Scaling to Multimodal and Multichannel Heart Sound Classification with Synthetic and Augmented Biosignals

Kayapanda Mandana; Matthew Fynn; Milan Marocchi; Yue Rong

REVIEW 4 major objections 4 minor 42 references

Scaling to Multimodal and Multichannel Heart Sound Classification with Synthetic and Augmented Biosignals

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Wav2Vec2 classifier fine-tuned on synthetic and augmented heart sounds reaches state-of-the-art accuracy on abnormal heart sound detection across single-channel, multimodal, and multichannel inputs.

desk verdict Solid applied ML with a load-bearing evaluation risk: the SOTA numbers hinge on a patient-level split that is never stated. read the letter →

arxiv 2509.11606 v4 pith:HM73O6WN submitted 2025-09-15 cs.SD cs.LGeess.SP

Milan Marocchi , Matthew Fynn , Kayapanda Mandana , Yue Rong This is my paper · ORCID

classification cs.SDcs.LGeess.SP

keywords heartsoundclassificationphonocardiogramWav2Vec2diffusionmodelsdataaugmentationmultimodallearningmultichannelPCGelectrocardiogram

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to show that data scarcity, not architecture, is what holds transformer-based heart-sound classifiers back. By combining traditional signal augmentations with synthetic heart-sound recordings generated by diffusion models, and fine-tuning a Wav2Vec2 encoder, the authors report state-of-the-art results on three setups: single-channel heart sounds, synchronised heart-sound and ECG signals, and six-channel heart sounds from a wearable vest. The practical payoff would be a single scalable pipeline that can take any number of heart-sound channels plus ECG as input and support automated pre-screening for cardiovascular disease. The method is aimed at realistic, noisy, free-breathing conditions, where prior work often required breath-holding and manual heart-cycle labels.

What carries the argument

The argument rests on coupling a Wav2Vec2 BASE encoder with an enlarged training set produced by two diffusion models and a multichannel-aware augmentation pipeline. WaveGrad and DiffWave synthesize PCG waveforms conditioned on ECG mel-spectrograms and disease labels; for multichannel data, the conditioning label specifies the source channel and the target channel being generated. The classifier concatenates features extracted from each input channel and passes them through an MLP, with LoRA fine-tuning and an SVM added for the small vest dataset. The model is trained in stages that alternate original, synthetic, and augmented data to prevent overfitting.

What would settle it

Inspect the CinC 2016 split by listing patients in each train and test partition; if any patient contributes recordings to both, recompute subject-level accuracy, UAR, and MCC after grouping all recordings by patient. Also check whether the conditioning ECG or PCG inputs used to train WaveGrad and DiffWave include any recordings from test-set subjects.

Watch

Extended reading notes

Core claim

The central discovery is that a raw-signal transformer encoder, normally too data-hungry for small medical audio sets, becomes state of the art once its training set is enlarged with a mix of traditional augmentations and synthetic PCG waveforms generated by WaveGrad and DiffWave. On the CinC 2016 single-channel PCG dataset, the paper reports subject-level accuracy of 92.48%, UAR of 93.05%, sensitivity of 93.63%, specificity of 92.48%, and MCC of 0.8283. On the synchronised PCG+ECG training-a subset, it reports 93.14% accuracy, 92.21% UAR, 94.35% sensitivity, 90.10% specificity, and MCC 0.8380. On a six-channel wearable vest mPCG dataset with 157 subjects, it reports 77.13% accuracy, 74.25%

Load-bearing premise

The reported subject-level numbers assume the 60/20/20 split separates patients, not just recordings, and that the diffusion generators never see test-subject data; the text never explicitly states either condition.

Editorial extensions

If this is right

Transformer-based raw-signal classifiers can be made practical for small medical audio datasets when supported by synthetic and augmented training data.
The same architecture scales from single-channel PCG to PCG+ECG to multichannel PCG with only feature-concatenation changes, supporting a unified pre-screening model.
Diffusion-based synthetic data makes the multimodal model trainable: without augmentation, the PCG+ECG model collapsed to roughly 51% UAR with nearly zero specificity.
The strong results on noisy, free-breathing vest data suggest the method tolerates realistic acquisition conditions better than prior work that required breath-holding and manual heart-cycle segmentation.
Sampling-rate tuning matters: 4.125 kHz matched to relevant heart-sound frequency bands outperformed 16 kHz on the multimodal and vest datasets, while 16 kHz worked best on the full CinC dataset.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If patient-level separation in the reported splits holds, a natural testable projection is that the augmentation benefit should grow as dataset size and signal quality shrink; ablating synthetic-data volume on the vest dataset could confirm this ordering.
The channel-to-channel conditioning scheme for the multichannel diffusion model could transfer to other multi-sensor physiological arrays, such as multi-lead ECG, where the model learns inter-channel dependencies rather than relying on hand-specified relationships.
A focused experiment generating synthetic-only training sets and probing transfer could reveal whether the gain comes from added morphological variability or from regularisation of the Wav2Vec2 encoder; the paper does not disentangle these two effects.
The reported near-parity with breath-held prior work under free-breathing, automatically segmented conditions suggests that the method's practical value may be higher than raw metric comparisons imply, since acquisition burden is much lower.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a genuinely useful applied paper with a load-bearing evaluation risk. The headline numbers should not be taken at face value until the authors clarify whether their splits are patient-disjoint and whether the diffusion generators ever saw test subjects.

