Prototype Learning to Create Refined Interpretable Digital Phenotypes from ECGs.

Area of Science:

• Artificial Intelligence
• Biomedical Informatics
• Cardiology

Background:

• Prototype-based neural networks offer interpretable predictions by learning representative signal patterns.
• While promising for physiological data classification, their ability to capture broader clinical phenotypes is unclear.

Purpose of the Study:

• To assess if prototypes from a deep learning model trained on ECG data align with clinical phenotypes (phecodes) in an external database.
• To evaluate the interpretability and clinical relevance of individual prototypes beyond standard classification.

Main Methods:

• A prototype-based deep learning model was trained for multi-label ECG classification on the PTB-XL dataset.
• The unmodified model was used for inference on the MIMIC-IV clinical database.
• Associations between individual prototypes and hospital discharge diagnoses (phecodes) were analyzed.

Main Results:

• Individual prototypes showed stronger, more specific associations with clinical outcomes than class predictions or NLP-extracted concepts.
• Prototype classes with mixed significance patterns displayed greater intra-class distances, indicating differentiation of meaningful variations.
• The model achieved high predictive performance for cardiac conditions (e.g., AUC 0.91 for heart failure) and showed signal for non-cardiac conditions like sepsis.

Conclusions:

• Prototype-based models can facilitate interpretable digital phenotyping from physiological time-series data.
• These models provide transferable intermediate phenotypes that capture clinically meaningful signatures beyond initial training objectives.
• Prototypes offer a valuable tool for understanding the link between physiological signals and diverse clinical conditions.