TLDR: CoMET (Cosmos Medical Event Transformer) is a new family of AI models trained on Epic Cosmos, the largest medical event dataset (115 billion events from 300M patients). It autoregressively generates future medical events, simulating patient health timelines. CoMET consistently matched or outperformed task-specific supervised models across 78 diverse clinical tasks, including disease prediction, differential diagnosis, and operational forecasting, without requiring fine-tuning. The research also confirmed that medical event models follow power-law scaling, meaning performance predictably improves with increased compute, data, and model size, indicating significant future potential.
The quest for personalized medicine, delivering the right care to the right patient at the right time, relies heavily on understanding a patient’s entire health journey. These journeys, often viewed as sequences of medical events, hold a wealth of insights. A new approach, leveraging large-scale foundation models, is showing significant promise in distilling this complex information to generate real-world evidence and adapt to various healthcare tasks.
Researchers have introduced the Cosmos Medical Event Transformer, or CoMET, a family of advanced artificial intelligence models. These models are built on a massive dataset called Epic Cosmos, which compiles de-identified health records from over 300 million unique patients and includes 16.3 billion medical encounters. CoMET was specifically trained on an unprecedented 115 billion discrete medical events, representing 151 billion tokens, making it the largest medical-event foundation model by training data volume to date. The core function of CoMET is to autoregressively generate the next medical event, effectively simulating a patient’s health timeline based on their real-world history.
Unveiling Predictive Power Across Clinical Tasks
The study rigorously evaluated CoMET’s performance across 78 real-world tasks, encompassing areas like diagnosis prediction, disease prognosis, and healthcare operations. Remarkably, for a foundation model with generic pretraining and simulation-based inference, CoMET consistently matched or even surpassed the performance of traditional task-specific supervised models on these diverse tasks. This was achieved without requiring any task-specific fine-tuning or even a few-shot examples, highlighting its inherent generalizability.
One of CoMET’s key strengths lies in its ability to generate realistic medical event sequences. Evaluations showed that the model produced valid multi-token events (like diagnosis codes or medication orders) with very low error rates, which decreased further with larger model sizes. It also accurately reflected the prevalence and co-occurrence rates of individual medical events, demonstrating a strong understanding of clinical patterns. Furthermore, CoMET proved adept at probabilistically forecasting the number and types of future medical encounters, such as office visits, emergency visits, and hospital admissions, with high calibration.
Beyond plausibility, CoMET demonstrated significant predictive capabilities. It could estimate future disease risk for conditions like Type 2 Diabetes, hyperlipidemia, and hypertension, often outperforming supervised models. The model also showed strong performance in predicting acute-on-chronic events, such as asthma exacerbations or sickle cell crises, and identifying the first instance of various disease states in the general population. A particularly innovative application is CoMET’s capacity to generate early, quantitative differential diagnoses, helping to flag patients at risk for specific conditions weeks or even months before a definitive diagnosis.
In terms of healthcare operations, CoMET proved valuable in forecasting patient interactions with the health system. It showed lower mean absolute error in predicting future encounter counts compared to supervised models. Additionally, it accurately predicted hospital length of stay and the risk of 30-day hospital readmissions, offering crucial insights for resource management and patient care planning.
Also Read:
- Building a Dynamic Medical Knowledge Graph with AI Agents
- Lean Language Models Master Reasoning and Retrieval for Private AI Applications
The Science of Scale: Following Predictable Laws
A significant contribution of this research is the largest scaling-law study conducted on medical event data. The findings indicate that, similar to large language models in natural language processing, medical event models exhibit power-law scaling relationships. This means that as computational resources, the number of training tokens, and model size increase, the model’s performance improves predictably. The study found an optimal token-to-parameter ratio of 1,000:1, suggesting efficient scaling for future medical foundation models. Crucially, the research established that minimizing training loss consistently translates into better performance on downstream clinical evaluations, indicating substantial room for further improvement by continuing to scale these models.
While CoMET represents a significant leap forward, the researchers acknowledge certain limitations. The model is trained on real-world data from Cosmos, which, despite rigorous quality control, can contain inherent imprecision or gaps. The tokenization method, while comprehensive, discretizes continuous values like lab results, potentially masking subtle but important changes. Future work aims to incorporate more diverse data types, explore counterfactual reasoning for actionable insights, and conduct prospective studies to assess real-world clinical alignment. For more technical details, you can refer to the original research paper: Generative Medical Event Models Improve with Scale.
In conclusion, CoMET stands as a powerful, general-purpose engine for real-world evidence. Its ability to screen populations for incident disease risk, forecast individual health timelines, surface differential diagnoses, and anticipate health system interactions—all from a single set of autoregressive generations—positions it as a transformative tool to support clinical decision-making, streamline healthcare operations, and ultimately improve patient outcomes.
