The Vote Hides the Failure: Aggregation Choice and Noise Robustness in Heart Murmur Detection

Q: What are the key findings of this study?

BiLSTM outperformed HMS-Net across all 16 noise conditions under matched majority-vote aggregation (accuracy 80.9% vs 21.2% at 0dB AWGN), while native aggregation masked window-level disagreement.

Q: What is the clinical evidence from this study?

Population: Heart murmur (n=942). Intervention: Bidirectional Long Short-Term Memory (BiLSTM) pipeline vs. Hierarchical Multi-Scale Convolutional Network (HMS-Net) pipeline. Primary outcome: Murmur classification Accuracy under 0dB Additive White Gaussian Noise (AWGN) with matched majority-vote aggregation.

Q: What does this research mean for the field?

The BiLSTM phonocardiogram murmur detection pipeline outperforms the HMS-Net pipeline across multi-severity noise conditions under matched aggregation, but standard aggregate metrics and voting rules mask underlying window-level degradation and training failures. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

Nicholaus Ladislaus; Damilare Emmanuel Olatunji; Chol Buol

doi:10.48550/arxiv.2609.37161

September 29, 2026Open Access

The Vote Hides the Failure: Aggregation Choice and Noise Robustness in Heart Murmur Detection

View Full Paper

Ask AI

Bookmark

Share

Key result

BiLSTM outperforms HMS-Net in noisy murmur detection, achieving ~60-point higher accuracy under 0dB noise.

n=942

Why the study?

Noise robustness in automated phonocardiogram murmur detection and how it is measured remain underexamined despite growing interest in low-resource screening.

Does the choice of prediction aggregation rule and evaluation metric alter noise-robustness conclusions for deep learning phonocardiogram heart murmur detection pipelines (HMS-Net and BiLSTM)?

Key Points

To examine how aggregation rules and evaluation metric choices influence reported noise robustness in automated phonocardiogram heart murmur detection pipelines.
Reimplemented two deep learning pipelines—Hierarchical Multi-Scale Convolutional Network (HMS-Net) and Bidirectional Long Short-Term Memory (BiLSTM)—using the CirCor DigiScope dataset.
Tested both pipelines under four synthetic noise types across multiple severities using native aggregation versus matched majority-vote aggregation, followed by noise-augmented fine-tuning.
Under matched majority-vote aggregation across all 16 noise conditions, the BiLSTM pipeline outperformed HMS-Net in Accuracy and Weighted Accuracy, with HMS-Net accuracy dropping to 0.212 versus BiLSTM 0.809 at 0 dB additive white Gaussian noise.
HMS-Net native aggregation masked window-level errors, exceeding majority-vote accuracy by 7.3, 16.8, and 13.4 percentage points at 1%, 5%, and 10% salt-and-pepper corruption.
Noise-augmented fine-tuning increased BiLSTM majority-vote Weighted Accuracy across all 8 evaluated conditions (+2.1 to +5.3 percentage points, two cells p < 0.001) while native accuracy simultaneously declined (-0.05 to -3.3 percentage points).

Structured PICO

Does the choice of prediction aggregation rule and evaluation metric alter noise-robustness conclusions for deep learning phonocardiogram heart murmur detection pipelines (HMS-Net and BiLSTM)?

P

Population

942 patients (3,046 phonocardiogram recordings) from the CirCor DigiScope dataset evaluated for heart murmur detection (19.0% Present, 7.2% Unknown, 73.8% Absent).

I

Intervention

Bidirectional Long Short-Term Memory (BiLSTM) pipeline and Hierarchical Multi-Scale Convolutional Network (HMS-Net) pipeline evaluated under controlled multi-severity synthetic noise (Additive White Gaussian, pink, uniform at 20, 10, 5, 0 dB; salt-and-pepper at 1, 3, 5, 10%) with noise-augmented fine-tuning (uniform-ablation and vulnerability-informed variants) and evaluated using native versus majority-vote (MV) aggregation.

