Abstract:
A major contributor to deaths worldwide, heart disease causes public health systems under great
strain and emphasizes medical resources. Reducing death and enhancing patient outcomes
depend much on early detection; yet, present diagnostic techniques are usually expensive, out of
reach, or need specialized knowledge and equipment. This project intends to provide a more
affordable and easily available diagnostic tool, especially appropriate for use in resource-limited
settings.
The project addresses this difficulty by using Phonocardiogram (PCG) data recordings of heart
sounds as a non-invasive input for heart disease prediction. Collected from several sites, PCG
recordings were processed to lower noise and highlight diagnostically relevant characteristics.
Techniques of feature extraction were used to identify audio patterns linked to cardiac anomalies.
Trained and assessed were several machine learning and deep learning models including Support
Vector Machines, Random Forests, and Convolutional Neural Networks. The best-performing
model was included into a web-based application allowing users to submit their PCG files and
get immediate diagnostic forecasts with related confidence scores.
Initial system tests showed encouraging results; the model's accuracy was 91% and its AUC
ROC score was 0.9753, indicating great dependability in differentiating between healthy and in
danger human beings. Strong sensitivity and specificity were verified by the confusion matrix.
These results suggest that the programme might be a useful and efficient early screening tool for
heart disease. Its accuracy is being improved by more tests and optimizations, therefore
guaranteeing more general real-world relevance.
