Twelve Lead Double Stacked Generalization for ECG Classification

The success of deep learning models has led to their widespread acceptance as automated decision support systems to aid in clinical diagnosis. Their prevalent adoption has led to their use in in-home patient monitoring and diagnosis to improve the quality of human life. The dependency on such systems requires the development of knowledge-intensive, high-performing models for superior healthcare outcomes. One of the most popular in-home telehealth systems is for monitoring cardiac events using Electrocardiogram (ECG) signals from the heart. ECG signals have a wide range of temporal, spatial, and frequency features which altogether account for clinical diagnosis. ECG recordings are usually obtained from 12 leads placed on the patient; however, the predominant body of research develops machine learning models using only a single lead or all 12 leads as 12 channels of a single model. We propose an approach to leverage knowledge from all 12 leads individually using double-stacked generalization which uses a meta-learner to develop an ensemble of single-channel and 12-channel deep learning models. Our proposed approach shows performance improvement of more than 5.2% in Area under Curve (AUC) and 7.94% in (Maximum F1 score) F
max
over individual lead models. Furthermore, our level 2 stacks 12-channel model with a level 1 meta-learner leveraging the features learned by the multi-channel and single-channel models leading to state-of-the-art F
max
and AUC.