Open-Source Demonstration Reproduction of MIRA

This repository provides an open-source demonstration reproduction of
MIRA: Medical Time Series Foundation Model for Real-World Health Data
(arXiv: https://arxiv.org/abs/2506.07584).

The goal of this demo is to offer a fully accessible and reproducible version of the MIRA framework, trained exclusively on publicly available time-series datasets, without using any proprietary or sensitive clinical data.
It is designed to showcase the core ideas of MIRA rather than replicate the full-scale training described in the original paper.


Overview

This demonstration model reproduces the key architectural components of MIRA:

  • Continuous-Time Rotary Positional Encoding (CT-RoPE)
    Enables geometric time encoding for irregularly sampled signals. The CT-RoPE temporal unit is set to 0.1.

  • Frequency-Specialized Mixture-of-Experts (MoE)
    Allows experts to specialize on low- vs. high-frequency clinical patterns.

  • Neural ODE–Based Continuous Dynamics Modeling
    Captures long-range physiological trajectories with continuous-time formulations.


Notes

  • This model is a demonstration reproduction, not the full-scale MIRA model.
  • Performance will differ from the original paper due to dataset differences.

Downloads last month
16
Safetensors
Model size
0.5B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support