app, push those changes back to the calendar. They automatically sync. I can also add notes to the event and ‘lock’ it so if I make other adjustments and push updates, it doesn’t overwrite it. This is key for travel-based constraints where I am definitively moving it to do when I get home from a trip, but sometimes I haven’t booked flights for the trip so looking at my calendar it may otherwise be confusing WHY the event is moved for the week. Having a lock shows it’s there for a reason, and the optional note also provides good context clue for why it’s moved. I can also add more events in the future with the tap of a button, if I’m still using this as my method for scheduling in the future (and I’m guessing I will be).

Here’s an animated demo showing the point of the app and why it’s useful:

software-shaped feelings. a blog post by Dana M. Lewis on DIYPS.orgThe point isn’t convincing you to want to use my app. Most people don’t have a use case for this. But I want more people to build the patterns of recognizing when their feelings are trying to tell them something (e.g. this is a problem we could work on improving, even if it is not solvable) and build the skills of identifying when this is a software-shaped feeling. Especially when we are dealing with health-related situations and especially chronic non-curable diseases, and especially ones with treatments or medications that are not fun, it is so important to solve friction and frustration whenever and wherever we can. Not everything is a software-shaped feeling, but the more we recognize and address, the better off we are. Or at least I am, and I hope to help others do the same.

PS – If you find yourself with a software-shaped feeling and aren’t sure where to start, feel free to reach out and I’m happy to help chat and brainstorm and help you figure out which LLM tools work for you to help build software that fit those feelings. (Even if you tried a particular LLM in the past, they are frequently updating models and tools and it is worth trying again. They have gotten way better and require way less specific prompting, like I outlined here – although there may be some tips you still find useful in this post.) It used to be that you had to fund someone or badger someone to help build something if you didn’t have previous technical skills (or invest a LOT of time and energy) to learn them to do a thing. Now, you don’t have to do that, or plan to have a company to get funding to build a prototype. You can just…build things. And it’s ok if it’s “just” a n=1 solution. If it works for you, it’s worth the time. And chances are, it’ll be something that works for someone else, too. And in the meantime, you’ll have made something that works for you. If you felt like you needed permission, you have it. It won’t fix everything, but that doesn’t mean you should fixing what you can. It feels really good to have forward progress and to do something productive.

Do something, anything, for someone else

by Dana Lewis

I’ve had a quote in the signature of my email for oh, something like 20 years. It says “Doing something for someone else is better than anything you could do for yourself.”

I discovered this a few months after I was diagnosed with type 1 diabetes. It sucked, I was miserable, even though I was figuring things out, because my 14 year old brain was very preoccupied with worrying that everyone would think (know) that I was different. I don’t know why that bothered me so much, except for the fact that I’m human and this is a very human thing. This is part of why I didn’t want an insulin pump, because I didn’t want people to look at me and know I had type 1 diabetes. (I grew up in Alabama: stigma against people with diabetes was strong then and honestly still is strong today.) But eventually I decided to drop that concern because I hated having to wake up every Saturday and Sunday morning to eat 60 grams of carbohydrates and do insulin shots. Waking up to eat felt stupid. So I got an insulin pump, and somewhat accepted that people would know that I have type 1 diabetes. And because of that, I decided to start volunteering in diabetes-related things, no longer worrying that people would “know” that I had diabetes. And it turns out…focusing on that, and other people, did so much more for me than anything I had been doing in the first months of living with type 1 diabetes.

I have retained this awareness and mental model ever since. Any time I start feeling overwhelming emotions, often related to or triggered by new health challenges (yay, immune systems that keep on giving new challenges) and feeling alone, I turn to a plethora of activities centered around 1) getting out and moving through space and 2) trying to do things for other people.

It works. It is very, very effective. Sometimes it’s only a momentary solution, but that brief respite in my mental suffering, so to speak, is enough to help break the cycle. I also then realize that if I find myself falling back into the same pattern, then instead of having brain cycles about that, I should instead use that brain cycle to PLAN to do my next thing, with/for other people.

I have observed lately that in these times (over say, the last 10 or so years), it seems like a lot of people would benefit from these types of strategies but maybe don’t realize they’re having these experiences and that they could intervene and break some of these recurring mental patterns. Maybe they don’t know what to do, so that’s the point of this post: a random assortment of things I’ve done over the years when I’ve hit these spots. Some of these activities I do over and over again and some are newly discovered that I’ve recently started. Some of them will feel silly to you or not a good fit for you – that’s to be expected. The kinds of things I am able to do and like to do are wildly different from what might be interesting and in your wheelhouse. But, consider this the ice breaker for you brainstorming YOUR list rather than staring at a blank page not knowing where to start – I predict looking at this list will spur a range of reactions from “ugh, that’s silly” or “I can’t do that/that doesn’t work for me” to “I could do that”.

And, why should you do this? Maybe you don’t perceive yourself to run into these situations, but…maybe you do. More and more often I am startled when, in professional contexts, someone will offer a happy holiday type wish or a “how was your X break” comment and the response is an Eeyore-style “everything is terrible ::gestures at the world::”.

From the outside, I personally have struggled with what to say in response to these (specifically talking about professional contexts). I recognize that the person saying this has a lot of feelings and feelings are valid. Yet. This is often rude and unprofessional. Why? Because you are projecting a lot of that negativity into professional spaces, which is not always the space for discussing those feelings & the situations that are contributing to them. It’s “drive-by dooming” which is contagious (negative feelings spreading) and unproductive. (I was actually struggling to figure out how to articulate why these situations were bothering me until Ken Jennings did a post thread about it – his point was about drive-by dooming on social media, but I see this pattern happen in professional contexts which has been bothering me and I think this framing applies here, too.) I still don’t know what to say, in the moment, in response to these, because I think people might be easily offended in the moment by trying to gently push back and say I hear your feelings but this isn’t a good place right now / this is not appropriate for this. (And honestly: I have observed this pattern for 6+ years so it is not political in one direction or the other, it is pretty common.)

