Abstract
Medical large language models (LLMs) have shown promise in medical knowledge retrieval and clinical reasoning. However, their capacity for quantitative reasoning over dense, longitudinal health data remains limited, largely due to the lack of effective training strategies that leverage large-scale, natural-language-based prospective cohorts. To address this, we present HealthFound, a medical LLM trained on 12.4 million self-constructed examples. These include 15‑year longitudinal biomedical records from 501,936 participants, as well as datasets for medical knowledge and general reasoning. We propose a novel three-stage training framework comprising progressive curriculum supervised fine-tuning, rejection-sampled retention fine-tuning, and task-verifiable policy optimization. This enables the model to effectively parse heterogeneous quantitative follow‑up data and perform robust quantitative reasoning. HealthFound achieves state‑of‑the‑art performance across multiple newly constructed benchmarks. It attains the highest accuracy on eight public medical benchmarks for general medical competence. On 1,162 UK Biobank tasks predicting future disease onset from biomedical measurements, it outperforms the second‑best model by over 10 percentage points on average. In two independent external validations (MIMIC‑IV and NHANES), it demonstrates superior zero‑shot generalization, achieving improvements of over 6% across 484 disease endpoints. The model also achieves top results on novel tasks for literature‑derived question answering and quantitative indicator imputation. Furthermore, HealthFound generates interpretable reasoning rationales that are consistent with established biomedical knowledge. Collectively, our results demonstrate that HealthFound enables medical LLMs to evolve from static knowledge repositories into dynamic quantitative reasoning engines over longitudinal health data.