What's actually new: they take the Wav2Vec2 + diffusion augmentation recipe from their prior work [32] and extend it in two directions: a multichannel PCG diffusion model that uses channel-conditioning labels, and a Wav2Vec2 classifier that fuses PCG+ECG or multiple PCG channels. That combination is not in the cited literature. The internal comparisons are consistent and large -- the unaugmented multimodal baseline collapses (TNR around 2%) while the augmented version reaches TNR around 90%, and vest MCC rises from 0.28 to 0.51. That gives real support for the claim that augmentation is doing the work. The architecture is clearly described and the training schedules are detailed enough to reproduce with effort.

The soft spots are real but mostly addressable. The main one: the CinC 2016 dataset has 3153 recordings from 764 patients, and the paper says only 'A 60-20-20 split is used.' It never states that all recordings from one patient stay in one split. Since subject-level metrics are computed by averaging fragment predictions per subject, recording-level splits would let the model exploit patient identity and inflate the reported 92.48% accuracy and 0.8283 MCC. The same ambiguity applies to the synthetic data: Section 2.1.1 says synthetic data is generated from training-a, but Section 3.2.1 says single-channel synthetic PCG was conditioned on Icentia ECG, and Section 3.2.2 says mPCG synthetic data used training-a and training-b. The paper doesn't say whether the diffusion generators were trained only on the training split. If they saw test subjects, the test set is indirectly observed. No code or data is released, so this can't be checked. The SOTA comparisons in Table 19 compare against a different vest dataset (fewer subjects, breath-held, hand-annotated), so the 'near-SOTA' claim there is softer than the text implies.

Minor issue: the sampling-rate grid search may have been done on validation, but it's not stated explicitly; not a major problem.

Who should read this: people working on scarce physiological audio who want to see how far diffusion augmentation can push transformer fine-tuning. It's a useful data point even if the exact numbers are conditional. It deserves a serious referee, but the referee needs to push on the split and generation boundary. If the authors clarify that and release code, this becomes a solid contribution.

My recommendation: send it to peer review. Don't desk-reject. The core method and internal comparisons are valuable; the missing split details are fixable, and the paper should not be rejected for them without giving the authors a chance to respond.

Referee Report

4 major / 4 minor

Summary. The paper proposes a scalable Wav2Vec 2.0-based classifier for abnormal heart sound detection, fine-tuned on augmented and diffusion-generated synthetic biosignals. It reports state-of-the-art subject-level results on the CinC 2016 single-channel PCG dataset (accuracy 92.48%, UAR 93.05%, MCC 0.8283), on the training-a PCG+ECG subset (accuracy 93.14%, MCC 0.8380), and on a multichannel mPCG vest dataset (accuracy 77.13%, MCC 0.5082). The method combines traditional signal augmentations with WaveGrad/DiffWave synthetic signals, and is evaluated on single-channel, multimodal, and multichannel inputs. The central claim is that the proposed pipeline enables transformer-based models to achieve state-of-the-art performance on these tasks.

Significance. If the reported results are valid, this is a practically significant contribution: it demonstrates a scalable way to apply transformer-based raw-signal models to small, imbalanced medical audio datasets. The internal augmented-versus-original comparisons are consistent and large, which is a clear strength: the training-a multimodal baseline collapses (TNR ≈ 2%) while the augmented model reaches TNR ≈ 90%; the vest MCC rises from 0.28 to 0.51; and the CinC subject-level MCC rises from 0.70 to 0.83. The interpretability analyses are also a useful addition. However, the subject-level SOTA claims rest on an evaluation protocol whose patient-level separation and synthetic-data boundary are not fully specified. Confirming those details is essential before the headline numbers can be accepted.

major comments (4)

