I wrote up my approach in a blog post: bennettwaxse.com/blog/bioinfo...
If you're a researcher curious about Claude Code, here's where I start.
If you have a different workflow, I'd love to hear about it.
I wrote up my approach in a blog post: bennettwaxse.com/blog/bioinfo...
If you're a researcher curious about Claude Code, here's where I start.
If you have a different workflow, I'd love to hear about it.
Why bother? Because LLM context windows are big enough to hold an entire projectβso use them. Don't hide how you derived a cohort when that derivation can inform what you're asking the model to do next.
3/
- #CLAUDE.md files that orient Claude to your project's structure, conventions, and domain quirks
- Reference materials (data dictionaries, trusted queries), so the model knows what's actually in your tables
- .claudeignore to prevent reading secrets, redundant files, or data you shouldn't share
2/
I've been using #ClaudeCode for #research #informaticsβcohort building, debugging pipelines, genomics in #AllofUs. Before any of that was useful, I had to set things up right.
The setup isn't complicated, but it matters. Here's how I do it:
1/
Clearly a big fan of VZV vaccination as an ID doc, but I need to dig into this deeper before it makes its way into patient counseling for me. Curious what others took away from the sex-specific differences, and wonder if other types of analysis (e.g., TTE) might be better suited for this...
5/5
Are we looking at detection bias? Do shingles episodes trigger cognitive screenings in women more readily? Is there a systematic difference in healthcare contacts between sexes? And we still don't know the mechanism: is it non-specific immunomodulation over VZV reactivation reduction?
4/
First, this applies only to the live-attenuated vaccine, not Shingrix, which is now standard. Second (what's really nagging at me), the dementia protection really only appeared in women, even though the vaccine prevented shingles and post-herpetic neuralgia equally in both sexes (Table S3).
3/
The 20% relative reduction in dementia risk is striking, and the regression discontinuity design is cool--particularly in the presence of no discontinuity among education (S15-S17) or in placebo temporal tests across other years (S12-S14). But, I was still left with more questions than answers.
2/
What I've been reading: a natural experiment to estimate the herpes #zoster vaccine's effect on dementia--did you know that the effect was predominantly observed in females?
pmc.ncbi.nlm.nih.gov/articles/PMC...
1/
My daughters remind me daily why I'm lucky to call science work--not the grants or papers, but the pure pleasure of careful observation leading to discovery. Happy New Year, y'all.
4/4
Getting back to work, I found similar joy. Staring at wearable data 180 days before COVID infection, discovering that people who developed Long COVID were already different 6 months before they got sick. Watching genetic signals shift depending on how we define disease in the same biobank.
3/
Parental leave let me slow down and observe. Our daughter tracking me across the room for the first time. Tummy time shifting from torture to exploration. Our 3yo stringing together complex sentences about her day or her feelings. Capacities build invisibly, showing up in surprising ways.
2/
There's something mesmerizing about watching an infant discover their smile. The way it floods the cheeks, squints the eyes, shakes down into the chest. They draw their gaze away--almost like it's too much--before turning back to do it all again.
1/
The death of any child is tragic. We always think it will not be my child, they are healthy. Severe influenza can be unpredictable & deadly. Getting vaccinated considerably reduces dz severity and deaths. Please make sure you are vaccinated & your children are too.
people.com/alabama-boy-...
3. What did this look like in other states? There was a more muted response for a prior outbreak, but that was during a time of social isolation (COVID-19). Would love to see local vs. national effects.
Overall, super cool study. As an ID doc, Iβd love to get my hands on realtime EHR data!
5/5
2. The study included a breakdown by age, but how does this look across social determinants of health? The study is restricted to those with regular healthcare access, so what happens to families not as well-resourced? Curious what data Truveta have in this domain.
4/
1. An early dose doesn't count toward the official 2-dose series. So of course itβs too early now, but do parents here also get the doses at 12-15 months and 4-6 years that actually confer immunity?
3/
Early MMR vaccination (6-11 months of age) jumped from a 0.7% baseline to 20.1% during the outbreak. That's a massive behavioral shift.
What caught my attention most though is what else they could look at:
2/
What Iβm reading: a super cool real-time RWE EHR study by Truveta that assessed vaccination behavior during the recent Texas measles outbreak.
jamanetwork.com/journals/jam...
1/
Very curious to see how this project evolves in peer review, and whether that will occur before we start to see this model in action. I heard at a talk that theyβre already thinking about deployment.
Would love to hear from those with expertise in this domain.
6/6
These research decisions, including the omission of social drivers of health, all matter and should be evaluated in work that can be as profound as something integrated in the most common US EHR. Overall, it reads like the project prioritized demonstrating scale over clinical utility.
5/
The 90/10 random split and absence of statistical tests for comparison with task-specific baselines are interesting choices. No prospective validation (coding changes, prevalence shifts, new treatments). The margins over supervised baselines are small without CIs or multiple testing correction.
4/
Figure 23 also stood out. Performance is best in low-risk populations (younger, healthier patients) compared to higher-risk groups (elderly with higher rates of heart disease or COPD). That's backwards from clinical utility. Is it just good at predicting healthy outcomes in healthy people?
3/
But here's what caught me: the model omits vitals, and not for computational reasons. When I think of differential severity for a given condition (e.g., COPD exacerbation, or sepsis), what better way to ascertain severity? Are lab quantiles and meds really enough?
2/
What I'm Reading:
Curiosity, a 1B-parameter model built on Epic Cosmos.
It's genuinely interesting: zero-shot outputs rivaled supervised baselines across 78 real-world tasks, and the scaling laws (Ξ±=0.52) mirror NLP. Cool to see how information scales across domains.
arxiv.org/abs/2508.12104
1/
12/ As an ID fellow doing EHR research, this represents my path toward independence - using informatics to understand host-pathogen interactions.
Huge thanks to Josh Denny lab at NHGRI, #AllofUs Research Program, and www.niaid.nih.gov/about/infect... for making this possible! π
#MedSky #IDSky
11/ For the #IDSky #EpiSky #AcademicSky community:
What other episodic infectious diseases need computable phenotypes? UTIs? Pneumonia subtypes? C. diff?
I'm building expertise in ID EHR methods - reach out for collaborations! Always happy to help with phenotyping approaches in and out of ID.
10/ What's next from our NHGRI lab:
𧬠GWAS to identify genetic risk factors for severe respiratory infections
π Clinical determinants of hospitalization across viruses
π Diagnostic stewardship for multiplex respiratory panels
Phenotypes are just the beginning! #IDSky
9/ Important caveats (transparency matters):
Low sensitivity = can't estimate true prevalence
#AllofUs demographics β general US population
Oversamples severe disease, misses mild infections
Remember: Always validate your phenotypes
But, methods likely applicable to any US EHR dataset
8/ Want to use these phenotypes or create your own EHR methods? Let's collaborate! π€
β¨ Code available in #AllofUs Community Workspace
𧬠Perfect for GWAS, health disparities, clinical outcomes research
π Framework adaptable to other episodic infectious diseases
DM me - my passion is EHR phenotyping