iiinova — where your digital self keeps evolving

David Burstein @davidbphd.bsky.social · 10/08/2026

Hi Monica, I’m David Burstein, an assistant professor at Mount Sinai working in psychiatric genetics, GWAS/PRS, and EHR data science. I’d love to contribute to the Science Feed. Here’s my Google Scholar profile. Thanks!

scholar.google.com

120

David Burstein @davidbphd.bsky.social · 30/07/2026

This is not theoretical. For-profit AI detector companies often sell subscriptions, API credits, or bulk checks. If users can buy repeated access, they can optimize against the detector. So the question is not just one-shot accuracy. It is whether detection survives adaptive use.

020

David Burstein @davidbphd.bsky.social · 30/07/2026

Is AI detection broken? The key limitation is adaptivity. If someone can see a detector score, ask an LLM to revise, and try again, millions of times if needed, the detector becomes feedback. A user can effectively search their way to a false negative. #AcademicSky #AI #AcademicIntegrity

110

David Burstein @davidbphd.bsky.social · 30/07/2026

Really interesting paper. I was struck by the ADHD association with dissolution, especially given the null for autism. Would adding parental PGIs for partnership dissolution as prognostic covariates, beyond the trait-matched parental PGIs, help improve precision?

000

David Burstein @davidbphd.bsky.social · 29/07/2026

🧬 Helpful recent preprint describing the All of Us “All by All” resource: common and rare variant association testing across 392,030 whole genomes and 3,602 phenotypes Preprint: www.medrxiv.org/content/10.6... Browser: allbyall.researchallofus.org

medrxiv.org

000

David Burstein @davidbphd.bsky.social · 24/07/2026

Hi Dani, I am a data scientist. If you could add me to the science feed, that would be great. Google Scholar: scholar.google.com/citations?us...

scholar.google.com

000

David Burstein @davidbphd.bsky.social · 24/07/2026

I don't know yet. Discerning meaningful discussion is a separate and more sophisticated analysis. Let me look into it and get back to you.

000

David Burstein @davidbphd.bsky.social · 22/07/2026

Thanks @altmetric.com! @rbly.bsky.social, hopefully LinkedIn will bring back its API. Until then, no idea. My hypothesis: LinkedIn's growth roughly tracks the decline in X, so its share is probably substantial. I'd be curious, but a little surprised, if it made X and Bluesky pale in comparison

120

David Burstein @davidbphd.bsky.social · 22/07/2026

Limitations: the counts extracted do not discern between the type of account (institutional, scientist, etc.) or uniqueness. Interesting blog post from last year relevant to the discussion where Altmetric inquired about the status of science on X and Bluesky, www.altmetric.com/blog/bluesky...

altmetric.com

Bluesky’s ahead, but is X a dead parrot?

Altmetric has been capturing and integrating Bluesky posts since October 2024. What does the data look like now?

011

David Burstein @davidbphd.bsky.social · 22/07/2026

So what does this mean? Science discussion is spreading across more platforms, not consolidating on one. No single site appears to hold the whole conversation, and the balance shifts by field/journal. Relying on only one platform may limit both your reach and what you see.

100

David Burstein @davidbphd.bsky.social · 22/07/2026

The 29% reflects the sample. Across the seven journals, Bluesky's median share ran from 14% (Nature Medicine) to 50% (Nature Machine Intelligence and npj Digital Medicine). So there is no single number for "science." Where your field's conversation lives may be specific to your specialty.

110

David Burstein @davidbphd.bsky.social · 22/07/2026

Where are scientists discussing papers now: Bluesky or X? I sampled 350 recent papers that had at least one post on X or Bluesky. The median paper got 29% of its X+Bluesky attention on Bluesky. Bluesky is substantial, but X appears active. #AcademicSky #ScienceSky #StatsSky

Visualization of the median per-paper Bluesky share of combined X and Bluesky attention for seven Nature-family journals (350 papers, 50 per journal, each with at least one X or Bluesky post), January to July 2026. Each dot is the journal's median; the bar shows the middle 50% of papers. The overall median is 29%. By journal: Nature Medicine 14%, Nature Genetics 25%, Nature 26%, Nature Communications 26%, Molecular Psychiatry 45%, npj Digital Medicine 50%, Nature Machine Intelligence 50%.

2124

David Burstein @davidbphd.bsky.social · 21/07/2026

This reminds me of doi.org/10.1111/sltb.... Very interesting to think about these potential health trajectories for interrelated traits, especially when “self-harm” can cover such different behaviors and relationships to suicidality.

doi.org

010

David Burstein @davidbphd.bsky.social · 19/07/2026

They just had to spoil the ending before I finished reading it...

010

David Burstein @davidbphd.bsky.social · 17/07/2026

This is interesting. I’ve been thinking about cases where between-study heterogeneity varies for concrete measurement reasons, e.g. misclassification bias in studies using self-report data. Nice to see a general framework for testing when the constant-heterogeneity assumption is doing too much work.

010

David Burstein @davidbphd.bsky.social · 14/07/2026

Really promising direction. A colleague is leading a related study we've been working on, using LLMs to generate "future-self" narratives for patients with suicidal ideation, to help flag who's at risk: osf.io/preprints/ps...

osf.io

OSF

010

David Burstein @davidbphd.bsky.social · 14/07/2026

Nice. The hard part for community curation is cold start from a handful of researchers. Could you bootstrap early rankings from Bluesky engagement on paper posts (likes/reposts)? It is an existing community-preference proxy, and unlike most platforms the open API makes it easy to get the data.

210

David Burstein @davidbphd.bsky.social · 14/07/2026

Might be less a tooling gap than a permissions one: commercial agents refuse CAPTCHAs by design because solving them circumvents bot-detection/site terms of use. An open model doesn't really sidestep that; it just shifts the terms of use (and institutional-policy) exposure onto the researcher.

030

David Burstein @davidbphd.bsky.social · 13/07/2026

Over 292,000 public comments already submitted on the proposed OMB rule revising federal grants guidance. The comment period closes today, July 13. Read it & comment: , www.regulations.gov/document/OMB... #SciencePolicy #SciSky #AcademicSky

regulations.gov

Regulations.gov

001

David Burstein @davidbphd.bsky.social · 11/07/2026

Kudos to @pangram.com for quantifying something real. But I think this solves the wrong problem: the issue was never AI-assisted content, it's slop. Don't filter out thoughtful human-AI science collaboration. Instead of flagging AI-written posts, why not flag slop (human or AI)?

000

David Burstein @davidbphd.bsky.social · 10/07/2026

Defining phenotypes across 10,000s of ICD codes is manual & subjective. Phecoder semi-automates this, surfacing clinically relevant codes existing curation missed. Analysis led by Jamie Bennett. 📝 Preprint: www.medrxiv.org/content/10.6... 🖥️ Code: github.com/DiseaseNeuroGenomics/Phecoder

input phenotype query + ICD descriptions → embedding space → ranked ICD code output

050