An Interpretable Model for ECG Data Based on Bayesian Neural Networks

Heart arrhythmia have been a life-threatening disease to human for a very long time, many techniques and methods have been developed by human experts since the advent of ECG machine, these approaches work well and most of the time patients with heart diseases can be diagnosed accurately, and yet not timely, human experts for heart disease are far from adequate. Deep learning models for heartbeat arrhythmia detection, therefore, have been proposed and investigated during the past few years, they have good results on some data-sets concerning arrhythmia detection and classification, some even outperformed human doctors. However, like most deep learning models, the underlying mechanism has not been justified and explained very well, for example, we cannot interpret the decision momentum of certain models. Concerning that, we developed a general feature extraction framework for ECG data, it can perform various kinds of feature engineering tasks, all of the features have their meaning under a clinical context. We distilled the features on a large ECG data-set, then these features were fed into a Bayesian Neural Network for training. The BNN has a sparsity-inducing prior distribution arranged in a tied manner to learn the importance of every feature concerning the outcome. We evaluated this method for a classification problem and made a comparison between the way our model makes decisions and how field experts do for the same problem, we were able to find out that a) the features extracted from raw ECG data using our framework can be used as an indication for specific heartbeat arrhythmia, b) the mechanism of the model can be interpreted as a decision making procedure by weighing relative metrics just like human experts do, c) the weight of the features extracted from raw ECG data can be used to build a knowledge tree for guidance on diagnosing of certain heartbeat disease.