# Generative Medical Event Models Improve with Scale

**Authors:** Shane Waxler, Paul Blazek, Davis White, Daniel Sneider, Kevin Chung, Mani Nagarathnam, Patrick Williams, Hank Voeller, Karen Wong, Matthew Swanhorst, Sheng Zhang, Naoto Usuyama, Cliff Wong, Tristan Naumann, Hoifung Poon, Andrew Loza, Daniella Meeker, Seth Hain, Rahul Shah  
**arXiv:** [2508.12104](https://arxiv.org/abs/2508.12104)  
**Date:** 2025-08-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2508.12104v1/

## Key points

- Personalized medicine at scale requires distilling insights from longitudinal patient journeys, which are sequences of medical events.
- Pretrained foundation models on large-scale medical event data, such as the Cosmos Medical Event Transformer (CoMET) models, offer a promising direction for scaling real-world evidence generation and generalizing to diverse downstream tasks.
- CoMET, pretrained on a dataset representing 118 million patients and 115 billion discrete medical events, can autoregressively generate the next medical event based on a patient's history and outperformed task-specific supervised models across 78 real-world tasks without requiring fine-tuning.
- The study established a methodology for pretraining transformer models and revealed power-law scaling relationships for compute, tokens, and model size in medical event data.
- CoMET effectively captures complex clinical dynamics as a generative medical event foundation model, supporting clinical decision-making, streamlining healthcare operations, and enhancing patient outcomes.

## Summary

In the pursuit of safe and effective medical care, personalized medicine at scale requires methods that can distill insights from longitudinal patient journeys. These journeys are essentially sequences of medical events. Pretrained foundation models on large-scale medical event data offer a promising direction for scaling real-world evidence generation and generalizing to diverse downstream tasks. Leveraging Epic Cosmos, a dataset comprising de-identified longitudinal health records for over 300 million unique patients and 16.3 billion encounters from 310 health systems, the Cosmos Medical Event Transformer (CoMET) models were introduced. These decoder-only transformer models were pretrained on a massive dataset representing 118 million patients and 115 billion discrete medical events. A comprehensive scaling-law study was conducted for medical event data, establishing a methodology for pretraining and revealing power-law scaling relationships for compute, tokens, and model size. Subsequently, a series of compute-optimal models with up to 1 billion parameters were pretrained. Conditioned on a patient's real-world history, CoMET can autoregressively generate the next medical event, simulating patient health timelines. The study encompassed 78 real-world tasks including diagnosis prediction, disease prognosis, and healthcare operations. Remarkably, CoMET outperformed or matched task-specific supervised models on these tasks without requiring task-specific fine-tuning or few-shot examples. The predictive power of CoMET consistently improved as the model and pretraining scale increased. Results demonstrate that CoMET effectively captures complex clinical dynamics as a generative medical event foundation model. This framework provides an extensible and generalizable approach to support clinical decision-making, streamline healthcare operations, and enhance patient outcomes. Furthermore,<kg>the introduction of Epic Cosmos</kg> has addressed challenges in leveraging real-world data for personalized medicine at scale by aggregating de-identified longitudinal health records across multiple health systems. The platform unifies various clinical data types to support patient care and accelerate scientific discovery while delivering actionable insights to clinicians at the point of care through features like Cosmos Median Length of Stay and Best Care Choices for My Patient™. Despite the vast potential of Cosmos data in informing healthcare decisions and research priorities such as understanding trends in healthcare utilization or investigating rare diseases, answering specific clinical questions still requires manual effort in crafting custom cohort definitions and feature-engineering pipelines. To enable routine clinical decision-making with personalized medicine at scale using RWE demands tools that can learn from integrated patient records efficiently answer complex medical inquiries across diverse contexts.