Designing a Machine Learning Solution to Atrial Fibrillation

Overview

A common heart arrhythmia that afflicts millions in the United States alone, atrial fibrillation (AF or AFib) increases the risk of serious health complications such as heart failure, stroke, blood clots, and more (American Heart Association). In patients, AFib presents its characteristic heartbeat quiver only part of the time, leaving many patients undiagnosed. My research takes on the task of using a machine learning algorithm in order to reliably detect AFib. In this unbiased trial, PhysioNet’s dataset of 8,528 patients’ ECG samples is first imported into the Python notebook that contains the convolutional neural network (CNN) architecture. The patients used in this research fall into three classifications: normal heartbeat (0), atrial fibrillation (1), and other heart arrhythmia(s) (2). Some preprocessing, including downsampling, filtering, and adding artificial noise, was performed on this data before training the model. For class splitting, the traditional 50:25:25 division was used between training, validation, and testing datasets. One of the three previously mentioned classifications was outputted to the GUI following the model’s decision on any input data. Several factors within the model, such as filtering procedures and CNN layer composition, were refined in order to improve the performance of the algorithm. This algorithm has generated some very promising results. The area under the Receiver Operating Characteristic (ROC) Curve averaged 0.86, showing high predictive value. When plotted, the algorithm’s Confusion Matrix is also encouraging, showing that most of the inaccuracies arise from misclassification between classes 1 and 2. False negatives are rare, accounting for only 3.9% of the predictions. False positives are more common, at 9.7% of the predictions, a future directive of focus for this research. To address the on-and-off issue that AFib poses, this algorithm can be implemented through wearable sensors such as FitBit and Apple watches, so that ECG data is continuously being collected and analyzed. The classifications made based on these data can be passed directly to both patient and doctor for better-informed medical care decisions. Overall, based on the results obtained in this study, this machine learning algorithm has the capability of detecting heart arrhythmias accurately and may in the future become a promising diagnostic tool for medical applications.

Competition history

• AJAS 2020 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• AJAS None projects
• Every AJAS 2020 project