Technical Summary: MorphologyFM

Problem Statement

While foundation models have revolutionized representation learning in natural language and computer vision, their application to physiological waveforms (specifically Electrocardiograms [ECG] and Pulse Oximetry [SpO2]) has been limited by the choice of self-supervised objectives. Existing approaches primarily optimize for signal reconstruction (e.g., Masked Autoencoders) or forecasting. These objectives successfully capture local temporal continuity but fail to explicitly preserve clinically meaningful waveform morphology.

The authors argue that the diagnostic value of physiological signals lies not in individual sample amplitudes, but in structural characteristics such as peak amplitude, interval duration, slope, inflection points, and beat-to-beat variability. Current methods may reconstruct a signal with high fidelity while failing to distinguish subtle morphological patterns critical for diagnosing arrhythmias, hypoxemia, or respiratory compromise. There is a need for a foundation model that treats waveform morphology as the primary inductive bias for learning transferable representations without requiring manual annotations.

Methodology: MorphologyFM

MorphologyFM is a multimodal foundation model pretrained on paired ECG and SpO2 waveforms from the MIMIC critical care database. It employs a self-supervised learning framework designed to learn latent representations that are invariant to noise but sensitive to morphological changes.

1. Architecture

• Tokenization: The model uses two modality-specific 1D convolutional tokenizers to convert raw ECG ($x_{ecg}$) and SpO2 ($x_{spo2}$) segments into patch embeddings. This allows each modality to learn specific low-level features (e.g., sharp depolarization in ECG vs. smooth vascular propagation in SpO2).
• Encoder: The token sequences are concatenated with modality and positional embeddings and processed by a shared Transformer encoder using multi-head self-attention. This enables the model to reason jointly about electrical cardiac activity and peripheral vascular responses.
• Decoder: A decoder reconstructs masked regions of the input waveform.

2. Core Components

The training objective combines three distinct components:

• Morphology-Aware Masking: Unlike standard random masking, this strategy identifies cardiac cycles (via R-peaks in ECG and pulse peaks in SpO2) and masks entire morphological components (e.g., P waves, QRS complexes, ST segments, systolic upstrokes). This forces the model to infer higher-order physiological relationships rather than relying on local interpolation.
• Cross-Modal Representation Learning: The model aligns the latent representations of synchronized ECG and SpO2 windows. A projection head maps modality-specific pooled representations into a shared space, treating synchronized pairs as positive examples and unrelated patients as negatives. This encourages the learning of shared cardiovascular states while preserving modality-specific morphology.
• Contrastive Morphology Objective: An InfoNCE loss is applied to embeddings derived from two independent morphology-preserving augmentations of the same waveform. This explicitly organizes the latent space so that morphologically similar signals cluster together.

3. Training Objective

The total loss function is a weighted sum of three terms:
$$L = \lambda_{rec}L_{rec} + \lambda_{con}L_{con} + \lambda_{align}L_{align}$$
Where $L_{rec}$ is the reconstruction error over masked regions, $L_{con}$ is the contrastive loss, and $L_{align}$ is the cross-modal alignment loss. Crucially, this process requires no diagnostic labels.

Key Contributions

1. Morphology-Centric Pretraining: The paper introduces a self-supervised framework where waveform morphology, rather than raw signal reconstruction, serves as the primary learning objective.
2. Morphology-Aware Masking Strategy: A novel masking technique that targets clinically relevant structural components (e.g., QRS complexes) rather than random time steps, forcing the model to learn physiological structure.
3. Multimodal Foundation Model: The demonstration that jointly modeling ECG and SpO2 produces more transferable representations than single-modality pretraining, leveraging the complementary nature of electrical and hemodynamic signals.
4. Comprehensive Evaluation: The model is evaluated across four diverse downstream tasks: arrhythmia classification, hypoxemia prediction, mortality prediction, and length-of-stay estimation.

Results

MorphologyFM was evaluated on the MIMIC database and compared against representative self-supervised baselines: Masked Autoencoders (MAE), Contrastive Learning (SimCLR-style), Barlow Twins, and Joint Embedding Predictive Architectures (JEPA).

• Downstream Performance: MorphologyFM achieved the highest performance across all tasks.
  • Arrhythmia (F1): 89.7% (vs. 86.8% for JEPA).
  • Hypoxemia (AUROC): 92.5% (vs. 89.3% for JEPA).
  • Mortality (AUROC): 90.4% (vs. 87.2% for JEPA).
  • Length of Stay (AUROC): 85.8% (vs. 82.7% for JEPA).
  • Average Score: 89.6, outperforming the next best method (JEPA) by 3.1 points.
• Multimodal Benefit: Joint pretraining on ECG + SpO2 yielded an average AUROC of 89.6, significantly outperforming ECG-only (87.4) and SpO2-only (85.8) models.
• Ablation Studies:
  • Replacing random masking with morphology-aware masking improved average AUROC from 87.9 to 89.6.
  • The full objective (Reconstruction + Contrastive + Alignment) outperformed any ablated version, confirming the complementary value of all components.
• Scaling: Performance improved consistently as the pretraining corpus size increased from 100k to 10M waveform segments, with the largest gains observed between 100k and 1M segments.
• Representation Quality: In nearest-neighbor patient retrieval tasks, MorphologyFM achieved a Recall@10 of 95.1%, outperforming all baselines, indicating superior preservation of clinically meaningful similarity.

Significance and Claims

The paper claims that waveform morphology constitutes a powerful inductive bias for self-supervised physiological representation learning. The authors argue that:

1. Reconstruction Fidelity $\neq$ Representation Quality: High-fidelity signal reconstruction does not guarantee that a model has learned the structural features clinicians use for diagnosis. Explicitly modeling morphology bridges this gap.
2. Universal Physiological Representation: Morphology provides a universal invariant across patients, devices, and clinical settings, making it a superior target for foundation models compared to generic time-series reconstruction.
3. Scalability without Labels: The approach leverages the millions of hours of unlabeled waveforms generated in routine clinical care, offering a practical path to foundation models without the bottleneck of manual annotation.

The authors position MorphologyFM as a step toward general-purpose physiological foundation models capable of learning robust representations directly from raw continuous signals, shifting the paradigm from treating biomedical signals as generic time series to treating them as morphology-rich clinical data.

Limitations Acknowledged:

• Experiments are currently limited to ECG and SpO2 from a single dataset (MIMIC).
• The model does not yet incorporate other critical modalities like arterial blood pressure, capnography, or EEG.
• Future work is needed to validate robustness across external healthcare systems and devices.