Leveraging Cardiologists Prior-Knowledge and a Mixture of Experts Model for Hierarchically Predicting ECG Disorders

Automatic methods to identify cardiovascular disorders based on Electrocardiograms (ECGs) have been used for decades and gained special attention with the breakthrough of deep learning.This paper proposes a new model based on a Mixture of Experts (MoE) to identify 6 physician-defined clinical labels spanning rhythm and conduction disorders.The MoE framework combines the opinions of multiple models ("experts") to make weighted decisions.Unlike typical MoE scenarios, where the gating network finds ways to weigh and combine the models, our approach leverages prior knowledge from cardiology specialists, who hierarchically divide the space, enabling us to tailor the MoE architecture to incorporate this information explicitly.The disorders are organized in a 2-level hierarchy, where specialists first separate normal ECGs from those rooted in rhythm and morphological abnormalities.The MoE is trained on the CODE-15 dataset, and the results represent the new state-of-the-art in this task.