[Section 2.1.1, Tables 12 and 15] The evaluation protocol does not state that the 60/20/20 split is patient-disjoint. CinC 2016 contains 3153 recordings from 764 patients, so multiple recordings per patient exist. Since subject-level results are obtained by averaging fragment predictions per subject, a recording-level split would allow the model to see fragments from the same patient in both training and test sets, inflating accuracy, UAR, and MCC. Please state explicitly that all recordings of each patient are confined to a single split, for all five shuffled splits, and confirm the same holds for the training-a subset. If this was not done, the experiments should be rerun with patient-disjoint splits.
[Section 2.1.1 vs Sections 3.2.1 and 3.2.2] The boundary of the synthetic-data pipeline is underspecified. Section 2.1.1 says synthetic data is generated from training-a, but Section 3.2.1 says single-channel synthetic PCG was generated using ECG from the Icentia dataset, and Section 3.2.2 says synthetic mPCG was generated from training-a and training-b. It is not stated whether the diffusion generators were trained only on the training partition, or whether any test-subject recordings were used as conditioning inputs for generation. If a test subject's ECG/PCG was used to synthesize training signals, test information indirectly enters training. Please specify, for each experiment, exactly which data were used to train the generative models and which data were used as conditioning inputs for the generated training set.
[Table 13 and accompanying text; Table 16 and accompanying text] The literature-comparison tables mix metrics from different configurations. In Table 12, the 16kHz All subject-level configuration gives Acc=92.48, UAR=93.05, TPR=93.63, TNR=92.48, F1=94.93, MCC=0.8283, while the 4.125kHz All configuration gives Acc=92.98, UAR=92.48, TPR=91.58, TNR=93.35, F1=95.42, MCC=0.8064. Table 13's 'This study' row lists Acc=92.98, UAR=92.48, TPR=93.63, TNR=92.48, F1=94.93, which is not obtainable from any single configuration in Table 12. Similarly, the text after Table 16 claims the proposed model has the highest metric 'other than TPR', but the RNN row reports TNR=91.10% compared with the proposed model's 90.10%, so TNR is not highest. The SOTA claim should be tied to a single, clearly identified configuration and accurate comparison rows.
[Section 4.1, Table 12 vs Table 13; abstract vs Section 4] The headline numbers are inconsistent across the abstract, conclusion, and tables. The abstract and conclusion report subject-level accuracy 92.48% and UAR 93.05% for CinC, matching the 16kHz All row in Table 12. However, Table 13 reports Acc 92.98% for 'This study' while also reporting UAR 92.48%, which is the 4.125kHz configuration's UAR, not the 16kHz configuration's. This inconsistency makes it difficult to determine which configuration is being claimed as state of the art. Please harmonize the reported values across the abstract, body, and comparison tables.

minor comments (4)

[Abstract and Conclusion] The abstract lists six numerical values but names only five metrics ('accuracy, UAR, sensitivity, specificity and MCC'). The sixth value (94.93%) is F1; please name F1 explicitly in the metric list.
[Section 2.1.1] The text says 'out of 409 recordings, 405 contain both signal types (288 abnormal, 117 normal)'. The counts sum to 405, but 288/405 = 71.1% abnormal, which is inconsistent with Table 1's training-a abnormal percentage of 67.5%. Please clarify which numbers refer to the subset of 405 versus the full 409.
[Sections 3.3 and 3.4.1] Section 3.3 states that for training-a and CinC, synthetic signals are limited to two segments per recording, while Section 3.4.1 states that only three segments are extracted from each synthetic subject. Please make the segment counts consistent.
[Section 3.2.1] The text says WaveGrad and DiffWave each generated '3,200 patients’ data' using Icentia ECG as conditioners. This is useful detail, but it would help to state whether these synthetic subjects overlap with the CinC training subjects in any way, particularly for the training-a experiments.

Circularity Check

1 steps flagged · score 6.0 of 10

Subject-level CinC/training-a results are not shown to be patient-disjoint; as described, the shuffled 60/20/20 split can put the same patient's fragments in both training and test, making the reported subject-level 'prediction' partly a within-patient fit.

fitted input called prediction [Section 2.1.1 (Dataset), Section 4.1/Table 12 (subject-level results), Section 3.3]
"There are 3153 recordings sourced from 764 patients [15]. ... A 60-20-20 split is used for training, validation, and testing in both the combined and training-a subsets. ... Each model was trained and tested ten times over five shuffled train-validation-test splits. ... Final classification was based on averaging fragment-level predictions."

With 3153 recordings from 764 patients (~4.1 recordings/patient), a recording-level 60/20/20 shuffle places many patients' recordings in both training and test. The paper does not state a patient-disjoint split or that synthetic data (also 'generated from training-a') was restricted to the training split. Subject-level Table 12 metrics are obtained by averaging fragment predictions per patient; the model's training directly includes fragments from the same test patient. The 'subject-level prediction' is therefore partly a function of that patient's training input—recognition of seen patients rather than prediction for unseen subjects—so the reported SOTA for CinC and training-a reduces, as described, to a training-set fit. The vest 7-fold CV is less affected because it is stratified and su

full rationale

The paper's main empirical claim is evaluated on external benchmark datasets, and it compares against non-overlapping literature, so most of the derivation chain is self-contained rather than circular. Self-citations to [5], [13], [22], and [32] are method inheritance: the augmentation and diffusion pipeline is adapted from the authors' prior work, but the current paper validates it with its own original-vs-augmented comparisons and external benchmarks, so the self-citation is not load-bearing in the sense of forcing the result. No uniqueness theorem or ansatz is smuggled in via citation; the Wav2Vec2 and diffusion architectures are public and independently specified. The significant circularity concern is the split construction: as described, the 60/20/20 split is a shuffled recording split, not an explicitly patient-disjoint split. Since CinC has multiple recordings per patient, the subject-level test metrics can be inflated by training on other fragments from the same patients, making the 'subject-level prediction' partly equivalent to the training input. The paper also does not specify whether the generative model that produces synthetic training-a data was restricted to the training partition, which would compound the same leakage. Because this affects the two headline SOTA claims (full CinC and training-a) but not the multichannel vest result to the same degree, the circularity is partial rather than total, hence a score of 6.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claim rests on standard supervised-learning assumptions plus several domain assumptions about split hygiene, synthetic data validity, label accuracy, and preprocessing. The largest burden is the unstated patient-level split and synthetic generation restriction, because benchmark numbers would be materially inflated if either is violated. No new physical entities are introduced; the main modeling freedom is in hyperparameters, sampling rate, augmentation schedule, and generation settings, which are tuned per dataset and partly inherited from the authors' prior work [32].