Since I haven’t figured out what to productively say in the moment that won’t hurt relationships – while recognizing that saying nothing is inadvertently hurting relationships, too, by not helping people see when they are ‘infecting’ other people with drive-by dooming – that’s the point of this post. Maybe it will help someone recognize a reflexive pattern of doing this and give them some ideas for other avenues to address these feelings and/or break some of the cycle of these feelings.

First and foremost – go for a walk/roll/ride (however you move through space). If you’re in your pajamas, fine. No one cares. (Bonus: if it’s at night, no one can see what you’re wearing! So if it’s dark out, that’s even more reason to grab a flashlight or a headlamp and go for a walk/roll/ride.). Maybe it’s a 60 second walk around your yard or down the street and back. It doesn’t have to be far/long/etc but the point is to feel the air moving on your face and to physically move you through space. I figured this out at the start of the pandemic, that physically moving myself out of my environment into an outdoor space and “away” from the physical space where I was having those emotions really helped. Did I still have those emotions? Sure. But it became at least infinitesimally easier to work through them or feel like I was able to “step away” from them slightly. It doesn’t have to be fast, and if the weather isn’t wonderful it doesn’t even have to be outside – go to a different room in your house, even, that isn’t your usual place to sit or lay down, to get some physical separation from your ‘usual’ spots. Bonus points if it is somewhere you can open your window and get fresh air (if you’re not in a bad outdoor air situation).

Second, find something or start thinking of something that you can do for someone else. It can be tiny, it can be for someone inside your house, it can be for someone you don’t know. Even the planning to do something can be powerful, even if it’s not something you can do right away (but do follow through on that). It can be related to the things contributing to your feelings, but you have to actually do things to take action – not just sit with the feelings. And, it can be things that prevent you from seeing things that give you feelings.

What do I mean? If you find yourself ruminating over the state of your country and politics, for example, consider whether this is actually useful to you to do so. Do you NEED to be up to date several times a day on what is going on and what everyone is saying about what is going on? Really? (Unless you’re a social media manager and this is your job, or you work in politics…you maybe actually don’t?). Consider putting a timer on your social media and/or news apps to help you see if you are compulsively checking these apps or news sources. If you do need to see them for another reason, consider muting words and accounts heavily. Turn off auto-play of video content so you’re not distracted by something you didn’t come into the app to see. Curate your feed by unfollowing (which you can often do separately than ‘unfriending’) or muting accounts that you don’t need to see right now. Yes, even if it’s someone you know including a friend or family member. If they are online drive-by dooming everywhere, you absolutely do not have to see that and know what they’re doing/saying and take on THEIR feelings and contribute to your own sense of dooming.

Some digital strategies to consider:

  • Mute/block/unfollow
  • (You can mute words and accounts forever or for things like 24 hours, 7 days, 30 days)
  • Put a timer on apps to limit, or at least help you gauge awareness of how long you are using them
  • Grayscale your phone or icons
  • Remove apps from phone
  • Try a different platform (i.e. pinterest style) where it’s easier to curate a feed and/or a new platform because you can start from scratch intentionally creating a feed

Outside of phones/social media/ digital stuff, though, think about doing other things:

  • Sign up to volunteer at your food bank
  • Join a buy nothing/purchase zero type group and find an “ask” you can meet. Post a few things you’d like to give away.
  • Find a few books you no longer want and look up/look for your nearest “Little Free Library” and go take your books there.
  • Don’t have any books you can stand to give away? Go to a used bookstore and get a few, read them, and then donate them.
  • Live in an apartment building? See someone’s package not delivered to their doorstep (i.e. left incorrectly in the lobby)? Go carry it to their doorstep for them. Bonus: this helps you move through space!
  • Make/cook/bake/create a batch of something and give it to someone else. (If you need ideas and you’ve never made fudge before, try this microwave fudge recipe. Nom. Or make something you love and give some to someone else to try.)
  • It doesn’t have to be home baked. This year at Christmas I was feeling the un-cheer of not having big holiday plans. One of the things I did was look through my boxes of gluten free cookies and realize I have a lot of great treats from other countries (shout out to my friends in New Zealand, Australia, UK, etc) that most people don’t get to try. Also sometimes it’s hard because I feel like I have to eat the whole thing when I open it and I feel guilty if I don’t because they’re special! So two birds with one stone: I opened 4 different kinds of cookies and made a variety holiday cookie plate of GF cookies and went out one night and drove around listening to holiday music and surprised a couple of my friends/their kids by delivering a GF cookie plate. Some of the families have someone who is GF, others aren’t GF. They maybe didn’t care about the cookies but that wasn’t the point. The point was the gesture which they DID appreciate, and I got to feel the holiday spirit (and have a few cookies myself that week without feeling pressured to eat the whole package once open). If you know someone with a food allergy or is GF (celiac or otherwise), this is an AWESOME thing to receive because it’s rare for us to get a “variety” versus getting a whole batch of the same thing (which is ALSO great, but this is slightly different kind of great). But the other point is: these were packaged cookies! I didn’t make them! And I didn’t get fancy plates or anything: I put them on paper plates and put them in clear gallon ziplock bags. And that was fine.
  • Use AI and start a new project and have it help you learn to do something new. This could be something like building your first app or website, anything digital that you maybe thought you couldn’t possibly figure out how to do. These tools have come a long way and it’s a lot more feasible now to tackle just about any project! I went from never having developed mobile apps to developing an iOS app, another one, my first-ever Android app (with stops and starts along the way until the models got good enough to help me through it), all kinds of research and data science steps, and more. Even if you tried a year or two ago and it didn’t work – try again. The models are much, much better, and you should even think about trying again after 3 months because things are changing that much! (And AI can help with non-digital projects, too, it’s helped me think step by step on some of my first paintings that I wouldn’t have known where to start without it.)

