Technical Summary: MedMamba

Problem Statement

Medical time series (MedTS) classification is critical for clinical decision support, yet it faces three primary challenges that existing methods struggle to address simultaneously:

1. Local-Global Dynamics: Models must capture discriminative local morphological features (e.g., sharp spikes) while modeling long-range temporal dependencies.
2. Nonstationarity: Real-world recordings often suffer from baseline drift and acquisition artifacts. These slow shifts can obscure diagnostic patterns and reduce generalization, particularly in subject-independent evaluation settings.
3. Latent Channel Interactions: Multichannel recordings (e.g., EEG, ECG) contain diagnostic cues in cross-channel interactions. However, these dependencies are often dynamic and sample-specific, making predefined anatomical graphs insufficient or inaccurate.

Existing approaches, such as CNNs (limited in long-range modeling), Transformers (quadratic complexity), and standard Graph Neural Networks (often reliant on static graphs or lacking nonstationarity handling), fail to balance these requirements efficiently.

Methodology

The authors propose MedMamba, an end-to-end architecture that integrates State Space Models (SSMs) with domain-specific inductive biases. The architecture consists of three core components:

1. Multi-Scale Convolutional Embedding (MCE)

To capture discriminative local morphology across different temporal scales, MedMamba employs parallel depthwise temporal convolutions with varying kernel sizes (e.g., $k \in \{3, 5, 7\}$). These features are concatenated and projected to a hidden dimension, providing a rich initial representation that preserves local waveform details before being fed into the SSM backbone.

2. Tri-Branch Differential State Space Encoder (TDSSE)

To mitigate nonstationarity and baseline drift, the model processes input through three complementary temporal views in parallel, fusing them via learnable gates:

• Raw View: A bidirectional Mamba block processes the raw signal to capture long-range context.
• Differential View: The input is transformed into first-order differences ($\Delta X_t = X_t - X_{t-1}$). This acts as a high-pass filter to suppress slow-varying baseline drift and emphasize dynamic changes.
• Frequency View: A learnable spectral modulation is applied in the frequency domain (via FFT) to isolate informative rhythmic bands (e.g., EEG rhythms) while attenuating artifacts.
  The three branches are fused using local element-wise gates (for denoising) and global view-selection gates, allowing the model to dynamically prioritize the most informative view for a given segment.

3. Spatial Graph Mamba (SGM) Module

To model latent cross-channel dependencies without relying on predefined anatomical graphs, MedMamba introduces a sample-conditioned adaptive graph learning mechanism:

• Graph Learning: For each input sample, the model learns a directed adjacency matrix $A$ by projecting pooled channel representations into "source" and "target" latent spaces. This allows the dependency structure to adapt to specific sample characteristics.
• Structure Priors: To ensure robustness and interpretability, the learned graph is regularized with two constraints:
  • Sparsity ($L_1$ norm): Encourages a sparse dependency structure to filter noise.
  • Acyclicity (DAG constraint): Uses a trace exponential penalty ($\text{tr}(e^{A \odot A}) - C$) to enforce a Directed Acyclic Graph structure, preventing unstable cyclic feedback loops.
• Spatial Fusion: The module combines a graph diffusion pathway (propagating information along the learned directed edges) with a channel-wise SSM pathway (modeling global spatial dependencies), fusing them via an adaptive gate.

The entire system is trained end-to-end with a cross-entropy loss augmented by the graph structure regularization terms.

Key Contributions

1. MedMamba Architecture: A novel framework combining SSMs with medical-domain inductive biases (multi-scale morphology, multi-view temporal encoding, and adaptive graph learning) for efficient and robust MedTS classification.
2. Tri-Branch Differential Encoder: A mechanism that explicitly addresses nonstationarity by fusing raw, differential, and frequency-domain views, significantly improving robustness to baseline drift.
3. Adaptive Spatial Graph Module: A data-driven approach to learning directed, sparse, and acyclic channel dependencies without prior topological knowledge, capturing complex cross-channel interactions.
4. Extensive Evaluation: Comprehensive experiments on five real-world medical datasets (three EEG, two ECG) under both subject-dependent and subject-independent protocols.

Experimental Results

MedMamba was evaluated on five datasets: ADFTD, APAVA, TDBRAIN (EEG), and PTB, PTB-XL (ECG).

• Performance: The model achieved State-of-the-Art (SOTA) performance across most metrics (Accuracy, F1, AUROC, AUPRC) on all five datasets. Notably, it secured the top F1 score across all datasets in the challenging subject-independent setting, demonstrating strong generalization to unseen subjects.
• Robustness: Experiments injecting controlled baseline drift and simulating missing channels showed that MedMamba degrades more gracefully than baselines (including MedGNN and raw-only variants), validating the efficacy of the TDSSE and adaptive graph learning.
• Efficiency: MedMamba maintains linear computational complexity with respect to sequence length ($O(T)$ or $O(T \log T)$ due to FFT), offering a favorable trade-off between accuracy and computational cost compared to quadratic-complexity Transformers and heavy graph-based models.
• Ablation Studies: Removing any component (MCE, TDSSE, or SGM) resulted in consistent performance drops, confirming the necessity of each module. Specifically, the TDSSE was crucial for nonstationary datasets, while SGM was vital for capturing cross-channel interactions.

Significance and Claims

The paper claims that MedMamba offers a practical solution for clinical deployment by addressing the specific limitations of current deep learning models in medical time series analysis. By integrating linear-complexity state space modeling with domain-specific inductive biases, MedMamba successfully bridges the gap between computational efficiency and the need for robust, interpretable modeling of nonstationary, multichannel physiological signals.

The authors emphasize that their approach does not rely on massive pre-trained parameters or static anatomical assumptions. Instead, it learns sample-adaptive functional connectivity that aligns with physiological plausibility (as shown in qualitative graph visualizations) while maintaining robustness against artifacts like baseline drift. This suggests that specialized, lightweight architectures can outperform general-purpose large models in high-stakes healthcare applications, paving the way for real-time, explainable diagnostic monitoring.