free parameters (5)

Per-dataset input sampling rate = CinC 16 kHz; training-a and vest 4.125 kHz
Selected by grid search over 1 to 16 kHz in 125 Hz steps using validation MCC (Section 3.3). The best rate differs per dataset, so reported metric differences between sampling rates are partly a tuned hyperparameter effect.
Bayesian-optimized classifier hyperparameters = Per dataset, e.g. CinC lr 0.001, wd 4.11e-5, momentum 0.576, gamma 0.167; multimodal and mPCG differ (Tables 5, 7, 10)
Optuna ran 150 trials with 5 repeats each, optimizing average validation MCC. These dataset-specific settings directly affect the reported absolute metrics.
Augmentation probabilities and per-class augmentation counts = e.g. HPSS 75%, white noise 7.5%, time stretch 25%, AM 75%, baseline wander 75%, EQ 25%, clinical 50%; augments 60/30 or
Hand-set in Section 3.1 and Tables 3 and 8, following [32]. Class-specific augmentation ratios shift the operating point and can inflate sensitivity or specificity.
Segmentation and windowing parameters = 4 s windows for CinC/training-a, 2 s for vest; 0.25 s overlap; 2 to 3 synthetic segments each
Chosen manually in Section 3.3. Window length, overlap, and fragment averaging affect subject-level results, and no ablation is shown for these choices.
Synthetic generation settings = 3200 subjects per model, 3:1 normal/abnormal, cycle shuffle 25%, mel 1024/256/80
Section 3.2.1. These settings determine the synthetic training distribution and are inherited from [32] rather than independently optimized on the target datasets.

assumptions (6)

domain assumption Patient-level independence of the 60/20/20 splits in CinC and training-a.
CinC has multiple recordings per patient, but Section 2.1.1 does not specify that all recordings of a patient are kept within one split. The subject-level metrics depend on this.
domain assumption Synthetic data generation and diffusion training are restricted to the training partition.
Sections 2.1.1 and 3.2.2 state synthetic data comes from training-a and training-b, but do not say generation excludes the test split. If test subjects contribute to synthetic or training data, the augmentation result is partly self-fulfilling.
domain assumption The diffusion-generated PCG signals are label-consistent and clinically realistic enough to act as valid training data.
No independent evaluation of synthetic signal quality is provided; only downstream classifier metrics are shown, and the generation pipeline is inherited from [32].
domain assumption Normal labels in the vest dataset are correct for the 21 subjects under 35 who were not angiographically confirmed.
Section 2.1.2 assumes CAD risk is low under 35; subclinical disease could mislabel some normal recordings and weaken the normal class.
domain assumption Wav2Vec2 speech-pretrained representations transfer usefully to PCG and ECG.
Speech initialization is used throughout. The multimodal baseline without augmentation collapses, so transfer alone is not sufficient and its contribution is entangled with augmentation.
domain assumption Resampling, filtering to 25-400 Hz PCG and 2-60 Hz ECG, and min-max normalization preserve diagnostic content.
Section 3.3 discards frequency content above 400 Hz, while the paper notes murmurs can extend to 600 Hz. No ablation on the filter band is shown.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Scaling to Multimodal and Multichannel Heart Sound Classification with Synthetic and Augmented Biosignals." pith.science (2026). https://pith.science/paper/HM73O6WN

@misc{pith2026250911606,
  author       = {Pith},
  title        = {Pith review of: Scaling to Multimodal and Multichannel Heart Sound Classification with Synthetic and Augmented Biosignals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HM73O6WN}},
  note         = {Machine review of arXiv:2509.11606}
}

read the original abstract

Cardiovascular diseases (CVDs) are the leading cause of death worldwide, accounting for approximately 17.9 million deaths each year. Early detection is critical, creating a demand for accurate and inexpensive pre-screening methods. Deep learning has recently been applied to classify abnormal heart sounds indicative of CVDs using synchronised phonocardiogram (PCG) and electrocardiogram (ECG) signals, as well as multichannel PCG (mPCG). However, state-of-the-art architectures remain underutilised due to the limited availability of synchronised and multichannel datasets. Augmented datasets and pre-trained models provide a pathway to overcome these limitations, enabling transformer-based architectures to be trained effectively. This work combines traditional signal processing with denoising diffusion models, WaveGrad and DiffWave, to create an augmented dataset to fine-tune a Wav2Vec 2.0-based classifier on multimodal and multichannel heart sound datasets. The approach achieves state-of-the-art performance. On the Computing in Cardiology (CinC) 2016 dataset of single channel PCG, accuracy, unweighted average recall (UAR), sensitivity, specificity and Matthew's correlation coefficient (MCC) reach 92.48%, 93.05%, 93.63%, 92.48%, 94.93% and 0.8283, respectively. Using the synchronised PCG and ECG signals of the training-a dataset from CinC, 93.14%, 92.21%, 94.35%, 90.10%, 95.12% and 0.8380 are achieved for accuracy, UAR, sensitivity, specificity and MCC, respectively. Using a wearable vest dataset consisting of mPCG data, the model achieves 77.13% accuracy, 74.25% UAR, 86.47% sensitivity, 62.04% specificity, and 0.5082 MCC. These results demonstrate the effectiveness of transformer-based models for CVD detection when supported by augmented datasets, highlighting their potential to advance multimodal and multichannel heart sound classification.

