Selection of Heart Sound Segments for Synchronous Classification of Multi-channel Heart Sounds
First work to classify patients via synchronous multi-channel heart sound analysis, mirroring how physicians perform cardiac auscultation at four spots simultaneously Introduces a segment selection algorithm that identifies optimal heart sound recordings from each of the four auscultation locations Combines the selection algorithm with a multi-input CNN to process all four channels concurrently rather than individually Achieves 96.5% overall accuracy on the CirCor DigiScope dataset (735 patients
Analysis
TL;DR
- First work to classify patients via synchronous multi-channel heart sound analysis, mirroring how physicians perform cardiac auscultation at four spots simultaneously
- Introduces a segment selection algorithm that identifies optimal heart sound recordings from each of the four auscultation locations
- Combines the selection algorithm with a multi-input CNN to process all four channels concurrently rather than individually
- Achieves 96.5% overall accuracy on the CirCor DigiScope dataset (735 patients), representing a 9.1% improvement over the best single-channel and asynchronous multi-channel baselines
- The proposed selection strategy significantly outperforms random segment selection (paired statistical test, p = 0.003)
Why It Matters
This work bridges a critical gap between clinical practice and automated cardiac screening by introducing synchronous multi-channel analysis—the first approach that replicates how physicians listen to heart sounds at multiple anatomical spots simultaneously. For AI practitioners working in medical diagnostics, it demonstrates that modeling inter-channel relationships can yield substantial accuracy gains over naive multi-channel aggregation, offering a blueprint for multi-sensor fusion in other clinical domains.
Technical Details
- Problem framing: Cardiac auscultation requires listening at four main anatomical spots; prior automatic systems either used a single channel or analyzed multiple channels independently (asynchronously), missing potential inter-channel correlations.
- Segment selection algorithm: A novel algorithm identifies the most informative heart sound segments from each of the four auscultation spots, replacing ad-hoc or random segment choices.
- Multi-input CNN architecture: The selected segments from all four channels are fed simultaneously into a multi-input convolutional neural network, enabling the model to capture cross-channel interference phenomena during classification.
- Dataset: Evaluated on the CirCor DigiScope dataset comprising 735 patients with complete four-spot recordings.
- Evaluation: Paired statistical significance testing confirms the selection algorithm's advantage over random selection (p = 0.003); the synchronous approach achieves 96.5% accuracy, a 9.1% gain over strongest single-channel and asynchronous baselines.
Industry Insight
- Synchronous multi-sensor fusion should be considered a standard design principle in clinical AI systems where multiple spatially distributed signals are available, as asynchronous aggregation discards valuable cross-channel information.
- The segment selection strategy highlights the importance of input curation in medical signal processing—model architecture alone is insufficient; intelligent preprocessing that mirrors clinical reasoning can drive significant performance gains.
- As digital stethoscopes and wearable cardiac monitors become more prevalent, this synchronous multi-channel paradigm is directly applicable to real-world screening pipelines, particularly in resource-limited settings where cost-effective cardiovascular disease screening is most needed.
Disclaimer: The above content is generated by AI and is for reference only.