The other thing I ended up doing a lot of this season was painting rocks. This started because I walk outside almost every day and several times over the past few years I will randomly spot painted rocks that people leave on the trail to find and move around and trade. I love seeing them. Last year someone did a few holiday ones, and this year I found myself thinking “I wish someone would do that again!”. Oh wait. Maybe *I* should do that?! After all, I have paints. I could get rocks and paint them. They probably wouldn’t be very good, but they don’t have to be. So I started looking up ideas (which is how I rediscovered Pinterest for the first time in 6 years) and found a few designs I thought I could maybe handle doing. And I started painting and realized that it would be fun to surprise people with and give them painted rocks and suggest THEY go leave them somewhere in their neighborhood/trail/wherever to make them smile. I ended up giving my first batch away without having any to put on my trail. So I made some more but they also turned into gifts. They’re not amazing (my hand muscles make it very challenging to paint straight lines for example) but a mix of practice, picking designs that tolerate mistakes, and separating how the rock actually looks (does it look like X?) rather than critiquing how I didn’t perfectly mimic the reference design, means that I’ve continued to paint rocks. I love doing this because 1) it’s fun to surprise people, 2) it gets me into a “flow” state while painting, 3) anytime I get into a negative feeling cycle I can now use the process of planning and picking out my next designs or starting my next batch of rocks (e.g. doing the base coat of paint that has to dry first) as an outlet, and 4) it encourages people to go out and put them out which is good for them (makes them happy) and will make someone else smile the way I get thrilled when I see painted rocks out in the world.

Examples of painted rocks including a fox peeking out from behind a snowbank next to a winter tree; a blue aurora reflected on a lake with trees; three cats doing christmas carols; a grinch; cats decorating a christmas tree; and a winter scene with sheep in a snowy field.Maybe painting isn’t your thing (it wasn’t my thing until I tried it this year!), but the point is to find something that you could learn to do or use something you can do and tie it into doing things for other people. It will feel better than having the cycles of negative feelings where you feel like you can’t do anything. And maybe you can’t do anything about the situation that causes the negative feelings, but maybe you could. And if not, doing something else to distract yourself or inject more positive feelings into your life could be a great strategy to try.

Doing something for someone else is sometimes better than anything you could do for yourself. A blog by Dana M. Lewis on DIYPS.orgThink “motion is lotion” and by doing something, anything, it will give you momentum so then you can do something else, including tackling bigger issues or problems or needs that you have.

(PS – if you have a favorite thing that you fall back to doing when you find yourself tipping into these patterns, please comment below – I’m always looking for things to add to *my* list to be able to do!)

Baseline Pilot – a leading indicator for biometric data and tool for reducing infection spread

by Dana Lewis

What if you could stop the spread of infection once it arrives at your house? This is easier to do now than ever before in the era of wearables like smart watches and smart rings that gather biometric data every day and help you learn what your baseline is, making it easier to spot when your data starts to deviate from baseline. This data often is the first indicator that you have an infection, even before you develop symptoms sometimes.

But, you need to know your baseline in order to be able to tell that your data is trending away from that. And despite the fact that these new devices making it easier to have the data, they actually make it surprisingly challenging to *quickly* get the new data on a daily basis, and the software they have programmed to catch trends usually only catches *lagging* trends of changes from baseline. That’s not fast enough (on either of those metrics) for me.

(The other situation that prompted this is companies changing who gets access to which form of data and beginning to paywall long-time customers out of previous data they had access to. Ahem, looking at you, Oura.)

Anyway, based on past experiences (Scott getting RSV at Thanksgiving 2024 followed by another virus at Christmas 2024) we have some rich data on how certain metrics like heart rate (RH), heart rate variability (HRV), and respiratory rate (RR) can together give a much earlier indication that an infection might be brewing, and take action as a result. (Through a lot of hard work detailed here, I did not get either of those two infections that Scott got.) But we realized from those experiences that the lagging indicators and the barriers in the way of easily seeing the data on a daily basis made this harder to do.

Part of the problem is Apple and their design of “Vitals” in Apple Health. They have a “Vitals” feature that shows you ‘vital’ health metrics in terms of what is typical for you, so you can see if you fall in the range of normal or on the higher or lower end. Cool, this is roughly useful for that purpose, even though their alerts feature usually lag the actual meaningful changes by several days (usually by the time you are fully obviously symptomatic). But you can’t reliably get the data to show up inside the Vitals section, even if the raw data is there in Apple Health that day! Refreshing or opening or closing doesn’t reliably force it into that view. Annoying, especially because the data is already there inside of Apple Health.

This friction is what finally prompted me to try to design something: if the data is there inside Apple Health and Apple won’t reliably show it, maybe I could create an app that could show this data, relative to baseline, and do so more reliably and quickly than the “Vitals” feature?

Turns out yes, yes you can, and I did this with BaselinePilot.

