A Large-scale Multimodal Study for Predicting Mortality Risk Using Minimal and Low Parameter Models and Separable...

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Biomedical studies often rely on limited datasets, hindering generalizability across diverse patient populations and long-term data.
• Predicting 1-year mortality is crucial for patient care, offering a critical timeframe for intervention.

Purpose of the Study:

• To develop and validate multimodal predictive models for 1-year mortality using a large-scale clinical dataset.
• To assess the contribution of individual factors and data modalities to overall mortality risk.
• To create a family of low-parameter models for clinical application.

Main Methods:

• Development and validation of multimodal models using over 25 million echocardiography videos and 2.9 million 8-lead electrocardiogram (ECG) traces.
• Utilized a massive dataset encompassing 316,125 patients for echocardiography and 631,353 patients for ECG.
• Employed optimized feature selection based on feature importance to create low-parameter models.

Main Results:

• Models demonstrated strong predictive performance, with AUC ranging from 0.72 (10 parameters) to 0.89 (under 105k parameters).
• The study successfully assessed the contribution of individual factors and modalities to mortality risk.
• A family of models was constructed and made available in the DISIML package.

Conclusions:

• The developed modular neural network framework provides insights into global mortality risk trends.
• These models can guide therapies and interventions aimed at reducing mortality risk.
• The approach enables the creation of highly performant, interpretable predictive models from large clinical datasets.