Generative Medical Event Models Improve with Scale: Curiosity Model Validates Scaling Laws for Medical Event Prediction on 115 Million Patients
In August 2025, a paper introduced the Curiosity series of models, pretrained on the Epic Cosmos dataset (310 million patients, 16.3 billion visits, 115 billion medical events). The largest model has 1 billion parameters and, on 78 real-world tasks (diagnosis prediction, prognosis, operations), outperforms or matches task-specific supervised models without fine-tuning. This is the first validation of power-law scaling relationships for computation, data, and model size on medical event data.
The paper demonstrates scaling laws for medical event sequence modeling: as model and data size increase, prediction performance consistently improves. Under zero-shot/few-shot settings, the Curiosity model surpasses traditional supervised models on tasks such as diagnosis, prognosis, and operations, showcasing the potential of general-purpose foundation models in healthcare. This shifts the previous belief that medical data is sparse and requires extensive domain customization, providing empirical evidence for building scalable medical AI infrastructure.
Curiosity uses a decoder-only Transformer architecture, pretrained on 115 billion medical events (151B tokens) with autoregressive next-event prediction. The study systematically tests different model sizes (from millions to 1 billion parameters) and training data volumes, finding that performance follows a power-law relationship with compute, and the optimal model aligns with the Chinchilla law. On 78 tasks, inference is performed by simulating patient history trajectories without task-specific fine-tuning. For example, on a diagnosis prediction task, the Curiosity 10B model achieves an AUC of 0.92, surpassing a supervised model trained on the same data (0.90).
This work has significant implications for the healthcare IT and AI-assisted decision-making industry. As the largest medical event dataset, the openness of Epic Cosmos will affect industry competition. The Curiosity model suggests that general-purpose medical foundation models can replace numerous task-specific models, reducing development and maintenance costs. Hospitals and insurers may consider deploying such models for risk stratification, resource planning, and clinical decision support.
Recommend that healthcare IT vendors evaluate the Curiosity model or similar foundation models to replace multiple single-task models and reduce integration costs. Hospitals can explore using it for patient risk prediction and resource scheduling. Investors should focus on companies with large-scale medical event data (e.g., Epic, Cerner) and startups building foundation models on such data.
Key points to watch: whether Epic will open the Cosmos dataset or collaborate with third parties; whether the Curiosity model will be open-sourced; and performance improvements from larger models (e.g., 100B parameters). Additionally, the model's generalization across different healthcare systems and demographic groups, as well as its deployment effectiveness and safety in real clinical workflows, need validation.