BaselinePilot is an iOS app that pulls in the HealthKit data automatically when you open it, and shows you that day’s data. If you have a wearable that pushes any of these variables (heart rate, heart rate variability, respiratory rate, temperature, blood oxygen) into Apple Health, then that data gets pulled into BaselinePilot. If you don’t have those variables, they don’t show up. For example, I have wearables that push everything but body temp into Apple Health. (I have it on a different device but I don’t allow that device to write to Health). Whereas Scott has all of those variables pushed to Health, so his pulls in and displays all 5 metrics compared to the 4 that I have.

What’s the point of BaselinePilot? Well, instead of having to open Apple Health and click and view every one of those individual metrics to see the raw data and compare it to previous data (and do it across every metric, since Vitals won’t reliably show it), now at a glance I can see all of these variables raw data *and* the standard deviation from my baseline. It has color coding for how big the difference is and settings to flag when there is a single metric that is very far off my baseline or multiple variables that are moderately off the baseline, to help me make sure I pay attention to those. It’s also easy to see the last few days or scroll and continue to see the last while, and I can also pull in historical data in batches and build as much history as I want (e.g. hundreds of days to years: whatever amount of data I have stored in Apple Health).

So for me alone, it’s valuable as a quick glance “fetch my data and make sure nothing is wonky that I need to pay attention to or tell someone about”. But the other valuable part is the partner sharing feature I built in.

With partner sharing, I can tap a button (it shows up as a reminder after your data is synced) and text a file over to Scott (or vice versa). Opening the file in iMessage shows a “open in BaselinePilot” button at the bottom, and it immediately opens the app and syncs the data. You can assign a display name to your person and thus see their data, too, in the same format as yours. You can hide or delete this person or re-show them at any time. This is useful for if you are going on a long vacation and sharing a house with family members, for example, and so you want to share/see your data when you’re going to be in physical proximity but then don’t need to see them after that – you can hide them from view until that use case pops up again.

Screenshots of BaselinePilot, showing 72 days of simulated data for the user (not actually my data) and the display with a simulated partner called "Joe".
Simulated data and simulated partner data displayed in BaselinePilot.

Ideally, I’d love to automate this data syncing with partners (who agree) over Bluetooth, so you don’t have to tap to share the file on a regular basis. But, that won’t work with the current design of phones and the ability to background sync automatically without opening the app, so I’ve stopped working on Bluetooth-based solutions given the technical/policy constraints in the phone ecosystems right now. Eventually, this could also work cross-platform where someone could generate the same style file off of their Android-based BaselinePilot and be able to share back and forth, but Scott and I both use iPhones so right now this is an iOS app. (Like BookPilot, I built this for myself/our use case and didn’t intend to distribute it; but if this sounds like something you’d use let me know and I could push BaselinePilot to the app store for other people to use.)

BaselinePilot: a leading indicator for biometric data and tool for reducing infection spread. A blog post by Dana M. Lewis on DIYPS.orgAll of this data is local to the app and not being shared via a server or anywhere else. It makes it quick and easy to see this data and easier to spot changes from your normal, for whatever normal is for you. It makes it easy to share with a designated person who you might be interacting with regularly in person or living with, to make it easier to facilitate interventions as needed with major deviations. In general, I am a big fan of being able to see my data and the deviations from baseline for all kinds of reasons. It helps me understand my recovery status from big endurance activities and see when I’ve returned to baseline from that, too. Plus the spotting of infections earlier and preventing spread, so fewer people get sick during infection season. There’s all kinds of reasons someone might use this, either to quickly see their own data (the Vitals access problem) or being able to share it with someone else, and I love how it’s becoming easier and easier to whip up custom software to solve these data access or display ‘problems’ rather than just stewing about how the standard design is blocking us from solving these issues!

What else have I built that you might like to check out? I just built “BookPilot”, a tool to help you filter book recommendations by authors you’ve already read, based on the list of books you’ve already read from your library data. If you use iOS, check out Carb Pilot to help get AI-generated estimates (or enter manual data, if you know it) to track carbs or protein etc and only see which macronutrients you want to see. If you have EPI (exocrine pancreatic insufficiency, also called PEI), check out “PERT Pilot” either on iOS or Android to help you track your pancreatic enzyme replacement therapy.

BookPilot: What Should I Read Next?

by Dana Lewis

Book hangovers are the worst. By book hangover, I mean that feeling you get at the end of a book when you come up for air and don’t know what you’re going to read next. In general, or because you have to ‘world switch’ from one type of book or one genre or one era to another (my most common cause), or because you don’t have any books picked out to read next.

I get this feeling, a lot, because I read a lot (hundreds of books per year). I also listen to a lot of audiobooks when I’m out doing long distance activities that take several hours (e.g. ultrarunning, hiking, or cross country skiing). So while I read 5x more books than I listen to audiobooks, I still consume several audiobooks per month, too, but I’m picky about what I like to read versus what I like for audiobook content. I regularly and periodically peruse catalog additions to my various* libraries’ digital collections under genres I like to read and under audiobook categories, putting things on hold so I always have a pipeline of things ready to read or listen to. But, this relies on me periodically doing so, and sometimes the influx of new things to read based on those effort dries up and I have fewer books on hand than I would like, especially when I know I’m going to be reading more (e.g. I have some downtime from not feeling well and will be reading more than normal or when I’m traveling/on vacation).

It occurred to me the other day that it would be nice to be able to automatically check authors I have read for new books and to be able to put those books on hold. The problem is that I read a lot. I have dozens if not hundreds of authors I like, and while I remember a few to check, I don’t have a good way to do this systematically. My brain finally put 2 and 2 together this week when it occurred to me that if Libby (the e-book/reading app now used by most of my libraries) allowed me to export my data, I could systematically list and look up authors and check for recent/new additions to their catalog. Ooooh.

