Source-linked AI summary
Selection of Heart Sound Segments for Synchronous Classification of Multi-channel Heart Sounds
Marcelo Nogueira, Jorge H. Oliveira, Carlos F. Ferreira, Miguel T. Coimbra, Alípio M. Jorge
TL;DR
The paper addresses whether synchronous analysis of four heart-sound channels can improve patient classification over single-channel and asynchronous approaches. It selects representative segments for each auscultation spot and feeds their MFCC features into a multi-input CNN. The approach achieves 96.54% overall accuracy, while the study’s scope is limited by incomplete recordings and baseline capacity differences.
Problem
The study asks whether synchronous four-channel analysis outperforms single-channel and asynchronous multi-channel heart-sound classification.
Method
A clustering-based algorithm selects one representative three-second segment per auscultation spot for simultaneous processing by a multi-input CNN.
Results
96.54% overall accuracy was achieved by the multi-input CNN with segment selection, exceeding the best single-channel and asynchronous results reported in the study.
Takeaways & Limitations
Synchronous multi-channel analysis and representative segment selection were associated with higher patient-classification performance than the evaluated single-channel, asynchronous, and random-selection approaches.
Takeaways & Limitations
Analysis was restricted to 735 of 1,568 patients with complete four-spot recordings, and baseline classifiers had considerably smaller capacity than the multi-input CNN.
Abstract
from arXiv · showhide
Cardiac auscultation remains the most cost-effective screening procedure for cardiovascular diseases, and requires listening at the four main auscultation spots. Despite this, automatic heart sound analysis algorithms mostly classify patients using a single heart sound (single-channel), or, when using more than one (multi-channel), analyze each channel individually. To our knowledge, no prior work classifies patients through the synchronous analysis of multi-channel heart sounds, following the procedure used by physicians. This motivates us to study whether synchronous multi-channel analysis outperforms single-channel approaches, and whether it holds an advantage over asynchronous multi-channel methods that analyze channels one by one, potentially by capturing inter-channel interference phenomena. To answer these questions, we introduce a selection algorithm that identifies optimal heart sound segments from each of the four auscultation spots, which are then fed into a multi-input CNN that classifies patients by analyzing the four selected sounds simultaneously. Our synchronous approach, combining the proposed selection algorithm with a multi-input CNN, achieves a superior overall accuracy of 96.5\%, a 9.1\% gain over the best-performing single-channel and asynchronous multi-channel methods. The benefit of the proposed segment selection strategy over random selection is confirmed by a paired statistical significance test ($p = 0.003$). These results were obtained on 735 patients from the CirCor DigiScope dataset with complete recordings from all four spots, and their scope and generalizability are discussed in light of this and other methodological considerations.
1 Introduction
The paper motivates synchronous analysis of four auscultation channels because clinical auscultation is multi-site, while prior automated methods largely use single-channel or asynchronous approaches. It evaluates whether synchronous fusion and representative segment selection improve patient classification.
- 32% of all deaths globally are attributed to cardiovascular diseases, motivating reliable low-cost screening methods.The paper describes cardiac auscultation as cost-effective and non-invasive, but dependent on physician skill and experience.
- Clinical auscultation listens at four primary locations: aortic, pulmonary, tricuspid, and mitral valves.These locations are denoted AV, PV, TV, and MV.
- Prior public datasets lacked synchronous recordings from all four auscultation spots, limiting algorithms that mimic comprehensive clinical examination.The CirCor DigiScope Dataset provides recordings from each auscultation spot for most patients.
- The proposed approach uses a multi-input CNN to analyze data from all four auscultation spots simultaneously.This mirrors the physician’s holistic approach to cardiac auscultation.
- The study asks whether synchronous multi-channel analysis outperforms single-channel and asynchronous methods, potentially by capturing inter-channel interference.It also examines whether representative three-second segment selection improves accuracy and robustness.
2 The PCG Dataset
The study uses the CirCor DigiScope Dataset, collected during Brazilian mass screening campaigns, and restricts analysis to patients with complete recordings from all four auscultation spots.
- The dataset contains 5,277 asynchronously collected recordings from 1,568 participants across the four main auscultation spots.Recordings were collected during mass screening campaigns in Brazil.
- 75% of individuals had normal heartbeats and 25% had diagnosed cardiac conditions.The dataset was recorded in an ambulatory environment containing various noisy components.
- 735 patients met the inclusion criterion of complete recordings from all four auscultation spots.The remaining 833 patients were excluded because recordings were available from only one, two, or three spots.
3 Methodology
The methodology preprocesses and segments heart sounds, extracts MFCC features, selects one representative segment per auscultation spot, and classifies the four selected inputs synchronously with a multi-input CNN.
- Pre-processing and feature preparation: The pipeline applies noise removal, three-second segmentation, MFCC extraction, representative segment selection, classifier training, and evaluation.Segmentation uses provided annotations, and one segment is selected for each auscultation spot.
- Pre-processing and feature preparation: Three-second segments start from the S1 state and contain at least one full cardiac cycle.Segmenting 2,940 PCG signals produced 49,833 three-second segments.
- Pre-processing and feature preparation: Each three-second segment is converted into a 6 x 300 MFCC feature matrix using overlapping 25 ms windows with 10 ms steps.Six MFCCs are computed for each of 300 time frames.
- Segment Selection: K-means clustering selects the representative segment by choosing the largest cluster and then the sample nearest its centroid.Silhouette analysis determines K before clustering segments from one heart sound.
