MedMamba: A New State Space Model for Medical Time Series Classification
MedMamba introduces a novel multi-scale bidirectional state space model specifically designed for classifying medical time series data, such as ECG and EEG signals. This innovative architecture overcomes drawbacks associated with existing convolutional, recurrent, and Transformer-based methods, which often face challenges due to their quadratic complexity. MedMamba integrates three key inductive biases: spatial centralization, multi-timescale temporal composition, and non-causal contextual dependency. Additionally, it features a lightweight channel-mixing module for effective cross-channel reparameterization. The research is accessible on arXiv under the identifier 2605.05214.
Key facts
- MedMamba is a multi-scale bidirectional state space architecture.
- It is tailored for medical time series classification.
- The model addresses limitations of convolutional, recurrent, and Transformer-based approaches.
- It incorporates three inductive biases: spatial centralization, multi-timescale temporal composition, and non-causal contextual dependency.
- A lightweight channel-mixing module is used for cross-channel reparameterization.
- The paper is on arXiv with identifier 2605.05214.
- Medical time series include ECG and EEG signals.
- Transformer-based approaches incur quadratic complexity.
Entities
Institutions
- arXiv