Like most things now, there are usually ways to export or access all sorts of data whenever you’re ready to use it. Libby actually now makes it easy in the app to do this, without having to request or log in to the web portal or do any obnoxious behaviors. Tap Shelf>Timeline>look at “Actions” button in the top right. It gives you not one but THREE ways to export your data, wahoo! If I have the data, what could I do with it?

(TL;DR the rest of this: in less than a day, I built a tool that I call BookPilot which allows me to take the Libby data, parse it for the books & authors I have read, check those authors for every book they’ve ever written, then display it for me. BookPilot solves the ‘what should I read next?’ problem by systematically finding all unread books from authors I have already read. I can then easily mark off books I’ve read elsewhere (e.g. read before my Libby history started or in physical format), kick out books I know I don’t want to read, and otherwise have an awesome list of books ready to go the next time I want to spend some time checking out books and/or putting more things on hold! It works really well and that’s just on the authors I’ve already read – I plan to also add recommendations for new authors that match my favorite authors, genres, etc. There is a web dashboard where you can view and interact with the recommendations as well as the ability to work with this data via the command line.)

The whole goal was to help fuel my book pipeline and to be able to recommend books that are highly probable that I want to read, based on my past reads, but to automatically filter by what I’ve already read. (That’s a problem with asking chatbots or using other tools: it’ll show you ‘books like this’ but when you read hundreds per year, you don’t remember the covers or names of all the books you’ve read so you spend a lot of time re-checking things you’ve in fact already read.)

I started by taking the Libby data export (you can get it as CSV, json, or html) and having Cursor write a script to pull the book title and author for everything I’ve read in my export. This is about 800 books and ~350 or so authors from the last 2.75 years (when I switched from Overdrive to Libby and it started recording my history; before I wasn’t having it log my reads). It then takes every single author on the list, queries the Open Library API to find the author’s Open Library ID, then fetches all their works (up to 100 books per author at a time) using the author’s works endpoint; it then supplements this data with Google Books API queries. All API responses are cached locally to avoid rate limits, and requests include built-in rate limiting delays between calls. It stores all books by each author with metadata: title, ISBN, publication date, series information, categories/genres, format availability. It then cross-references catalog books with my reading history to mark read/unread status, attempts to filter out duplicates and non-English versions (e.g. “Anne of Green Gables” and “Anne of Green Gables (German Edition)” versus Anne of Green Gables in the actual german language title and also “Anne of Green Gables / Anne of Avonlea (Box Set)” etc). It takes a while if you have hundreds of authors but the script is set up so you can start/stop it and it will pick up where you left off. The next time you check, it will check and only add books that are more recently published (customizable thresholds, of course) so future catalog checks will be quick to look for recent additions.

Then I have a dashboard that sorts and shows me all these books, ordered by series (when series info is available) so I can see a) series I have not read from authors I have already read and b) any books in a series that I have partially read; I can also sort by author’s name or highest count of the books I haven’t read, e.g. maybe I’ve read one book from an author but they have another 30 books to consider!

(On first pass, I still need to go through and mark books that I have read from >=3 years ago digitally and/or in physical form. But a quick look has already shown that this system works great and I’ve already been able to instantly add a bunch of books to my to-read list and holds already!)

I added a thumbs up feature which automatically ports a book to my “books to read” tab, which answers my “what should I read next?” question. If I thumbs down a book, it goes away from the dashboard. I can also tag “already read” (also disappears it), “not in English” (e.g. my filters missed that this is a non-English version on the first pass) or “duplicate” for duplicate titles that have somehow crept into the listing.

I have separate recommendation lists for authors I’ve listened to (audiobooks) versus authors I’ve read (ebooks), so I can also have a good queue to fill for similar audiobook content.

I have a ton (hundreds of audiobooks, thousands of ebooks) to review, based on my ~800 books read in the last <3 years, from 350+ distinct authors. It’s delightfully overwhelming to have this many recommendations! There is nothing worse than a book hangover combined with having nothing in the reading queue and having to fight the book hangover and having to simultaneously search for what’s available and what you want to read next. (It’s like feeling hangry…book hangover hangry?) This system now means I won’t have that combination problem again: I will always have books lined up to read and if not I can quickly and more easily identify off this list from my known authors where they’ve added books or they have additional series or I can (eventually) find series from the same “era”/”genre”/”world” to continue reading and pick up where my brain left off.

Book Pilot series recommendation of books you have not read from authors you have already read Book Pilot dashboard recommendation of books you have not read from authors you have already read.

Eventually, I’m going to add some recommendations for new authors that ‘match’ my most-read authors in different categories. But I don’t have to, yet, because I have so many high-ROI options based on my current author list! And the next time I drop in an updated Libby export, it will add the new authors I’ve read *and* automatically remove books from my “books to read” list that appear in my reading history.

BookPilot: answering 'what should I read next' by filtering what I have already read. A blog post from Dana M. Lewis on DIYPS.orgI haven’t open-sourced BookPilot yet, but I can (UPDATE: it’s now open source!)) – if this sounds like something you’d like to use, let me know and I can put it on Github for others to use. (You’d be able to download it and run it on the command line and/or in your browser like a website, and drop your export of Libby data into the folder for it to use). (And did I use AI to help build this? Yes. Could you one-shot a duplicate of this yourself? Maybe, or otherwise yes you could in several hours replicate this on your own. In fact, it would be a great project to try yourself – then you could design the interface YOU prefer and make it look exactly how you want and optimize for the features you care about!)

