Neural Network Regularization for Generalized Heart Arrhythmia Classification

Sammanfattning: Background: Arrhythmias are a collection of heart conditions that affect almost half of the world’s population and accounted for roughly 32.1% of all deaths in 2015. More importantly, early detection of arrhythmia through electrocardiogram analysis can prevent up to 90% of deaths. Neural networks are a modern and increasingly popular tool of choice for classifying arrhythmias hidden within ECG-data. In the pursuit of achieving increased classification accuracy, some of these neural networks can become quite complex which can result in overfitting. To combat this phenomena, a technique called regularization is typically used. Thesis’ Problem Statement: Practically all of today’s research on utilizing neural networks for arrhythmia detection incorporates some form of regularization. However, most of this research has chosen not to focus on, and experiment with, regularization. In this thesis we measured and compared different regularization techniques in order to improve arrhythmia classification accuracy. Objectives: The main objective of this thesis is to expand upon a baseline neural network model by incorporating various regularization techniques and compare how these new models perform in relation to the baseline model. The regularization techniques used are L1, L2, L1 + L2, and Dropout. Methods: The study used quantitative experimentation in order to gather metrics from all of the models. Information regarding related works and relevant scientific articles were collected from Summon and Google Scholar. Results: The study shows that Dropout generally produces the best results, on average improving performance across all parameters and metrics. The Dropout model with a regularization parameter of 0.1 performed particularly well. Conclusions: The study concludes that there are multiple models which can be considered to have the greatest positive impact on the baseline model. Depending on how much one values the consequences of False Negatives vs. False Positives, there are multiple candidates which can be considered to be the best model. For example, is it worth choosing a model which misses 11 people suffering from arrhythmia but simultaneously catches 1651 mistakenly classified arrhythmia cases?