HealthFound learns from longitudinal health records: it beats the strongest comparator by 10 points on average across 1,162 UK Biobank endpoints
Synopsis
Medical language models handle knowledge questions but struggle to infer disease trajectories from an individual's repeated measurements, because literature and exams give little supervision on how indicators covary over time. HealthFound converts records from 501,936 UK Biobank participants into readable longitudinal health profiles and mixes them with medical and general reasoning data, training in three stages: curriculum supervised fine-tuning, rejection-sampled retention fine-tuning and verifiable policy optimization. Across 1,162 endpoints it reaches about 69% mean accuracy, 10.0 percentage points above the strongest comparator within each clinical chapter on average, and it also leads in zero-shot transfer to MIMIC-IV and NHANES.
Interpretation
HealthFound extends a language model's training signal from descriptive medical text to numerical follow-up records measured repeatedly in individuals, so one model both answers medical exam questions and judges future high or low risk from pre-onset measurements. Earlier medical language models learned mainly from biomedical literature, clinical notes and medical exam questions, which are largely qualitative and give little supervision on how measurements covary across individuals, change over time, or relate pre-disease profiles to later outcomes. On 1,162 UK Biobank endpoints, HealthFound reached about 69% overall mean accuracy, with a mean advantage of 10.0 percentage points over the strongest comparator within each clinical chapter, ranging from 7.9 to 11.5 points, and the difference appeared across endocrine/metabolic, cardiovascular, genitourinary, respiratory, musculoskeletal, mental, neurological, digestive, ophthalmic and neoplastic groups.
The advantage does not come from longer reasoning chains: the model generates a mean of 138.6 reasoning tokens per high-risk or low-risk judgment. DeepSeek-V3.2, Qwen3 and MiniMax-M2.7 used means of 985.6, 806.7 and 775.9 reasoning tokens respectively, all more than HealthFound, so generation length cannot explain its accuracy advantage. Of the 1,162 endpoints, 444 reached at least 70% accuracy with HealthFound, compared with 47 for Gemini-3-Flash, 41 for DeepSeek-V3.2, 41 for GPT-5.4, 28 for Qwen3 and none for MiniMax-M2.7.
Training proceeds in three steps: curriculum supervised fine-tuning gradually introduces cohort tasks while interleaving medical reasoning data, rejection-sampled retention fine-tuning recovers weakened medical abilities, and policy optimization with verifiable rewards strengthens quantitative reasoning. Feeding raw cohort tables to a model for next-token prediction is not feasible, because tens of thousands of coded fields must become readable sequences without losing numerical values, units or temporal order, and extensive cohort training risks eroding pretrained medical knowledge. A component ablation shows curriculum supervised fine-tuning raised UK Biobank accuracy from 57.1% to 61.9% and mean public-benchmark accuracy from 66.3% to 80.0%; retention fine-tuning raised the public benchmarks to 84.3%, while policy optimization raised UK Biobank accuracy from 62.8% to 69.0% and changed public benchmarks by 0.4 percentage points.
What is learned from UK Biobank transfers zero-shot to datasets collected for different purposes, with no further training, reward adjustment or prompt change at transfer. MIMIC-IV supplies hospitalization trajectories and NHANES supplies cross-sectional population measurements, and both differ from UK Biobank in sampling, observation density and label definitions. On 475 MIMIC-IV incident endpoints, HealthFound reached about 77% overall accuracy, Gemini-3-Flash about 71% and the remaining models 64% or lower; on the nine NHANES prevalent-disease tasks it achieved the highest accuracy for every reported endpoint and the aggregate comparison.
Perspective
The demonstrated scope is high-risk or low-risk classification under a retrospective, matched case-control construction, scored by accuracy and positive recognition, covering 1,162 UK Biobank endpoints, eight public medical benchmarks, 475 MIMIC-IV endpoints and nine NHANES tasks; every modality enters as text-serialized derived values, with imaging contributing only as imaging-derived phenotypes and no raw images, waveforms or free-text clinical notes processed directly.
UK Biobank is affected by healthy-volunteer selection and limited demographic representativeness, and disease labels can reflect coding practices, healthcare contact and unequal follow-up; text serialization preserves values, units and ordering but simplifies irregular sampling, repeated measurements and informative missingness; the principal disease evaluations use balanced or matched samples and emphasize accuracy and positive recognition rather than calibration, the qualitative cases provide no error rate, and the representation projection was not compared against the unadapted backbone; current evidence does not establish clinical benefit, treatment utility or safety under prospective deployment.