* Pro tip: if you live in Washington state, most county libraries have “reciprocal” agreements if their counties offer services to each other. This means that because I lived in Seattle and have a SPL card and a King County (KCLS) card, I can also get Pierce County, Sno-Isle, etc etc etc…. It’s mainly digital collection access, but this is amazing because you can have all of these library cards added to your Libby account and when you want to go get a book, you can check ALL of your catalogs. Sometimes a book is on most of these systems and I can get on the hold list for the one with the shortest wait time. Other times, it’s only in ONE catalog and I wouldn’t have been able to get it without these reciprocal county cards because it wasn’t in SPL or KCLS (my primary local libraries)! And even when books are not in any collection yet you can add a smart tag and be notified automatically whenever something IS added to *any* of your libraries. It’s amazing. Not sure if other states or areas have these types of setups but you should look to see if you can access other catalogs/get cards, and if you live in Washington and love to read, definitely do this!

What else have I built that you might like to check out? If you use iOS, check out Carb Pilot to help get AI-generated estimates (or enter manual data, if you know it) to track carbs or protein etc and only see which macronutrients you want to see. If you have EPI (exocrine pancreatic insufficiency, also called PEI), check out “PERT Pilot” either on iOS or Android to help you track your pancreatic enzyme replacement therapy.

What you shouldn’t take away from my talk about patient experiences of using AI in clinical healthcare

by Dana Lewis

I was asked to contribute a talk in a session at Stanford’s recent AI+Health conference. I spoke alongside two clinician researchers, and we organized the session so that I would talk about some individual patient experiences & perspectives on AI use in health, then scale up to talk about health system use, then global perspectives on AI for health-related use cases.

One thing I’ve been speaking about lately (including when I was asked to present in DC at the FDA AI workshop) is about how AI…is not one thing. There are different technologies, different models, and they’re going to work differently based on the prompts (like our research showed here about chart notes + LLM responses, as evaluated by a patient and clinician) as well as whether it’s a one-off use, a recurring use, something that’s well-defined in the literature and training materials or whether it’s not well-defined. I find myself stumbling into the latter areas quite a bit, and have a knack of finding ways to deal with this stuff, so I’m probably more attuned to these spaces than most: everything from figuring out how we people with diabetes can automate our own insulin delivery (because it was not commercially available for years); to figuring out how to titrate my own pancreatic enzyme replacement therapy and building an app to help others track and figure things out themselves; to more recent experiences with titration and finding a mechanistic model for an undiagnosed disease. The grey area, the bleeding edge, no-(wo)man’s land where no one has answers or ideas of what to do next…I hang out there a lot.

Because of these experiences, I see a lot of human AND technology errors in these spaces. I see humans make mistakes in chart notes and also in verbal comments or directions to patients. I see missed diagnoses (or wrong diagnoses) because the medical literature is like a game of telephone and the awareness of what conditions are linked to other conditions is wrong. I see LLMs make mistakes. But I see so many human mistakes in healthcare. One example from my recent personal experiences – it’s 2025 and a clinical team member asked me if I had tried cinnamon to manage my glucose levels. This was after she had done my intake and confirmed that I had type 1 diabetes. I looked at her and said no, because that would kill me. She looked surprised, then abashed when I followed that comment that I have to take insulin because I have type 1 diabetes. So I am maybe less bothered than the average person by the idea that LLMs sometimes make mistakes, say a wrong (or not quite right, even if it’s not wrong) thing, or don’t access the latest evidence base. They can, when prompted – and so can the human clinical teams.

A big point I’ve been bringing up in these talks is that we need everyone to care about the status quo of human, manual healthcare that is already riddled with errors. We need everyone to care about net risk reduction of errors and problems, not focus on additive risk.

We saw this with automated insulin delivery – people without diabetes were prone to focus on the additive risk of the technology and compare it to the zero risk they face as a person without diabetes, rather than correctly looking at the net risk assessment of how automated insulin delivery lowers the risk of living with insulin-managed diabetes even though yes, there is some additive risk. But almost a decade later, AID is now the leading therapy choice for people with insulin-managed diabetes who want it, and OS-AID is right there in the standards of care in 2026 (and has been for years!!!) with a strong evidence base.

I also talked about my experiences observing the real use of clinical scribe tools. I talked more in detail about it in this blog post, but I summarized it in my talk at Stanford, and pointed out how I was surprised to learn later – inadvertently, rather than in the consent process – that the scribe tool had access to my full medical record. It was not just transcribing the conversation and generating chart notes from it.

I also pointed out that my health system has never asked me for feedback about the tool, and that I’ve actually seen my same clinician use multiple different scribe technologies, but with no different consent process and no disclosure about chart access or any differences in privacy policy, retention timeline, etc. (Don’t pick on my clinician: she’s great. This is a point about the broader systematic failures.)

This is where I realized later people might have taken away the wrong point. This was not about me being a whiny patient, “oh they didn’t ask for my feedback! My feedback is so important!” and self-centered.

It was about flagging that these technologies do not have embedded feedback loops from ALL stakeholders, and this is more critical as we roll out more of these technologies related to healthcare.

It’s one thing to do studies and talk about user perspectives and concerns about this technology (great, and the other two presenters did an awesome job talking about their work in these spaces).

But we need to do more. We need to have pathways built in so all stakeholders – all care team members; patients; caregivers/loved ones; etc. – have pathways to talk about what is working and what is not working, from everything from errors/hallucinations in the charts to the informed consent process and how it’s actually being implemented in practice.

