Final Design Report: Smart Stethoscope

The Smart Stethoscope project integrates signal processing and artificial intelligence (AI) to improve the diagnostic accuracy of heart abnormalities, especially in remote environments where access to skilled professionals may be limited. Cardiovascular diseases remain a leading cause of mortality worldwide, highlighting a need for innovative solutions in healthcare. This initiative seeks to address that need by developing a machine learning model capable of classifying 42 different heart abnormalities with an accuracy of 90%, using a dataset composed of stethoscope recordings. The proof of concept of this project was carried out during the Fall semester, using a 2D-Convolutional Neural Network (CNN) approach on a dataset of 10 different heart abnormalities. During the Spring semester, the team seeked to simplify and optimize the proof of concept by replacing the 2D-CNN approach for two approaches: 1D-CNN and Wavelet Transforms. As per these approaches, the same dataset was used for their implementation in its full length, consisting of 42 unique heart conditions, each represented by a 10-15 second sample of continuous heartbeats. To better simulate a real-world dataset of heart sounds collected from a wide range of patients, this data was procedurally expanded using MATLAB. This dataset was splitted into 70% training data, 15% validation, and 15% testing data for both implementations. The first model is a one-dimensional convolutional neural network (CNN) implemented in Python, trained on the full length of each of the 42 audio samples. The second model, implemented with MATLAB, is a feedforward neural network trained on wavelet-transformed embeddings of the same 42 samples. In terms of results, both approaches showcased promising results above the accuracy threshold set at the beginning of the project (90%). The CNN implementation achieved a 99.7% without overfitting with a size of 3 Mb. At the same time, the feedforward implementation yielded an accuracy of 96.14%, using three fully connected layers, and three dropout layers, which helped optimize the model. The results and implementation of this project showed that it was possible to classify many heart sounds, and it has the potential to be expanded for other applications such as ECG analysis. Overall, the work done in this project sets a basis for the creation of a valuable tool for improving heart condition diagnoses in areas with limited resources.