- Segment Selection: The selected segment is intended to reduce the influence of noisy or unrepresentative portions of recordings.Noise may arise from stethoscope movement and contaminate the beginning, end, or middle of recordings.
- Multi-input CNN: Each patient contributes four selected segments, one per auscultation spot, to the multi-input CNN.Using one segment per spot also simplifies the model compared with approaches requiring dozens of segments per patient.
- Multi-input CNN: The CNN feeds four MFCC matrices into separate parallel branches, concatenates their flattened outputs, and classifies the patient through fully connected layers.The selected architecture contains 9,287,426 parameters and was trained with ten-fold stratified cross-validation.
4 Experimental Setup and Evaluation Metrics
The evaluation compares synchronous multi-input CNN classification with single-channel, asynchronous multi-channel, and random-selection baselines using patient-level cross-validation and standard diagnostic metrics.
- Experimental protocol: 735 patients are represented by four three-second segments, each converted into a 300×6 MFCC matrix for the classification network.The experimental protocol uses one segment per auscultation spot for the proposed model.
- Experimental protocol: Patient-level ten-fold stratified cross-validation preserves the approximately 80% normal and 20% abnormal class distribution across folds.This addresses the dataset’s class imbalance during evaluation.
- Baselines: The single-channel baseline uses one heart sound per patient and classifies segment-level outputs with SVM, Random Forest, XGBoost, and CNN classifiers.Patient classification labels a patient abnormal if any segment is classified as abnormal.
- Baselines: The asynchronous multi-channel baseline processes all four heart sounds independently and aggregates sound-level predictions by majority vote.It uses the same classifier family at the segment level.
- Baselines: The random-selection variant performs synchronous four-channel analysis but replaces representative segment selection with randomly selected segments.This isolates the contribution of the proposed selection strategy.
- Evaluation metrics: Sensitivity, specificity, and overall accuracy are defined from TP, TN, FP, and FN for comparison with prior evaluation standards.The metrics align with studies using the PhysioNet/CinC Challenge 2016 Database.
5 Results and Discussion
The synchronous multi-input CNN with selected segments achieved the strongest reported classification performance, outperforming single-channel and asynchronous multi-channel approaches. Segment selection further improved accuracy, although differences in model capacity and the restricted dataset limit how specifically the gains can be attributed and generalized.
- Results: 96.54% overall accuracy was achieved by the multi-input CNN with segment selection, establishing a benchmark for synchronous multi-channel heart sound classification.No prior work had explored this synchronous approach, according to the reported results.
- Comparison with single-channel methods: 14 points separated the multi-input CNN from the best single-channel method, XGBoost at 82.5%.The authors attribute this gain to capturing murmur-related interference patterns across channels.
- Comparison with asynchronous multi-channel methods: 96.54% versus 87.5% showed that the synchronous multi-input CNN outperformed the best asynchronous multi-channel SVM.Because both methods used the same four heart sounds, the authors suggest the gap reflects inter-channel interference captured synchronously.
- Segment selection: 3.7 points improved accuracy over random segment selection, increasing performance from 92.8% to 96.5%.The authors associate careful selection with reduced inter-segment variability and minimized noise.
- Statistical analysis: p = 0.003 supported a significant advantage for cluster-based over random segment selection across 10 paired cross-validation folds.The paired t-test gave p = 0.004, with Cohen’s d = 0.49.
- Class-imbalance evaluation: 0.955 sensitivity and 0.976 specificity were reported for the selected-segment multi-input CNN despite the dataset’s ≈80% normal versus 20% abnormal imbalance.A patient-level qualitative analysis of misclassified murmur types and pathologies was outside the study’s scope.
6 Conclusion
The study combines representative segment selection with synchronous four-channel CNN analysis, supporting gains over alternative approaches while identifying scope and capacity-related caveats.
- 6 Conclusion: 96.5% overall accuracy was achieved by the multi-input CNN, outperforming baseline approaches through synchronous analysis of four auscultation spots.The model integrates information from multiple anatomical perspectives and may capture inter-channel interference phenomena such as murmur wave patterns.
- 6 Conclusion: 3.7 percentage points separated selected segments from random segments, with the improvement statistically significant by paired tests.Performance increased from 92.8% with random selection to 96.5% with selection; the Wilcoxon test gave p = 0.003 and the paired t-test gave p = 0.004.
- 6 Conclusion: The findings support better performance for multi-channel over single-channel methods, synchronous over asynchronous processing, and selected over random segments.The study reports these three hypotheses as validated by its experiments and statistical testing.
- 6 Conclusion: The conclusions are limited to 735 patients from a single dataset and the experimental constraints discussed by the authors.External validation on additional multi-spot datasets is identified as a way to strengthen generalizability.
- 6 Conclusion: Future work will test heterogeneous feature types, attention mechanisms, matched-capacity baselines, and additional multi-spot datasets.These directions aim to simplify segment selection, assess alternative architectures, separate model capacity from synchronous fusion, and improve generalizability.
Declarations
The declarations report no competing interests, use of a publicly accessible de-identified dataset under applicable ethical principles, and no clinical trial registration.
- Declarations: The authors declare no competing interests.
- Declarations: The study used the publicly accessible CirCor DigiScope Dataset without individual identifiers, so specific consent was not required.The authors state that the work adheres to ethical principles and guidelines for publicly available datasets.
- Declarations: No clinical trial number applies to this study.