It matters where these pathways are implemented. The reason I haven’t forced feedback into my health system is two-fold. Yes, I have agency and the ability to give feedback when asked. Because I’ve worked at a health system before, I’m aware there’s a clinic manager and people I could go find to talk to about this. But I have bigger problems I’m dealing with (and thus limited time). And I’m not bringing it up to my clinician because we have more important things to do with our time together AND because I’m cognizant of our relationship and the power dynamics.

What do I mean? The clinician in question is new-ish to me. I’ve only been seeing her for less than two years, and I’ve been carefully building a relationship with her. Both because I’m not quite a typical patient, and because I’m not dealing with a typical situation, I’m still seeking a lot from her and her support (she’s been great, yay) to manage things while we try to figure out what’s going on. And then ditto from additional specialists, who I’m also trying to build relationships with individually, and then think about how their perceptions of me interplay with how/when they work with each other to work on my case, etc.

(Should I have to think about all this? Do I think about it too much? Maybe. But I think there’s a non-zero chance a lot of my thinking about this and how I’m coming across in my messages and appointments have played a role in the micro-successes I’ve had in the face of a rubbish and progressive health situation that no human or LLM has answers for.)

So sure, I could force my feedback into the system, but I shouldn’t have to, and I definitely shouldn’t have to be doing the risk-reward calculation on feedback directly to my clinician about this. It’s not her fault/problem/role and I don’t expect it to be: thus, what I want people to take away from my talk and this post is that I’m expecting system-level fixes and approaches that do not put more stress or risk on the patient-provider relationship.

The other thing I brought up is a question I left with everyone, which I would love to spur more discussion on. In my individual case, I have an undiagnosed situation. A team of specialists (and yes, second opinions) have not been able to diagnose/name/characterize it. I have shifted to asking providers pointedly to step away from naming and thinking broadly: what is their mental model for the mechanism of what’s going on? That’s hard to answer, because providers aren’t often thinking that way. But in this situation, when everything has been ruled out but there is CLEARLY something going on, it’s the correct thing to do.

And for the last year, the LLMs had a hard time doing it, too. Because they’re trained on the literature and the human-driven approach to make differential diagnoses, the LLMs have struggled with this. I recently began asking very specifically to work on mechanistic models. I then used that mechanistic model to frame follow up questions and discussions, to rule things in/out, to figure out where there are slight overlaps, to see where that gives us clues or evidence for/against our mechanistic model hypothesis, and to see what treatment success / failure / everything in between tells us about what is going on. Sure, a name and a diagnosis would be nice, but it’s been so relieving to have at least a hypothetical mechanistic model to use and work from. And it took specific probing (and the latest thinking/reasoning models that are now commonly available). But why am I having to do it, and why are my clinicians not doing this?

Do clinicians have an obligation to disclose when they are not using AI?I know some of the answers to the question of why clinicians aren’t doing this. But, the question I asked the Stanford AI+Health audience was to consider why we focus so much on informed consent for taking action, but we ignore the risks and negative outcomes that occur when not taking action.

And specifically in rare/undiagnosed diseases or edge cases of known diseases…do clinical providers have an obligation now to disclose when they are NOT using AI technology to facilitate the care they are providing?

It’s an interesting question, and one I would love to keep discussing. If you have thoughts, please share them!

Vein Pilot: a solution for finding your veins so you can NOT stick them

by Dana Lewis

Introducing Vein Pilot, a solution for finding your veins so you can NOT stick them.

Wait, what?

Most people who want to find their veins do so in order to administer medications that are intravenous (in their veins). I remember getting an infusion and having a nurse who preferred using a vein finder light. It was fancy: it displayed a red light on a rectangle of skin and the veins showed up as darkly colored, so it was very easy to assess them.

Fast forward to more recently when I’m trying to do a home subcutaneous infusion, with several needle sites, every week. At one point, I had inserted a needle and then realized blood was flowing up the tube. Oops, hit a vein. I had to pull that site out and not use it for administering the medication. (And, as someone who is a delightful bruiser, I had quite the bruise for weeks).

This made me nervous knowing that could happen again, because my needles/sites are split from the same set of tubing, and if I hit a vein and have to pull one out, I have wasted that needle/site and it then messes up how I would do the rest of my infusion. It would require additional time and using a full other set of infusion needles, since I can only administer so much per site and dropping one site would put the others over the limit. That means then I have to figure out how to split the cartridges across the remaining sites and keep a cartridge or two to split over another set of sites. This costs an extra hour of time, a backup set of sites (which luckily I have but not everyone has backups) AND causes additional ‘lump’ risk post-infusion (because every needle causes trauma to the tissues that about a third of the time leaves a loose or consolidated lump that can take anywhere from one to a bunch of weeks to go away.) So this nervousness isn’t “I have to stick another needle in!” but a more complex, nested “I have to rethink and redesign this infusion and it’s non-optimal, time-consuming, and has long-term post-infusion impacts, too”. It’s a risk for every single needle which is multiple per week, every week, for the foreseeable future. (Although it’s only happened 1/88 so far…)

As a result, I became very cautious about trying to see where the veins were in the areas I was trying to infuse in, but this is challenging. I have limited areas to infuse in, because I can’t reach some areas that are allowed – and also because I have to juggle real estate for my insulin pump, which also infuses into subcutaneous tissue, and CGM. This leaves my outer thighs, my inner thighs, and my abdomen as possible infusion sites that I can actually reach for these bigger volume infusions. And…these thigh/abdomen zones have a lot of veins. It’s not always easy to see them, depending on the light or how I’m positioned, but I definitely want to avoid them.

