MIRA: Medical Time Series Foundation Model for Real-World Health Data

A foundation model for medical time series, pretrained on ethically approved<br/>clinical datasets, can substantially reduce annotation burdens, minimize the need<br/>for task-specific tuning, and promote reliable transferability across healthcare<br/>institutions, data modalities, and clinical tasks, especially in data-scarce or privacysensitive environments. However, existing generalist time series foundation models<br/>struggle to handle medical time series data due to their inherent challenges, including irregular intervals, heterogeneous sampling rates, and frequent missing values.<br/>To address these challenges, we introduce MIRA, a unified foundation model<br/>specifically designed for medical time series forecasting. MIRA incorporates aContinuous-Time Rotary Positional Encoding that enables fine-grained modeling<br/>of variable time intervals, a frequency-specific mixture-of-experts layer that routes<br/>computation across latent frequency regimes to further promote temporal specialization, and a Continuous Dynamics Extrapolation Block based on Neural ODE that<br/>models the continuous trajectory of latent states, enabling accurate forecasting at<br/>arbitrary target timestamps. Pretrained on a large-scale and diverse medical corpus<br/>comprising over 454 billion time points collect from publicly available datasets,<br/>MIRA achieves reductions in forecasting errors by an average of 8% and 6% in outof-distribution and in-distribution scenarios, respectively, when compared to other<br/>zero-shot and fine-tuned baselines. We also introduce a comprehensive benchmark<br/>spanning multiple downstream clinical tasks, establishing a foundation for future<br/>research in medical time series modeling. Our code is available at MIRA.