AI- Enhanced Stethoscope in Remote Diagnostics for Cardiopulmonary Diseases

Technical Summary: AI-Enhanced Stethoscope in Remote Diagnostics for Cardiopulmonary Diseases

Problem Statement
Cardiovascular and pulmonary diseases represent a leading cause of global mortality, yet timely diagnosis remains a significant challenge, particularly in remote and under-resourced regions. Traditional auscultation relies heavily on the subjective interpretation of skilled practitioners, a resource often scarce in developing areas. While digital stethoscopes exist, their high cost limits accessibility, and many existing AI-driven diagnostic models are restricted to analyzing either heart or lung sounds individually, rather than concurrently. Furthermore, many current models struggle with class imbalance in datasets and lack the lightweight architecture necessary for deployment on low-cost embedded devices.

Methodology
The authors propose a hybrid deep learning model designed to run on low-cost embedded hardware, integrated with a traditional stethoscope modified with a microphone for audio capture. The system workflow involves the following technical stages:

1. Data Acquisition and Pre-processing:

   • Hardware: A standard stethoscope connected via a 3.5mm jack to a device hosting a web application.
   • Signal Processing: Raw auscultation signals (25–400 Hz) are filtered using a 2nd-order Butterworth band-pass filter, resampled to 1000 Hz, and truncated to 2.5-second sequences (2500 data points). Signals are normalized to the range (-1, 1).
   • Augmentation: To address class imbalance (particularly in the ICBHI 2017 lung dataset), techniques such as noise addition, temporal shifting, and pitch/speed alteration are applied to underrepresented classes.

2. Feature Extraction:

   • The system utilizes Mel-Frequency Cepstral Coefficients (MFCCs) to capture essential audio features mimicking human auditory perception.
   • The process involves windowing, Fourier transformation to the frequency domain, application of Mel filter banks, logarithmic compression, and Discrete Cosine Transform (DCT).
   • The input shape for the model is defined as (1, 52) MFCC features.

3. Model Architecture (Hybrid CNN-GRU):

   • The core innovation is a hybrid architecture combining Convolutional Neural Networks (CNN) for spatial feature extraction and Gated Recurrent Units (GRU) for temporal sequence learning.
   • CNN Layers: Two convolutional layers (256 and 512 filters respectively) with kernel size 11, followed by max pooling, batch normalization, and ReLU activation.
   • GRU Layers: Five sets of GRU layers with varying units (32, 64, 128) using 'tanh' activation, with outputs combined via an additive layer.
   • Classification: Dense layers with Leaky ReLU activation lead to a final output layer with 11 units and softmax activation for multi-class classification.

4. Training and Deployment:

   • The model is trained using the Adam optimizer (learning rate 0.0002) and Categorical Crossentropy loss.
   • Techniques such as Early Stopping (patience of 300 epochs) and Model Checkpointing are employed to prevent overfitting.
   • The trained model is deployed via a Streamlit-based web application, allowing real-time analysis of audio recorded from the low-cost stethoscope and enabling telemedicine through email report generation.

Key Contributions

• Concurrent Diagnosis: Unlike previous works that focus on either cardiac or pulmonary sounds, this model simultaneously diagnoses 11 distinct classes: 5 cardiovascular conditions (Aortic Stenosis, Mitral Stenosis, Mitral Regurgitation, Normal, Mitral Valve Prolapse) and 6 pulmonary conditions (Pneumonia, Bronchiectasis, Bronchiolitis, URTI, COPD, Healthy).
• Cost-Effective Accessibility: The system is explicitly designed for low-cost embedded devices, aiming to bridge the gap in resource-limited settings where high-end digital stethoscopes are unaffordable.
• Hybrid Architecture: The integration of CNN and GRU leverages both frequency-based and temporal dependencies, addressing limitations of standalone models.
• End-to-End System: The paper presents a complete workflow from hardware setup to a functional web interface for remote consultation.

Results
The study evaluated the proposed hybrid model against standalone CNN and GRU models using metrics including Precision, Recall, F1-Score, and Accuracy.

• Lung Sounds: The standalone CNN achieved an average accuracy of 74.16%, and the GRU achieved 75.83%. The hybrid CNN+GRU model significantly improved performance, reaching an average accuracy of 94% (with specific class accuracies ranging from 75% for Pneumonia to 94% for Bronchiolitis).
• Heart Sounds: The standalone CNN achieved 81% accuracy, and the GRU achieved 83%. The hybrid model achieved an average accuracy of 94% across all cardiac classes (e.g., 94.01% for Aortic Stenosis, 94.09% for Normal).
• Comparison: The hybrid model demonstrated superior performance in handling class imbalance and complex feature extraction compared to individual models, as evidenced by confusion matrices and comparative metric tables.

Significance and Claims
The authors claim that this work represents a transformative step toward standardized, accessible healthcare. By integrating a cost-effective stethoscope with an efficient AI model, the system aims to:

• Reduce human error in diagnosis.
• Provide real-time analysis in remote areas lacking skilled practitioners.
• Facilitate telemedicine workflows through a user-friendly interface that allows non-experts to record and transmit data to doctors.

The paper concludes that while the current results are promising, the system requires large-scale clinical testing and further data expansion across various hospitals to fully validate its effectiveness and refine the model for broader deployment. The authors maintain a modest tone, acknowledging that the current study is a proof-of-concept for a low-cost, AI-empowered diagnostic tool.