Using Machine Learning to Analyze Image-Based Volume Characteristics of Left Atrial Volumes in Atrial Fibrillation Patients

Abstract

Detecting atrial fibrillation (AF) using artificial intelligence models is much more efficient than current clinical management methods and can reduce financial burdens on the healthcare industry. The purpose of this study was to train an AI model to detect whether a given sample of left atria (LA) showed signs of AF and the difference between LA volumes of people with and without AF. Using LA image datasets from the Cardiac Atlas Project and MSD Cardiac Dataset, I was able to analyze the images using ITK Snap, load them onto a Python console and use those images to train the AI segmentation model. The implementation of a convolutional neural network was essential in training the model, as it allowed the model to break down visual aspects of the LA and distinguish between AF and Non-AF LAs. The main characteristic used to determine whether a given LA had AF was its volume. After calculating the volume of the LA of both AF and non-AF patients, as well as conducting a significance test, I was able to conclude that the larger the volume of the LA was, the more likely the patient was to be diagnosed with AF. The average volume of LAs in patients without AF was 21839.77 mm³, while the average LA volumes in patients diagnosed with AF was 55649.42 mm³. The significance test showed that there is a significant difference between these two averages. Taken together, these results show that analyzing the volume of the LA is a very effective way of determining whether a patient is diagnosed with AF.