Figures

Figures reproduced from arXiv: 2509.11606 by the authors.

**Figure 2.** mPCG of a patient with an abnormal heart condition. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Wav2Vec 2.0 encoder module architectures. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (35 more)

**Figure 4.** Figure 4: WaveGrad architecture. (a) UBlock module architecture. (b) DBlock module architecture. (c) FiLM module architecture [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: WaveGrad module architectures. 1.3.2. DiffWave DiffWave is a diffusion model for raw audio synthesis with both conditional and unconditional variants [7]. The model uses 1D convolutions and fully connected layers, with its core comprising bi-directional dilated convolu…

**Figure 6.** Figure 6: DiffWave architecture. 2. Materials All data processing and model training were conducted using a Ryzen 7 3800X CPU and an Nvidia RTX 3090 (24 GB), with Python 3.10 and PyTorch 2.1.2. Diffusion models were trained on an RTX 4090 using the vast.ai cloud service. 2.1. Da…

**Figure 7.** Figure 7: 3.1. Augmented Signals 3.1.1. Single channel augmentations The PCG and ECG augmentation process is shown in [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 7.** Figure 7: Classification model creation procedure. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: PCG and ECG augmentation procedure 10 [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

**Figure 9.** Figure 9: Multichannel PCG augmentation procedure 11 [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]

**Figure 10.** Figure 10: Synthetic signal generation procedure. Other than this modification, the process for training the multichannel diffusion models remained the same. Synthetic mPCG signals were generated from conditioning PCG data from the CinC 2016 dataset, utilising training-a and tra…

**Figure 11.** Figure 11: Single input model architecture [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 12.** Figure 12: PCG and ECG model architecture. This model was trained using the SGD optimiser with an exponential learning rate decay, with the same hyperparameters as the single channel model. The initial hyperparameters of the multimodal models are found in [PITH_FULL_IMAGE:figur…

**Figure 13.** Figure 13: Multichannel PCG model architecture. The multichannel model followed the schedule in [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]

**Figure 14.** Figure 14: PCG 16kHz no augments CinC 2016 model ROC plots [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]

**Figure 15.** Figure 15: PCG 16kHz CinC 2016 model ROC plots (a) Fragment level. (b) Subject level [PITH_FULL_IMAGE:figures/full_fig_p020_15.png]

**Figure 16.** Figure 16: PCG 4.125kHz CinC 2016 model ROC plots Augmentation was shown to yield better ROC curves (less threshold sensitivity) versus the noaugment setting, along with less variation. Additionally, the area under the curve (AUC) is increased with augmentation, with 4.125kHz a…

**Figure 17.** Figure 17: PaCMAP for average no augment 16kHz CinC PCG model. [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]

**Figure 18.** Figure 18: PaCMAP for average 16kHz CinC PCG model. [PITH_FULL_IMAGE:figures/full_fig_p021_18.png]

**Figure 19.** Figure 19: PaCMAP for average 4.125kHz CinC PCG model. [PITH_FULL_IMAGE:figures/full_fig_p021_19.png]

**Figure 20.** Figure 20: CinC interpretability images for abnormal subject e01791. [PITH_FULL_IMAGE:figures/full_fig_p022_20.png]

**Figure 21.** Figure 21: CinC interpretability images for normal subject e01427. [PITH_FULL_IMAGE:figures/full_fig_p023_21.png]

**Figure 22.** Figure 22: PCG+ECG 16kHz no augments training-a model ROC plots [PITH_FULL_IMAGE:figures/full_fig_p024_22.png]

**Figure 23.** Figure 23: PCG+ECG 16kHz training-a model ROC plots [PITH_FULL_IMAGE:figures/full_fig_p024_23.png]

**Figure 24.** Figure 24: PCG+ECG 4.125kHz training-a model ROC plots [PITH_FULL_IMAGE:figures/full_fig_p024_24.png]

**Figure 25.** Figure 25: PaCMAP for average no augment 16kHz multimodal model. [PITH_FULL_IMAGE:figures/full_fig_p025_25.png]

