MIRA: Medical Time Series Foundation Model for Real-World Health Data
- Hao Li ,
- Bowen Deng ,
- Chang Xu ,
- Zhiyuan Feng ,
- Viktor Schlegel ,
- Yu-Hao Huang ,
- Yizheng Sun ,
- Jingyuan Sun ,
- Kailai Yang ,
- Yiyao Yu ,
- Jiang Bian
NeurIPS 2025 |
A unified foundation model for medical time series — pretrained on open access and ethics board-approved medical corpora — offers the potential to reduce annotation burdens, minimize model customization, and enable robust transfer across clinical institutions, modalities, and tasks, particularly in data-scarce or privacy-constrained environments. However, existing generalist time series foundation models struggle to handle medical time series data due to their inherent challenges, including irregular intervals, heterogeneous sampling rates, and frequent missing values. To address these challenges, we introduce MIRA, a unified foundation model specifically designed for medical time series forecasting. MIRA incorporates a Continuous-Time Rotary Positional Encoding that enables fine-grained modeling of variable time intervals, a frequency-specific mixture-of-experts layer that routes computation across latent frequency regimes to further promote temporal specialization, and a Continuous Dynamics Extrapolation Block based on Neural ODE that models the continuous trajectory of latent states, enabling accurate forecasting at arbitrary target timestamps. Pretrained on a large-scale and diverse medical corpus comprising over 454 billion time points collect from publicly available datasets, MIRA achieves reductions in forecasting errors by an average of 10% and 7% in out-of-distribution and in-distribution scenarios, respectively, when compared to other zero-shot and fine-tuned baselines. We also introduce a comprehensive benchmark spanning multiple downstream clinical tasks, establishing a foundation for future research in medical time series modeling.
Related Tools
MIRA: Medical Time Series Foundation Model for Real-World Health Data
January 29, 2026
MIRA is a foundation model for medical time-series, designed to learn a unified representation space across heterogeneous clinical datasets and support zero-shot forecasting in real-world healthcare settings. Unlike conventional time-series models that operate on fixed sampling rates or task-specific feature spaces, MIRA is built to handle irregular and clinically diverse signals natively. By combining continuous-time encoding, frequency-aware specialization, and neural dynamics modeling, MIRA generalizes robustly across conditions. MIRA is pretrained on 454B time points collected from large-scale clinical corpora spanning both ICU physiological signals and hospital EHR time-series, covering a rich range of sampling frequencies (minute-level vitals, hourly labs, waveform segments, and multi-day clinical indicators). This large and heterogeneous training distribution allows MIRA to serve as a unified backbone capable of strong out-of-distribution generalization. In extensive evaluations, MIRA achieves state-of-the-art zero-shot forecasting performance across diverse clinical benchmarks. Compared with existing foundation models, MIRA obtains SOTA results on 4 of 5 out-of-distribution evaluation settings on standard baselines—demonstrating strong robustness under dataset shift, irregular sampling, and multimodal temporal variations. Key features Continuous-Time Rotary Positional Encoding (CT-RoPE) Provides a principled way to embed irregular timestamps while preserving temporal geometry, enabling robust reasoning across arbitrary sampling patterns. Frequency-specialized Mixture-of-Experts Allows different experts to specialize on physiological signs, improving transfer across diverse clinical signals. Neural ODE Extrapolation Models latent dynamics continuously over time, enabling forecasting at arbitrary future timestamps.