C

Comparator

Comparison between HMS-Net and BiLSTM pipelines, native aggregation versus matched majority-vote (MV) aggregation, and baseline models versus noise-augmented fine-tuned models.

O

Outcome

Classification performance for heart murmur detection measured by Accuracy, Weighted Accuracy (W.acc), and Balanced Accuracy across noise conditions and aggregation methods.surrogate

Main Result

Absolute Event Rate: 80.9% vs 21.2%

In automated phonocardiogram heart murmur detection, patient-level aggregation rules and accuracy metric choices can mask window-level prediction failures or artificially inflate improvements after noise-augmented training, demonstrating that evaluation conventions critically shape noise-robustness conclusions.

Limitations

Per-recording predicted-class distributions across severity are not reported
Both pipelines' native aggregation is Present-biased by construction, contributing a baseline gap independent of noise
Preprocessing alters effective input SNR differently between pipelines, preventing isolation of architecture versus preprocessing effects
Per-fold seed variance cannot be separated from fold variance since each fold used one seed
Unknown-class recall estimates rest on only 10-16 patients per fold and are noisy
All injected noise is synthetic and may not transfer to real-world acoustic interference
Comparison cannot isolate whether robustness differences stem from architecture, preprocessing, or their interaction without a matched-SNR comparison or preprocessing-stage ablation.
Per-recording predicted-class distributions across severity are not reported.
Per-fold seed variance cannot be separated from fold variance because each fold used one seed.
Unknown-class recall estimates rest on only 10–16 patients per fold and are noisy.
All injected noise is synthetic and may not transfer to real-world interference.

Ask about this paper

Answers are grounded in this paper and cite their sources.

Implication

Aggregation strategies can mask true model performance in noisy murmur detection; leaves open standardized evaluation before clinical adoption.

Abstract

Noise robustness in automated phonocardiogram (PCG) murmur detection, and how it is measured, remains underexamined despite growing interest in low-resource screening. We evaluate two independently reimplemented pipelines, Hierarchical Multi-Scale Convolutional Network (HMS-Net)--CNN, and Bidirectional Long Short-Term Memory (BiLSTM)--LSTM, under controlled, multi-severity noise with noise-augmented fine-tuning and held-out generalization testing. Under matched aggregation, the complete BiLSTM pipeline outperforms the complete HMS-Net pipeline across all conditions in accuracy and Weighted Accuracy. A stable aggregate accuracy score can misrepresent what individual predictions show: HMS-Net's native aggregation degrades under salt-and-pepper noise far less than majority-vote (MV) aggregation at the same severity, a gap reflecting window-level disagreement its native rule absorbs, while BiLSTM's MV accuracy rises after noise-augmented training even though its individual predictions do not improve. HMS-Net's training effect is significant under one accuracy metric but not another. Noise-robustness conclusions can depend as much on evaluation choices as on the models themselves.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Ladislaus et al. (2026) studied Heart murmur (n=942). Bidirectional Long Short-Term Memory (BiLSTM) pipeline vs. Hierarchical Multi-Scale Convolutional Network (HMS-Net) pipeline was evaluated on Murmur classification Accuracy under 0dB Additive White Gaussian Noise (AWGN) with matched majority-vote aggregation. BiLSTM outperformed HMS-Net across all 16 noise conditions under matched majority-vote aggregation (accuracy 80.9% vs 21.2% at 0dB AWGN), while native aggregation masked window-level disagreement.

synapsesocial.com/papers/6ae28dedcbadf70daf6c3f98 https://doi.org/10.48550/arxiv.2609.37161

View Full Paper

Ask AI

Bookmark

Share

The Vote Hides the Failure: Aggregation Choice and Noise Robustness in Heart Murmur Detection

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

The Vote Hides the Failure: Aggregation Choice and Noise Robustness in Heart Murmur Detection

Key Points

Structured PICO

Main Result

Limitations

Ask about this paper

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading