PCSCN: A quality assessment model for large-scale phonocardiogram data cleaning and enhancement

Background:

The acquisition of large-scale phonocardiogram (PCG) data is crucial for clinical research and has significantly advanced the application of data-driven heart sound classification models. Owing to the complexity of clinical environments, the quality of collected PCG data cannot be guaranteed. The exclusion of low-quality signals is essential for conducting a reliable PCG analysis.

Methods:

This study introduces a novel quality assessment model, the Parallel Channel Sequence Convolutional Network (PCSCN). The PCSCN automatically and accurately detects and removes low-quality PCG signals, thereby improving dataset reliability and usability. Unlike previous methods that rely on manually extracting numerous statistical features, the PCSCN employs low-complexity features and a multi-channel sequence architecture, offering greater accuracy and efficiency. In addition, this study applies the PCSCN in a controlled experiment for database cleaning and enhancement to verify the impact of this process on downstream data-driven PCG classification models.

Results:

When tested on a public dataset, PCSCN achieved an accuracy of 95.45% and an F1-score of 95.44%. In a database cleaning task involving PCG data from 7220 subjects, PCSCN completed the task in 515.76s. Furthermore, in the controlled experiment, the PCG classification model trained with PCSCN-enhanced data demonstrated enhanced performance across multiple metrics.

Conclusions:

The PCSCN is both reliable and efficient, and is poised to play a pivotal role in the cleaning of large-scale PCG databases. By enhancing and improving the quality of the training data, the PCSCN significantly strengthens the clinical decision-making capabilities of PCG classification models, thereby elevating their value for clinical research and application.