Deep Learning Analysis of ECG Signals for Early Myocardial Infarction Detection

This thesis investigates the development of a robust machine learning pipeline for the early detection of myocardial infarction (MI) using elec- trocardiogram (ECG) data. Recognizing MI promptly is vital for effective treatment, potentially saving lives by reducing the mortality and morbidity associated with this condition. The research leverages the comprehensive PTB-XL dataset, which includes a vast array of clinical ECG recordings annotated by medical professionals, making it an ideal resource for train- ing and validating the proposed neural network models. The methodology encompasses several stages, starting with the preprocessing of ECG data to enhance signal quality and remove noise and artifacts. Various neural network architectures were explored, including Convolutional Neural Net- works (CNNs), and LSTM networks, to determine the most effective model for ECG analysis. The study employed data augmentation techniques such as noise addition, time warping, and signal shifting to address the issue of overfitting and improve the generalizability of the models. Training was executed using Python and TensorFlow, with an emphasis on optimizing the neural networks through meticulous hyperparameter tuning and the application of advanced optimizers like Adam and SGD. The effectiveness of different loss functions and learning rate schedulers was also evaluated to enhance model training dynamics. The models exhibited high accuracy and precision in detecting MI from ECG signals. The LSTM model, in particular, showed a significant improvement in performance, achieving an accuracy of 96%, under augmented data conditions. These results high- light the potential of advanced neural networks in the automatic detection of cardiac event. The machine learning pipeline developed in this thesis marks an advancement in the automated detection of myocardial infarc- tion using ECG data. The findings suggest that such models can be inte- grated into clinical settings to provide real-time, accurate assessments of MI, thereby facilitating prompt medical intervention. Future work will fo- cus on ECG image analysis and introducing techniques such as fine-tuning to potentially increase the pipeline’s utility in clinical settings.