Temporal convolutional neural network-based feature extraction and asynchronous channel information fusion method for heart abnormality detection in phonocardiograms

Background and Objective

Auscultation-based cardiac abnormality detection is valuable screening approach in pediatric populations, particularly in resource-limited settings. However, its clinical utility is often limited by phonocardiogram (PCG) signal variability and a difficulty in distinguishing between pathological and innocent murmurs.

Methods

We proposed a framework that leverages temporal convolutional network (TCN)-based feature extraction and information fusion to integrate asynchronously acquired PCG recordings at the patient level. A probabilistic representation of the pathological state was first extracted from segmented PCG signals using a TCN-based model. These segment-level representations were subsequently averaged to generate record- or patient-level features. The framework was designed to accommodate recordings of varying durations and different auscultation locations. Furthermore, we addressed domain adaptation challenges in cardiac abnormality detection by incorporating transfer learning techniques.

Results

The proposed method was evaluated using two large, independent public PCG datasets, demonstrating robust performance at both record and patient levels. While its initial performance on an unseen external dataset was modest, likely due to demographic characteristics and signal acquisition, transfer learning significantly improved the model's performance, yielding an area under the receiver operating characteristic curve of 0.931±0.027 and an area under the precision-recall curve of 0.867±0.064 in external validation. Combining internal and external datasets further enhanced model generalizability.

Conclusion

This proposed framework accommodates multi-channel, variable-length PCG recordings, making it a flexible and accurate solution for detecting pediatric cardiac abnormalities, particularly in low-resource settings. The source code is publicly available on Github (https://github.com/baporlab/pcg_pathological_murmur_detection).