An AI-Based Diagnostic Framework for Cardiopulmonary Health Monitoring from Clinical Dataset Analysis

Abstract

Cardio-respiratory diseases remain a major global health concern, accounting for a significant proportion of mortality, particularly in India where they contribute to nearly 30% of total deaths. Early and accurate diagnosis is critical, yet traditional auscultation using acoustic stethoscopes relies heavily on physician expertise, often leading to subjective and inconsistent interpretations. This study presents an automated cardio-respiratory sound classification system designed to enhance diagnostic reliability through machine learning and deep learning techniques. A novel dataset was developed using a digital stethoscope to capture high-quality heart and lung sounds from a clinical manikin, enabling the collection of both individual and mixed recordings. This dataset provides combined cardiorespiratory audio signals, facilitating dual classification tasks. The system incorporates a user-friendly graphical interface with role-based access, allowing administrators to train models and users to perform predictions seamlessly. Audio signals are processed using Librosa for feature extraction, including MelFrequency Cepstral Coefficients (MFCC), Chroma features, and Mel spectrograms. Multiple traditional machine learning models such as Quadratic Discriminant Analysis (QDA), Gradient Boosting Classifier (GBC), Naive Bayes Classifier (NBC), and Logistic Regression Classifier (LRC) are implemented and evaluated. In addition, a custom deep learning architecture combining a Bi-directional Convolutional Neural Network (BiCNN) with a Bi-directional Gated Recurrent Unit (BiGRU) is proposed to capture spatial and temporal characteristics. The system performs simultaneous prediction of heart and lung sound types from mixed recordings, achieving robust performance across evaluation metrics.