CoMET (Cosmos Medical Event Transformer)

Summary

CoMET treats a patient's medical history as a token sequence and trains a decoder-only transformer to predict the next medical event. Conditioned on real history, it autoregressively simulates forward patient timelines, and predictions for arbitrary clinical questions are read off those simulations rather than from a task-specific head. Across 78 real-world tasks — diagnosis prediction, disease prognosis and healthcare operations — it generally matched or outperformed task-specific supervised models without fine-tuning or few-shot examples.[1]

Source Status

Preprint: compiled from an arXiv preprint not yet peer reviewed, authored by the vendor of the underlying data platform. confidence: medium. See the note on evaluation independence below.

Data and training

• Source: Epic Cosmos — de-identified longitudinal health records covering 16.3 billion encounters across 300 million unique patients from 310 health systems.[1]
• Training subset: 115 billion medical events from 118 million unique patient records, January 2012 – April 2025, rendered as 151 billion tokens.[1]
• Architecture: Qwen2 transformer, randomly initialised — no pretrained language-model weights are loaded, so this is not an LLM adapted to health records but a model trained from scratch on medical events.[1]
• Tokenization: compact per-event tokens, adapted from the ETHOS approach for Cosmos's scale and heterogeneity.[1]
• Three compute-optimal models up to 1 billion parameters.[1]

Scaling laws

The paper's most transferable contribution is a Chinchilla-style scaling analysis for medical event data: 10 model sizes from 2 million to 1 billion parameters trained over 136 billion tokens, with isoFLOP parabolas fitted at four fixed compute budgets.[1]

The best-fit exponents were α = 0.520 and β = 0.512 for compute-optimal parameter count and token count respectively — closely mirroring the natural-language values (α = 0.49, β = 0.51).[1] Because α ≈ β, model size and training tokens should be scaled proportionally as compute increases, exactly as for text.

Inference: that medical event streams follow essentially the same compute-optimal scaling exponents as natural language is a non-obvious empirical result, and it implies the scaling playbook developed for text transfers to this modality without re-derivation. The authors report the coefficients and the similarity but do not draw this generalisation explicitly.[1]

Downstream clinical performance improves predictably as training loss falls, with a sigmoid fitted across checkpoints — so loss is a usable proxy for clinical utility within the regime tested.[1]

Evaluation

CoMET was compared against three classes of task-specific supervised baselines trained on each downstream task: linear/logistic regression, gradient-boosted trees (XGBoost), and supervised transformers trained from random initialisation.[1] Task families covered:

• disease-specific outcome prediction, acute-on-chronic incidence, and incident disease risk;
• early quantitative differential diagnosis generation;
• health-system operations — hospital length of stay and 30-day readmission;
• generation plausibility (encounter types and frequency, single-encounter realism).

Performance consistently improved with model and pretraining scale.[1]

Note on evaluation independence: the study was conducted by Epic Systems — the vendor of the Cosmos platform — with Microsoft Research and academic co-authors, and the benchmark tasks, baselines and data are all internal to that platform. The comparison against task-specific models is the paper's own; no independent external replication is cited. The reported result (a generic generative model matching supervised task-specific models zero-shot) is strong and should be treated as vendor-reported until replicated.

Relation to other trajectory models in this vault

Synthesis: CoMET, Delphi-2M and Aladynoulli all model disease trajectories over time, but differ in what they treat as the unit of prediction — CoMET generates the next medical event token in an operational health-record stream, whereas the others are built around diagnosis-level disease progression. CoMET is additionally distinguished by using simulation-based inference: rather than fitting a head per outcome, it samples forward timelines and computes the statistic of interest from them, which is what allows one model to serve 78 tasks. No cited study compares these frameworks directly.

The trade-off this creates is worth naming: simulation-based inference gives task generality but makes calibration a property of the sampling procedure rather than of a fitted output layer, and the paper evaluates a conservative n = 20 simulations for some analyses.[1]

See Also

• Delphi-2M — generative disease-trajectory modelling from health records.
• Aladynoulli — latent disease-trajectory modelling.
• Disease Embedding Space — representation learning over diagnosis histories.
• Multimodal Disease Risk Prediction — the broader risk-prediction context.
• Expert Oversight of AI in Bioinformatics — relevant to the vendor-evaluation caveat above.

Citations

[1] Waxler, S., Blazek, P., White, D., Sneider, D. et al. (2025). Generative Medical Event Models Improve with Scale. arXiv preprint arXiv:2508.12104. Supports: dataset scale and training subset; Qwen2 random-initialisation architecture and ETHOS-derived tokenization; the scaling-law exponents α = 0.520 / β = 0.512 and their natural-language comparison; the 78-task evaluation and the three supervised baseline classes; the loss-to-performance sigmoid; the n = 20 simulation setting. Location: Abstract; Results §2.1 (training and inference), §2.3–2.5 (prediction tasks), §2.6 (scaling laws), §2.7 (scaling performance); Table 1; Figs. 15–16. Source paper: 2508.12104v1.pdf