what can a deeply phenotyped human cohort tell us, and what can today’s AI models do with that information? we built PhenoBench around the Human Phenotype Project: 90 tasks across 15 clinical domains, using clinical, imaging, molecular, and wearable data. having those tasks let us ask a few things: which measurements actually help? and do more capable models make better use of them? on these tasks, tabular foundation models performed better overall, but the gains over ridge regression were small. the LLM results felt like a useful example of the “jagged frontier”: they did relatively well on some tasks and poorly on others. models trained on the same input fields generally did better. btw, most of the work was figuring out what to ask and how to tell whether a model’s answer was any good. that took much longer than running the models. having done that, trying another model or measurement is a lot easier.

In our new preprint, we show that prognostic signals can be found in DXA scans acquired routinely to measure bone density and body composition Our model, LeDXA, is a JEPA-based vision model that learns by predicting latent representations rather than pixels, trained from scratch on 11,540 unlabeled Human Phenotype Project scans and tested on 47,400 UK Biobank scans It beats both scanner-derived DXA measures and other models at cross-cohort prediction of diseases and biomarkers, with ~150,000× fewer training images and ~40× fewer parameters. Over 4.3 years of follow-up it improves prediction of incident hip and knee arthrosis and type 2 diabetes Its representations accurately predict age (r = 0.88, MAE 2.9y), and the biological age gap tracks disease burden and a 45% higher mortality hazard, and appears modifiable, decreasing in women after starting hormone-replacement therapy Preprint: arxiv.org/abs/2608.02208

Our new preprint about PulseOx-FM: a foundation model trained on ~7M overnight pulse-ox segments from >10,000 Human Phenotype Project participants We show that a wearable sleep signal carries broader physiology than standard HR, SpO₂, or apnea features PulseOx-FM beat engineered PPG and proprietary feature sets on age prediction, generalized to an external surgical cohort, and performed well across 64 phenotypic cardiometabolic, sleep, imaging, medication & neuropsychiatric targets It also captured 2-year hypertension incidence from baseline, and next-day glycemic, dietary & activity state within individuals Preprint: medrxiv.org/content/10.64898…

Our new preprint shows that walking after a meal significantly lowers its blood glucose response Based on glucose responses to 400,000 meals across 10,000 Human Phenotype Project participants medrxiv.org/content/10.64898…

Our new preprint is a significant milestone for us We built "HealthFormer" by training on our deeply phenotyped cohort from the Human Phenotype Project data. Healthformer is a multimodal generative transformer model that tokenizes each participant's physiological trajectory across 667 modalities (biomarkers, body comp, sleep, CGM, microbiome, wearables, meds) and is trained with a single objective: predict the next measurement Forecasting, risk stratification, and intervention-conditioned simulation all arise as queries from one shared representation Key findings: → Matches direction of effect in 41/41 published RCTs; 30/41 within the reported 95% CI → Reconstructs biomarkers at r > 0.9; forecasts 2 yrs ahead → Validation on external data from UK Biobank, NHANES, PNP3, Framingham → Outperforms Framingham CVD & PREVENT-ASCVD on 27/30 endpoints → Predicts individual 6-mo responses in a held-out RCT Paper: arxiv.org/abs/2604.27899 Great work by Guy Lutsker, Gal Sapir, Jordi Merino, Smadar Shilo, Anastasia Godneva, Eli Meirom, Shie Mannor, Hagai Rossman, Gal Chechik

Our new preprint on a gait foundation model that we developed from videos of over 3,400 people walking and doing other motor tasks. This 3D skeletal motion model can predict diverse human traits and disease conditions arxiv.org/abs/2603.25283

Our two new arxiv papers achieve state-of-the-art (SOTA) performance on multiple long-term time-series forecasting tasks, using way fewer parameters than previous SOTA models. Our models are based on mixture of expert transformer architectures Papers: arxiv.org/abs/2601.21641 arxiv.org/abs/2601.21866 Great work by Evandro Ortigossa and @GLutsker

The future is already here: An AI model built on continuous glucose monitoring (CGM) data predicts Type 2 diabetes and cardiovascular deaths better than the standard biomarker, HbA1c

Nature research paper: A foundation model for continuous glucose monitoring data go.nature.com/4602eqO

🚀 GluFormer, a transformer model trained on 10M+ glucose readings, predicts diabetes and cardiovascular risk up to 12 years ahead - outperforming HbA1c. Built by NVIDIA Israel, @WeizmannScience, @Pheno_AI, and clinical partners. 📄 nvda.ws/4rnWIa7

Sharing Gluformer, our latest paper @Nature: The first generative foundation model for blood glucose data, trained on data from 10,812 adults of the Human Phenotype Project Gluformer predicts risk of diabetes and cardiovascular outcomes better than the standard of care Full paper: rdcu.be/eY5fH Collaboration of @WeizmannScience @mbzuai @nvidia @Pheno_AI Work by @GLutsker Gal Sapir, @smadarshilo, Jordi Merino, @nastya_godneva, Jerry R. Greenfield, Dorit Samocha-Bonet, Raja Dhir, Francisco Gude, Shie Mannor, Eli Meirom, @ericxing, Gal Chechik, and @H_Rossman

Continuous glucose monitoring sensor data, with a foundation model, predicts risk of Type 2 diabetes and cardiovascular outcomes better than HbA1c New @Nature @segal_eran @WeizmannScience @GLutsker "66% of incident diabetes cases and 69% of cardiovascular deaths occurred in the top risk quartile, compared with 7% and 0%, respectively, in the bottom quartile." nature.com/articles/s41586-0…

In the Human Phenotype Project, home sleep apnea testing data was collected for a total of 16,812 nights in 6,410 individuals, allowing for a comprehensive study of the association of sleep traits with physiological features across 16 body systems. nature.com/articles/s41591-0…

Our lab at the Weizmann Institute was hit tonight by Iran We will rebuild and return 💪