Anish Salvi
Data-Driven Bioengineer
Cardiac Condition Classifier
As part of the Yale CBIT Healthcare Hackathon 2021, I submitted my project classification of electrocardiogram (ECG/EKG) signals. The objective of this work was to develop a convolutional neural network (CNN) that classified ECG images as myocardial infarctions (MI) or as healthy.
Unfortunately, the model behavior was somewhat inconsistent for testing accuracy. There are a few key improvements I would make, such as changing the Loss function. Nevertheless, I improved my intuition in Data Mining and Machine Learning. Inspired by this project, I hope to develop a web-based application for Arrhythmia classification using TensorFlow and Arduino sensors.
MI, or heart attacks, claimed the lives of over 100,000 people in 2019. Due to the difficulty of interpreting ECG signals, a CNN approach was required. Data was derived from the PTB Diagnostic ECG Database and visualized via the WFDB library. Coding was in Python on Google Colab.
By manipulating NumPy arrays and utilizing Matplotlib, ECG signal images and their corresponding Healthy or MI classifications were derived. The training set consisted of 64 Healthy/64 MI images and the testing set consisted of 16 Healthy/16 MI images.
Via the Keras library, a TensorFlow model consisting of Conv2D, MaxPooling2D, and Dense layers was encoded. To avoid overfitting, Data Augmentation and Dropout layers were added. The Adam optimizer and Crossentropy loss function were implemented as well.