I started venting to myself the other day about the situation – limited real estate, lean areas that are hard to infuse in that further limit where on the thighs and abdomen I can comfortably infuse in, plus the veins limit the real estate further and cause additional stress at the time I’m going to insert the sites. Ugh! But, one thing I’ve gotten progressively better at is trying to categorize which of the things that are bothering me are modifiable. Maybe not “fixable” completely, but could they be modified to be a little bit better? I have a lot of experience now doing this with software solutions (PERT Pilot, MacrosOnTheRun, spreadsheets galore, Baseline Pilot, Book Pilot, etc. – many of my ‘software-shaped feelings’) but I also increasingly, thanks to LLMs, have started to consider physical, hardware solutions for things, too.

It’s not my default skillset, so I don’t always think about it, but in this scenario I remembered and was thinking through the light that the nurse used. I looked it up, thinking maybe I could buy one (or insurance might rent one). LOL, nope: these things are $2000+. But I looked into how they work – by shining light, identifying the contrast, then displaying a dark ‘map’ of the veins back. That helped me realize that the consumer grade “solutions” that are shining a light aren’t doing quite the same thing. I also looked at the reviews for the consumer grade options, and they are terrible for the $200 devices. Plus the shape didn’t seem like it would be usable for my use case. I tried to see if anyone had made DIY versions that were better, and there were a few 3D printed type devices that you add light strips into, and that made me think – could I try to build one?

Same lights, different purpose.

I ended up brainstorming something more simple than the ‘vein finder’ DIY projects. Their goal is to find the veins to stick them directly – I want to avoid them! But maybe the lights could be used similarly.

Instead of a “circle” (or open circle) which most of them are designed around, I decided I wanted to try a simple ‘fork’ or a rectangle (with light strips on two sides and the ends blocked) to position light pointing inward to a spot, to try to illuminate the tissue so I could see the veins more clearly than with bare eyes and room light. My LLM brainstorming partners had suggested using a combination of red + amber lights. I decided I would buy cheap light strips (you can find them in the range of $5-20 depending on length, color, etc.) and see if they worked, then iterate on a 3D printed type ‘fork’ or rectangle to use to help limit light leakage away from the area I was trying to illuminate. But I never ended up finishing iterating because…

I bought red light strips ($5 set) and amber (orange, $13 set) light strips. I got them and plugged them in to the 12-volt dimmer switch we happened to have, plus some light strip LED strip-to-wire connectors. (We originally got and had these connectors when we put up bright white LED light strips in our closet to add closet lighting. Having these on hand made it so I only had to buy the new colored light strips to prototype this idea, otherwise I would’ve bought the dimmer and strip-to-wire connectors.)

I first took the red light and put it against my leg. It didn’t really help, I couldn’t see the veins well. I couldn’t tell if that’s because I didn’t have it set up in the intended fork design – a light on each side, slightly angled inward, with tape on the outside to limit light leakage – or if the red light was non-optimal for lighting up my skin or if it’s the particular design of the light strip I had gotten.

I tried the amber (orange) light next, expecting it also to not work for the same reasons but, as I moved it across my leg, I SAW THE VEINS! They were dark. It is actually the *motion* of the light that helps, because my eyes then ‘see’ the dark vein based on the light moving, so my eyes pick up on the dark ‘line’ of the vein on either side of the light strip as it moves. By moving it over an area in different directions, I can pick out a vein, then move the light and watch for other veins nearby.

Sometimes the veins fork and I can see a good spot in the “V” of two veins, or similarly they look like they’re running in parallel to each other and I can see where there’s a good spot in between. I dim or turn off the light, holding my finger in the ‘good’ zone and look on my bare skin for freckles or marks to help keep track of this location. (I could mark it with a skin-safe marker, but I have to scrub the skin with an alcohol swab before infusing and I assume that would remove the mark, so I need other landmarks to use regardless.) I can then swab/scrub it, wait for it to dry, then place the needle.

I tried it for the first time this week. I had chosen two abdomen sites and two inner thigh sites. I didn’t actually use it with both abdomen sites, because I tried the light strip with the first side of my abdomen and didn’t see anything. That could be because the light strip is currently still too long (I need to cut and shorten it for easier handling); because of the angle of my view down on my abdomen I can’t really see both sides of the light against the skin – I only see one side; or because there weren’t any surface level veins right there. So I didn’t try with my other abdomen side, either. But I did use it with both of my inner thigh sites and it was so much easier and less stress to use it to pick an area I thought I wanted to use (generally); check the area for the nearest veins, then narrow down the exact spot I wanted to target and use.

Woohoo, less stress!

Vein Pilot: a solution for finding your veins so you can NOT stick themIf anyone ran into this post because you are looking for finding your veins so you can stick them – it would work for that, too. You’d probably want to play around with an actual 3D printed (or even popsicle sticks) structure for the lights so you could try to keep the lights in the spot and keep the vein highlighted. In my use case, I am trying to AVOID hitting the veins and don’t need to keep it highlighted beyond figuring out where they are and picking my spot. (Then I move the lights away so I can clean the site and insert, which I couldn’t do with anything within 3-4 inches of the site, because of how I hold my hand to insert this needle at a 90 degree angle – it’s different from IV access angles.) You may get benefit out of using different combinations of colored lights and orienting them differently than I do.

But if you’re like me, want to AVOID veins, an amber/orange light strip and moving it around may be good enough for this purpose. You could iterate on different shapes/structures to hold the lights, but since a single amber light strip and the motion of it works for my use case, I haven’t iterated (yet) on a better design that would improve it.