Technical Summary: Modality vs. Morphology: A Framework for Time Series Classification for Biological Signals

Problem Statement

Time series classification (TSC) of biological signals (e.g., EEG, ECG, EMG, PPG, ocular signals) has historically progressed from handcrafted, modality-specific approaches to deep learning architectures. However, a disconnect remains between the unique challenges of specific signal modalities and the underlying waveform structures (morphologies) that define them. Traditional methods often struggle to generalize across populations or conditions because they rely on domain expertise that may not transfer, while deep learning models sometimes fail when their inductive biases do not align with the specific temporal dynamics of the signal. The core problem is the lack of a unified framework that connects waveform structure to methodological design, leading to suboptimal preprocessing, feature engineering, and model selection.

Methodology: The Morphology–Modality Framework

The authors propose a dual framework that unifies two traditional perspectives:

1. Modality-Oriented Perspective: Examines specific signals (EEG, EMG, ECG, PPG, ocular) individually, addressing their unique noise profiles, sensor constraints, and preprocessing needs.
2. Morphology-Oriented Perspective: Organizes temporal dynamics by five distinct waveform structures, regardless of the signal source:
   • Spikes/Bursts: Isolated, high-amplitude, short-duration deflections (e.g., epileptic discharges, muscle activations).
   • Oscillations: Regular, repeating rhythms (e.g., cardiac cycles, alpha/beta bands).
   • Slow Drift: Gradual baseline shifts over long time horizons (e.g., cognitive load in pupillometry).
   • Hierarchical: Multiscale structures where local waveforms are nested within larger cycles (e.g., P-QRS-T complexes, nested cross-frequency coupling).

The framework analyzes how these morphologies dictate specific strategies across the TSC pipeline:

• Preprocessing: Strategies must be tailored to morphology. For instance, aggressive normalization can flatten transient spikes, while standard bandpass filtering may remove critical slow-drift information.
• Feature Engineering & Extraction: Different morphologies require different descriptors. Time-domain features suit spikes/bursts; frequency-domain features suit oscillations; decomposition techniques are vital for mixed morphologies; and symbolic representations work for oscillations but fail for precise spike timing.
• Modeling: The paper maps morphologies to optimal model classes:
  • Spikes/Bursts: Local feature-based models (SVMs, ROCKET, CNNs) or hybrid CNN-Transformers.
  • Oscillations: State-space models (SSMs), frequency-domain methods, and cycle-synchronous attention mechanisms.
  • Slow Drift: Recurrent networks (LSTMs) and residual architectures that model gradual changes.
  • Hierarchical: Hybrid architectures (CNN-LSTM, Transformers, SSM-Attention) that integrate local and global context.
• Evaluation: Standard metrics (accuracy, F1) are insufficient. The authors advocate for morphology-specific metrics (e.g., spike detection precision, beat-to-beat accuracy, drift estimation error) and temporal responsiveness.

Key Contributions

1. Unified Framework: The introduction of a morphology–modality framework that serves as a structured guide for comparing methodologies and identifying cross-domain opportunities in biological signal analysis.
2. Morphology-Driven Design Principles: The demonstration that morphology, not model class, is the primary determinant of performance and interpretability. The paper argues that deep models succeed when their inductive biases (e.g., locality, periodicity, long-range dependency) align with the underlying waveform dynamics.
3. Operational Guidelines: The provision of detailed, actionable recommendations for preprocessing, feature selection, and model deployment based on waveform structure (summarized in Tables I and II).
4. Comprehensive Review: A synthesis of evidence across five major biological signal modalities, illustrating how the dual framework enhances analysis for tasks ranging from seizure detection to stress monitoring.
5. Identification of Gaps: The paper highlights critical limitations in current practices, including the lack of standardized benchmarks for behavioral signals, the risks of morphology-agnostic data augmentation, and the need for evaluation metrics that reflect real-world temporal and clinical scenarios.

Results and Evidence

The paper synthesizes findings from existing literature to validate the framework:

• Spikes/Bursts: Studies on EEG and EMG show that performance improves when localized convolutional or probabilistic filters are combined with temporal-attention mechanisms to preserve fine-scale alignment. Hybrid models (e.g., EEG Conformer) achieve state-of-the-art results by coupling local inductive biases with global context.
• Oscillations: For ECG and PPG, classical frequency-domain features remain competitive, but deep learning models (e.g., PhysMamba, ECG DETR) excel when they preserve phase coherence and spectral stability. State-space models are shown to bridge physiological realism with computational efficiency.
• Slow Drift: In ocular signals, architectures that incorporate baseline normalization and temporal recurrence (e.g., BiLSTMs) outperform static approaches by isolating trends from noise.
• Hierarchical: Models that employ layered abstractions (e.g., CNN-LSTM hybrids, multiresolution attention) consistently yield better physiological interpretability and robustness for signals like ECG and EEG.
• Multimodal Fusion: Integrating complementary morphologies (e.g., EEG + PPG) via morphology-aware synchronization and adaptive channel weighting leads to superior performance compared to single-modality baselines.

Significance and Claims

The paper claims that the proposed framework offers a foundational shift in how biological signal classification is approached. Its significance lies in:

• Generalizability: By focusing on morphology, the framework enables the development of models that generalize across different modalities and populations, moving beyond "one-size-fits-all" deep learning solutions.
• Interpretability: Aligning model design with physiological morphology ensures that models rely on meaningful features rather than artifacts, which is critical for clinical adoption and regulatory approval.
• Future Direction: The authors posit that future progress depends on three converging priorities: (1) expanding and standardizing datasets to include diverse morphologies and behavioral contexts; (2) developing evaluation metrics that reflect real-world temporal dynamics; and (3) designing models that balance accuracy with interpretability.

The paper concludes that the goal is not to identify a single optimal model per modality, but to bridge modality-specific characteristics with morphology-based principles to achieve generalizable, interpretable, and clinically meaningful solutions.