**Figure 26.** Figure 26: PaCMAP for average 16kHz multimodal model. [PITH_FULL_IMAGE:figures/full_fig_p025_26.png]

**Figure 27.** Figure 27: PaCMAP for average 4.125kHz multimodal model. [PITH_FULL_IMAGE:figures/full_fig_p026_27.png]

**Figure 28.** Figure 28: CinC interpretability images for abnormal subject a0327. [PITH_FULL_IMAGE:figures/full_fig_p027_28.png]

**Figure 29.** Figure 29: CinC interpretability images for normal subject a0352. [PITH_FULL_IMAGE:figures/full_fig_p027_29.png]

**Figure 30.** Figure 30: mPCG 16kHz no augments vest dataset model ROC plots [PITH_FULL_IMAGE:figures/full_fig_p028_30.png]

**Figure 31.** Figure 31: mPCG 16kHz vest dataset model ROC plots 27 [PITH_FULL_IMAGE:figures/full_fig_p028_31.png]

**Figure 32.** Figure 32: mPCG 4.125kHz vest dataset model ROC plots [PITH_FULL_IMAGE:figures/full_fig_p029_32.png]

**Figure 33.** Figure 33: PaCMAP for average 16kHz mPCG model. 28 [PITH_FULL_IMAGE:figures/full_fig_p029_33.png]

**Figure 34.** Figure 34: PaCMAP for average no augment 16kHz mPCG model. [PITH_FULL_IMAGE:figures/full_fig_p030_34.png]

**Figure 35.** Figure 35: PaCMAP for average 4.125kHz mPCG model. Embedding separation is weaker than in CinC and training-a, consistent with dataset difficulty; nonetheless, augmentation improves clustering, which aligns with the higher metrics. The attention importance and GradCAM++ for a CA…

**Figure 36.** Figure 36: Vest data interpretability images for abnormal CAD subject. [PITH_FULL_IMAGE:figures/full_fig_p031_36.png]

**Figure 37.** Figure 37: Vest data interpretability images for a normal subject. [PITH_FULL_IMAGE:figures/full_fig_p031_37.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 4 linked inside Pith

[32]

Generative deep learning and signal processing for data augmentation of cardiac auscultation signals: Improving model robustness using synthetic audio,

L. Abbott, M. Marocchi, M. Fynn, Y. Rong, and S. Nordholm, “Generative deep learning and signal processing for data augmentation of cardiac auscultation signals: Improving model robustness using synthetic audio,” 2024. [Online]. Available: https://arxiv.org/abs/2410.10125 34

arXiv 2024
[1]

Cardiovascular Diseases (CVDs)

WHO, "Cardiovascular Diseases (CVDs)".Geneva, Switzerland: WHO, 2021

2021
[2]

Cardiac Auscultation: Rediscovering the Lost Art,

M. A. Chizner, “Cardiac Auscultation: Rediscovering the Lost Art,”Current Problems in Cardiology, vol. 33, no. 7, pp. 326–408, Jul. 2008

2008
[3]

The Lost Art of Clinical Skills,

C. A. Feddock, “The Lost Art of Clinical Skills,”The American Journal of Medicine, vol. 120, no. 4, pp. 374–378, Apr. 2007

2007
[4]

Accuracy of cardiac auscultation in detection of neonatal congenital heart disease by general paediatricians,

Q.-M. Zhao, C. Niu, F. Liu, L. Wu, X.-J. Ma, and G.-Y. Huang, “Accuracy of cardiac auscultation in detection of neonatal congenital heart disease by general paediatricians,”Cardiology in the Young, vol. 29, no. 5, pp. 679–683, May 2019

2019
[5]

Abnormal heart sound classification and model interpretability: A transfer learning approach with deep learning,

M. Marocchi, L. Abbott, Y. Rong, S. Nordholm, and G. Dwivedi, “Abnormal heart sound classification and model interpretability: A transfer learning approach with deep learning,”Journal of Vascular Diseases, vol. 2, no. 4, pp. 438–459, 2023. [Online]. Available: https://www.mdpi.com/2813-2475/2/4/34

2023
[6]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2017

2017
[7]

Diffwave: A versatile diffusion model for audio synthesis,

Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Diffwave: A versatile diffusion model for audio synthesis,” 2021

2021

Show all 42 references

[8]

Wavegrad: Estimating gradients for waveform generation,

N. Chen, Y. Zhang, H. Zen, R. J. Weiss, M. Norouzi, and W. Chan, “Wavegrad: Estimating gradients for waveform generation,” 2020

2020
[9]

Listen2yourheart: A self-supervised approach for detecting murmur in heart-beat sounds,

A. Ballas, V. Papapanagiotou, A. Delopoulos, and C. Diou, “Listen2yourheart: A self-supervised approach for detecting murmur in heart-beat sounds,” 2022

2022
[10]

Transfer learning based heart valve disease classification from phonocardiogram signal,

A. Maity, A. Pathak, and G. Saha, “Transfer learning based heart valve disease classification from phonocardiogram signal,”Biomedical Signal Processing and Control, vol. 85, p. 104805, 2023

2023
[11]

wav2vec 2.0: A framework for self-supervised learning of speech representations,

A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech representations,” 2020

2020
[12]

Wav2vec-based detection and severity level classification of dysarthria from speech,

