Sensor analytics for interpretation of EKG signals

Abstract Motivation and Objectives Smartphones are emerging as personal fitness assistants, collecting data through in-built or external sensors. The next frontier for these devices is to use advanced sensors and machine learning algorithms to offer more personalized and advanced medical assessments. Along these lines, the objective of this paper is to develop a multi-label classification model to detect heart complications through electrocardiograms (EKGs) collected by an FDA-approved single-lead EKG sensor attached to a smartphone. The EKG sensor produces a standard EKG chart, but for such a sensor to be useful to a consumer, an interpretation of the graph is necessary. Materials and Method We adapt a machine-learning approach to detect multiple heart conditions simultaneously from the generated EKG graph. Three different multi-label machine learning models (binary relevance, label powerset and multi-perceptron neural network) were built and compared to categorize five different heart states: Normal, Atrial Fibrillation, Atrioventricular Block, Sinus Bradycardia and Sinus Tachycardia. The binary relevance model was selected based on the accuracy. Results and Implications The model generated rules inductively from the data to interpret nine out of every ten heart conditions correctly. Our model is being adapted for commercial use by a company (as a part of their App) that markets the EKG sensor for smartphones. Our model is usable in a cardiovascular disease alert expert system that will potentially allow users to monitor their heart health continuously and prevent a serious illness by providing this information in the early stages.