Predicting the Effects of Drugs and Comorbidities on Arrhythmogenic Risk using Deep Learning
Emma Maliar

Abstract
Arrhythmias affect millions of people worldwide. In particular, arrhythmias cause 200,000 - 300,000 sudden deaths in the US per year. Doctors diagnose arrhythmias by looking at electrocardiograms that represent a patient’s heartbeat. However, these diagnoses are not always accurate. For example, the average cardiologist accuracy for diagnosing atrial fibrillation is about 50%. We developed an artificial intelligence statistical model that diagnoses arrhythmias by analyzing the patients’ electrocardiogram values, comorbidities and drugs. Our model, written in python, relies on a multilayer neural network which is trained by deep learning optimization methods. We analyze the database ECG-ViEW II from South Korea that contains information on 461,000 patients, including 10,081 comorbidities and drugs. To reduce the runtime, because of the large dataset, we run the code on a supercomputer Bridges 2. Our deep learning model diagnoses arrhythmias with an overall accuracy of 83.87%, thus overperforming trained medical doctors. Using our deep learning model to evaluate how drugs and comorbidities contribute to a patient's risk of suffering from an arrhythmia, we find that most common drugs such as aspirin and Vitamin C do not significantly affect the incidence of arrhythmias. But we also distinguished a number of drugs and comorbidities that have a strong statistically significant effect on the incidence of arrhythmias, particularly, drugs related to pregnancy, skin eruptions, and the stroke. Our deep learning analysis can aid doctors diagnose and prevent arrhythmias by informing prescriptions and by treating comorbidities that increase the risk of arrhythmias.

Full Text: PDF