F. Javanmardi, S. Tirronen, M. Kodali, S. R. Kadiri, and P. Alku, “Wav2vec-based detection and severity level classification of dysarthria from speech,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, Jun. 2023

2023
[13]

Y. Rong, M. Fynn, and S. Nordholm,A Pre-Screening Technique for Coronary Artery Disease with Multi-Channel Phonocardiography and Electrocardiography. Taylor & Francis, 2023, ch. 9

2023
[14]

Leatham,Auscultation of the Heart and Phonocardiography, 2nd ed

A. Leatham,Auscultation of the Heart and Phonocardiography, 2nd ed. Edinburgh: Churchill Livingstone, 1975

1975
[15]

An Open Access Database for the Evaluation of Heart Sound Algorithms,

C. Liu, D. Springer, Q. Li, B. Moody, R. A. Juan, F. J. Chorro, F. Castells, J. M. Roig, I. Silva, A. E. Johnson, Z. Syed, S. E. Schmidt, C. D. Papadaniil, L. Hadjileontiadis, H. Naseri, A. Moukadem, A. Dieterlen, C. Brandt, H. Tang, M. Samieinasab, M. R. Samieinasab, R. Samen...

2016
[16]

Acoustic features for the identification of coronary artery disease,

S. E. Schmidt, C. Holst-Hansen, J. Hansen, E. Toft, and J. J. Struijk, “Acoustic features for the identification of coronary artery disease,”IEEE Transactions on Biomedical Engineering, vol. 62, no. 11, pp. 2611–2619, Nov. 2015

2015
[17]

Logistic regression-hsmm-based heart sound segmentation,

D. B. Springer, L. Tarassenko, and G. D. Clifford, “Logistic regression-hsmm-based heart sound segmentation,”IEEE transactions on biomedical engineering, vol. 63, no. 4, pp. 822–832, 2016

2016
[18]

Electrocardiogram signal analysis-an overview,

R. Rajni and I. Kaur, “Electrocardiogram signal analysis-an overview,”International Journal of Computer Applications, vol. 84, no. 7, pp. 22–25, 2013

2013
[19]

G. D. Clifford, F. Azuaje, and P. McSharry,Advanced Methods and Tools for ECG Data Analysis. Artech House, 2006

2006
[20]

Xie,Biomedical Signal Processing: An ECG Application

C. Xie,Biomedical Signal Processing: An ECG Application. Cham: Springer International Publishing, 2020, pp. 285–303

2020
[21]

Prognostic value of ischemic electrocardiographic findings for cardiovascular mortality in men and women,

D. De Bacquer, G. De Backer, M. Kornitzer, K. Myny, Z. Doyen, and H. Blackburn, “Prognostic value of ischemic electrocardiographic findings for cardiovascular mortality in men and women,” Journal of the American College of Cardiology, vol. 32, no. 3, pp. 680–685, 1998

1998
[22]

Practicality meets precision: Wearable vest with integrated multi-channel pcg sensors for effective coronary artery disease pre-screening,

M. Fynn, K. Mandana, J. Rashid, S. Nordholm, Y. Rong, and G. Saha, “Practicality meets precision: Wearable vest with integrated multi-channel pcg sensors for effective coronary artery disease pre-screening,” p. 109904, 2025. [Online]. Available: https: //www.sciencedirect.com/...

2025
[23]

Librispeech: An ASR corpus based on public domain audio books,

V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public domain audio books,” inProc. of ICASSP. IEEE, 2015, pp. 5206–5210

2015
[24]

Deep unsupervised learning using nonequilibrium thermodynamics,

J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” 2015

2015
[25]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” 2020

2020
[26]

Auto-encoding variational Bayes,

D. P. Kingma and M. Welling, “Auto-encoding variational Bayes,” 2022

2022
[27]

Generative adversarial networks,

I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,” 2014

2014
[28]

High fidelity speech synthesis with adversarial networks,

M. Bińkowski, J. Donahue, S. Dieleman, A. Clark, E. Elsen, N. Casagrande, L. C. Cobo, and K. Simonyan, “High fidelity speech synthesis with adversarial networks,” 2019

2019
[29]

Icentia11k: An unsupervised representation learning dataset for arrhythmia subtype discovery,

S. Tan, G. Androz, A. Chamseddine, P. Fecteau, A. Courville, Y. Bengio, and J. P. Cohen, “Icentia11k: An unsupervised representation learning dataset for arrhythmia subtype discovery,” 2019

2019
[30]

Ephnogram: A simultaneous electrocardiogram and phonocardiogram database,

A. Kazemnejad, P. Gordany, and R. Sameni, “Ephnogram: A simultaneous electrocardiogram and phonocardiogram database,” 2021

2021
[31]

The impact of the mit-bih arrhythmia database,

G. Moody and R. Mark, “The impact of the mit-bih arrhythmia database,”IEEE Engineering in Medicine and Biology Magazine, vol. 20, no. 3, pp. 45–50, 2001

2001
[33]

The advantages of the matthews correlation coefficient (MCC) over f1 score and accuracy in binary classification evaluation,

