RT @NeurIPSConf: NeurIPS is pleased to officially endorse EurIPS, an independently-organized meeting taking place in Copenhagen this year,….

RT @SymposiumML4H: 🚨 Machine Learning for Health (ML4H) is back and better than ever!.🌴 Join us in San Diego on December 1–2, 2025, right b….

Join millions who have switched to Grok.

RT @Lucas_Arnoldt: Current multimodal single-cell integration methods act as black boxes, lacking meaningful interpretability. We introduce….

RT @Lucas_Arnoldt: NetworkVI is a group effort by Julius Upmeier zu Belzen, Luis Herrmann, Khue Nguyen, supervised by @fabian_theis & @nebw….

RT @EricTopol: How does @deepseek_ai compare with other LLMs for clinical diagnosis and treatment plan—125 patient case benchmarks, not rea….

RT @Lucas_Arnoldt: Thrilled to share our @NatureBiotech paper “Multistate and functional protein design using RoseTTAFold sequence space di….

RT @stefanhgm: Does removing unsupported facts in the training or prompting data effectively reduce hallucinations?. We tested this for GPT….

Amazing work from @Youssef_M_Nader!.

RT @natfriedman: Ten months ago, we launched the Vesuvius Challenge to solve the ancient problem of the Herculaneum Papyri, a library of sc….

RT @SteinfeldtJakob: Over the past few months, @gogothorr and I have been heads down building @PheironInc . Our platform extracts disease m….

RT @tlandgraf: Echo chambers in the honeybee dance communication! We automatically detected 100K waggle phases of dancing bees and found a….

RT @tomlincr: 💡Diffsurv: Differentiable sorting for censored time-to-event data.👏Great work from @cdt_ai_health @UCL_IHI colleagues @AndreV….

@SteinfeldtJakob @gogothorr @pietznerm @juluzb @AndreVauvelle @stefanhgm @SpirosDenaxas @profhhemingway @CaptainSysBio For another perspective on this work, see the thread from @SteinfeldtJakob here:

Thanks to all colleagues and collaborators involved in this endeavor, @SteinfeldtJakob, @gogothorr, @pietznerm, @juluzb, @AndreVauvelle, @stefanhgm, @SpirosDenaxas, @profhhemingway, Claudia Langenberg, Ulf Landmesser, John Deanfield, and @CaptainSysBio!.

In summary, we can predict disease onset based on medical history for ~ 1.900 endpoints at once, without additional measurements required. The approach generalizes across healthcare systems and populations and is competitive with selected disease-specific risk scores.

Absolute risk predictions are likely underestimated due to the UK Biobank cohort being healthier than the general population. If routine health records are to be used, robust governance rules to protect individuals, such as opt-out and usage reports, must be implemented.

There are still important challenges: Health records are subject to biological, procedural, and socio-economic biases, and not all diagnoses are captured explicitly in the records but can be inferred e.g. from medications; thus, discriminative performance can only be approximated.

For illustration, we also present individual phenome-wide risk profiles: Predisposition (Estimated 10-year risk relative to a risk model based on age and sex only) is displayed in the inner circle, and absolute estimated 10-year risk in the outer circle.

We then show that predictive models can generalize across healthcare systems and populations, including communities historically underrepresented in biomedical research. We find discriminative improvements for 1.310 (83.5%) endpoints without retraining in the US All of Us cohort.

We directly compare the model to established risk scores for a subset of cardiovascular diseases. The medical history is competitive to SCORE2, ASCVD, and QRISK3 for cardiovascular diseases, but at no additional effort and cost (no measurements and lab tests required).