SpectroNet-LSTM: An interpretable deep learning approach to cardiac anomaly detection through heartbeat sound...

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Cardiac anomalies pose significant health risks, necessitating early detection for improved patient outcomes.
• Current diagnostic methods for heart valve diseases often require specialized expertise and equipment.
• Over 13 million individuals annually are affected by heart valve diseases, highlighting the need for advanced diagnostic tools.

Purpose of the Study:

• To develop an automated framework, SpectroNet-LSTM, for detecting cardiac anomalies from heartbeat sound recordings.
• To enhance the interpretability of automated cardiac anomaly detection systems for clinical use.
• To leverage deep learning and advanced feature extraction for improved diagnostic accuracy.

Main Methods:

• Utilized Mel-frequency cepstral coefficients (MFCCs) and spectrogram analysis for acoustic feature extraction.
• Trained deep learning models including ResNet101, VGG16, and Inception V3 on extracted heartbeat features.
• Integrated Explainable AI (XAI) techniques, SHAP and LIME, for model interpretability.

Main Results:

• The SpectroNet-LSTM model demonstrated superior performance compared to benchmarks in detecting heart anomalies.
• The system successfully captured critical acoustic features for accurate anomaly identification.
• Explainable AI techniques provided visualization and understanding of the model's decision-making process.

Conclusions:

• SpectroNet-LSTM offers a novel, automated, and interpretable solution for cardiac anomaly detection.
• The integration of feature extraction, deep learning, and XAI enhances cardiovascular diagnostics.
• This research promotes accessible healthcare solutions and efficient patient outcomes globally through automation.