D. Chicco and G. Jurman, “The advantages of the matthews correlation coefficient (MCC) over f1 score and accuracy in binary classification evaluation,”BMC Genomics, vol. 21, no. 6, 2020

2020
[34]

Optuna: A next-generation hyper- parameter optimization framework,

T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyper- parameter optimization framework,” inProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2019

2019
[35]

LoRA: Low-rank adaptation of large language models,

E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” 2021. [Online]. Available: https://arxiv.org/abs/2106.09685

2021 arXiv
[36]

Understanding how dimension reduction tools work: An empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization,

Y. Wang, H. Huang, C. Rudin, and Y. Shaposhnik, “Understanding how dimension reduction tools work: An empirical approach to deciphering t-sne, umap, trimap, and pacmap for data visualization,” 2021. [Online]. Available: https://arxiv.org/abs/2012.04456

2021 arXiv
[37]

Audio for audio is better? an investigation on transfer learning models for heart sound classification,

T. Koike, K. Qian, Q. Kong, M. D. Plumbley, B. W. Schuller, and Y. Yamamoto, “Audio for audio is better? an investigation on transfer learning models for heart sound classification,” in 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Soc...

2020
[38]

Time–frequency-domain deep learning framework for the automated detection of heart valve disorders using pcg signals,

J. Karhade, S. Dash, S. K. Ghosh, D. K. Dash, and R. K. Tripathy, “Time–frequency-domain deep learning framework for the automated detection of heart valve disorders using pcg signals,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–11, 2022

2022
[39]

Heart sound classification considering additive noise and convolutional distortion,

F. B. Azam, M. I. Ansari, I. Mclane, and T. Hasan, “Heart sound classification considering additive noise and convolutional distortion,” 2021. [Online]. Available: https://arxiv.org/abs/2106.01865

2021 arXiv
[40]

Enhancing cross-domain robustness in phonocardiogram signal classification using domain-invariant preprocessing and transfer learning,

A. Maity and G. Saha, “Enhancing cross-domain robustness in phonocardiogram signal classification using domain-invariant preprocessing and transfer learning,”Computer Methods and Programs in Biomedicine, vol. 257, p. 108462, 2024. [Online]. Available: https://www.sciencedirect...

2024
[41]

Prediction of cardiovascular diseases by integrating multi-modal features with machine learning methods,

P. Li, Y. Hu, and Z.-P. Liu, “Prediction of cardiovascular diseases by integrating multi-modal features with machine learning methods,”Biomedical Signal Processing and Control, vol. 66, p. 102474, 2021

2021
[42]

Cardiovascular disease detection based on deep learning and multi-modal data fusion,

J. Zhu, H. Liu, X. Liu, C. Chen, and M. Shu, “Cardiovascular disease detection based on deep learning and multi-modal data fusion,”Biomedical Signal Processing and Control, vol. 99, p. 106882, 2025. 35

2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [32]

Generative deep learning and signal processing for data augmentation of cardiac auscultation signals: Improving model robustness using synthetic audio,

L. Abbott, M. Marocchi, M. Fynn, Y. Rong, and S. Nordholm, “Generative deep learning and signal processing for data augmentation of cardiac auscultation signals: Improving model robustness using synthetic audio,” 2024. [Online]. Available: https://arxiv.org/abs/2410.10125 34

arXiv 2024

[2] [1]

Cardiovascular Diseases (CVDs)

WHO, "Cardiovascular Diseases (CVDs)".Geneva, Switzerland: WHO, 2021

2021

[3] [2]

Cardiac Auscultation: Rediscovering the Lost Art,

M. A. Chizner, “Cardiac Auscultation: Rediscovering the Lost Art,”Current Problems in Cardiology, vol. 33, no. 7, pp. 326–408, Jul. 2008

2008

[4] [3]

The Lost Art of Clinical Skills,

C. A. Feddock, “The Lost Art of Clinical Skills,”The American Journal of Medicine, vol. 120, no. 4, pp. 374–378, Apr. 2007

2007

[5] [4]

Accuracy of cardiac auscultation in detection of neonatal congenital heart disease by general paediatricians,

Q.-M. Zhao, C. Niu, F. Liu, L. Wu, X.-J. Ma, and G.-Y. Huang, “Accuracy of cardiac auscultation in detection of neonatal congenital heart disease by general paediatricians,”Cardiology in the Young, vol. 29, no. 5, pp. 679–683, May 2019

2019

[6] [5]

Abnormal heart sound classification and model interpretability: A transfer learning approach with deep learning,

M. Marocchi, L. Abbott, Y. Rong, S. Nordholm, and G. Dwivedi, “Abnormal heart sound classification and model interpretability: A transfer learning approach with deep learning,”Journal of Vascular Diseases, vol. 2, no. 4, pp. 438–459, 2023. [Online]. Available: https://www.mdpi.com/2813-2475/2/4/34

2023

[7] [6]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” 2017

2017

[8] [7]

Diffwave: A versatile diffusion model for audio synthesis,

Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro, “Diffwave: A versatile diffusion model for audio synthesis,” 2021

2021