An Active Role for Machine Learning in the Diagnosis of Atrial Fibrillation Quentin Hughes Minnetonka High School

Overview

Abbott Laboratories Atrial Fibrillation (AF) is a cardiac disease affecting over three million Americans, with a prevalence of 1% (above 6% in the elderly).1 AF is an irregular heartbeat that can lead to blood clots, stroke, heart failure and death. Annually, AF is responsible for over 750,000 hospitalizations and 130,000 deaths in the U.S.2 Accurate and early diagnosis is key, however, the disease is significantly under-diagnosed. Monitoring devices can acquire large amounts of real-time ECG images but evaluating this big data remains a challenge. Diagnostic accuracy with current technologies is high but false positive rates are between 27-90%. There is a growing need for automated image analysis. This study aims to create a supervised machine learning algorithm that will reliably identify AF and reduce the false positive rate. 5,761 pre-classified ECG images were collected, augmented and classified as either AF or NOT for binary analysis. It was hypothesized that a machine learning algorithm could be built to diagnose AF that maintains sensitivity while significantly reducing the false positive rate. The optimal model included seven convolution neural networks, four dense layers and nine epochs. Testing the final model on the validation images resulted in sensitivity of 94.7%, specificity of 95.1% and a false positive rate of 4.9%. With a consecutive-image trigger, the false positive rates are dramatically reduced (< one per 10 years). Using this algorithm to diagnose AF can potentially achieve similar accuracy and a significantly lower false positive rate than the reported algorithms used today.

Competition history

JSHS 2020 Category not listed

Resources

National JSHS abstract book

Related projects

ISEF · 2020