REVIEW 5 major objections 5 minor 17 references
Blind Source Separation in Biomedical Signals Using Variational Methods
T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read An unsupervised VAE can separate overlapping heart and lung sounds from stethoscope mixtures.
desk verdict A plausible but under-evidenced application of a standard VAE to heart/lung sound separation; the separation claim is not entailed by the training objective as described. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The mechanism is the variational autoencoder itself: an encoder $q_\phi(z|x)$ compresses mixed signals into a low-dimensional latent code $z$, and a probabilistic decoder $p_\theta(x|z)$ reconstructs the input, trained by maximizing the evidence lower bound $\mathcal{L}(\theta,\phi;x)=\mathbb{E}_{q_\phi(z|x)}[\log p_\theta(x|z)]-D_{\mathrm{KL}}(q_\phi(z|x)\|p(z))$. Separation is performed by decoding from distinct regions of the latent space, namely the clusters that emerge for heart versus lung sources. The t-SNE (a dimensionality-reduction visualization) projections of the latent space are the evidence offered that such distinct regions exist.
What would settle it
A concrete check: use recordings with separately known heart and lung sounds, train the same model, and measure how closely the separated outputs match the true sources (for example, signal-to-distortion ratio). If the decoded clusters do not match the isolated sources better than the raw mixture does, the claimed separation is not happening.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is that the structure of a VAE latent space, learned only from mixed signals, encodes the acoustic sources separately. The encoder maps a mixed signal $x$ to a variational posterior $q_\phi(z|x)$; the decoder learns $p_\theta(x|z)$; training maximizes the evidence lower bound, balancing reconstruction quality with KL regularization toward a standard normal prior. After training, t-SNE projections show distinct latent clusters for heart and lung sounds, and decoding from those clusters gives reconstructions whose spectrograms match the true source components. The authors take this as evidence that unsupervised blind source separation is achievable without labels and without prior knowledge of the sources.
Load-bearing premise
The whole method rests on the VAE's internal representation splitting cleanly by which sound produced it, with a known mapping from each cluster to heart or lung; the paper does not prove that split or explain how the mapping is assigned.
Editorial extensions
If this is right
- Unsupervised separation of heart and lung sounds becomes feasible without labeled training pairs, lowering a barrier to clinical deployment.
- A VAE trained this way could be embedded in a digital stethoscope or portable monitor, returning separated heart and lung audio during auscultation.
- The emergence of interpretable latent clusters gives a visual check that the model has separated sources rather than merely memorizing mixtures.
- The method offers a generative alternative to periodicity-based non-negative matrix factorization, with potential to handle overlapping, nonstationary acoustics.
Reading between the lines
- A practical extension is to label the latent clusters automatically by matching each cluster's spectral content to known heart and lung frequency ranges, since the paper leaves that assignment implicit.
- Because the data come from a clinical manikin, the strongest next test is real human recordings with body sounds, motion artifacts, and variable stethoscope placement.
- If the latent structure is source-specific rather than clip-specific, the same training scheme could transfer to other overlapping biosignals, such as bowel and respiratory sounds or speech embedded in noise.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript proposes an unsupervised method for separating heart and lung sounds from mixed recordings using a variational autoencoder (VAE). The VAE is trained to reconstruct mixtures, and separation is attempted by decoding 'distinct regions' of the latent space after t-SNE visualization suggests clusters. The paper reports distinct latent clusters and spectrograms of predicted components that visually match ground truth sources.
Significance. If substantiated, the approach would offer a practical unsupervised tool for cardiorespiratory sound separation, with application to digital stethoscopes and portable diagnostics. The paper's strengths include use of real clinical-manikin recordings, a publicly described dataset (HLS-CMDS, Ref [17]), and a video presentation. However, the current evidence is qualitative and the separation procedure is not specified, so the central claim is not yet established.
major comments (5)
- [Methods] The sentence "The separated signals are obtained by decoding distinct regions in the latent space associated with different acoustic sources" does not specify how the regions are determined, how clusters are assigned to heart versus lung, or how the decoder produces a single source when it was trained to map z to the full mixture. Provide an algorithmic description of the separation step and explain why a decoder trained on the ELBO for mixtures should yield source-specific outputs when restricted to a latent region.
- [Results / Figures 1-2] Provide quantitative evaluation metrics (e.g., SDR, SIR, correlation coefficient, spectral error) with error bars over repeated runs, and compare against the NMF baselines cited in Refs [1]-[3]. Visual t-SNE and spectrogram comparisons are not sufficient to support the claims of "high fidelity" and "strong alignment."
- [Results / Methods] The manuscript does not state whether ground-truth labels or source signals were used to identify which latent clusters correspond to heart and lung. If any ground-truth information was used for cluster labeling or region selection, the method is not fully blind; disclose this and argue why the procedure remains unsupervised.
- [Methods] Missing implementation details: latent dimension k, encoder/decoder architectures, input representation (raw waveform vs spectrogram), training hyperparameters, dataset size, and train/test split. Without these, the experiments cannot be reproduced or evaluated.
- [Methods / Results] The t-SNE visualizations show clusters, but t-SNE is a nonlinear dimensionality-reduction technique that can introduce spurious clusters. Since the VAE prior is a standard normal and no identifiability conditions are imposed, the latent space is not guaranteed to be interpretable; relate the clusters to a quantitative measure of separation in the original latent space (e.g., clustering metrics on z, or decoding accuracy per cluster).
minor comments (5)
- [Methods] The ELBO formula is garbled in the text; it should be typeset with proper notation such as L(θ,φ;x) = E_{q_φ(z|x)}[log p_θ(x|z)] − KL(q_φ(z|x) || p(z)).
- [Figure 1] The caption mentions "successive training epochs" but the figure has no epoch numbers or legend identifying which cluster is heart and which is lung.
- [Introduction] The text states "operates without prior knowledge about the nature or number of sources," but the method appears to assume two sources (heart and lung) for the clustering assignment; clarify this assumption.
- [References] Several references (e.g., [7]-[16]) are only tangentially related to the method and are not discussed in the text; consider trimming to keep the paper focused.
- [Discussion] The Discussion claims "improved interpretability, flexibility, and generalization" relative to other methods without supporting experiments; temper these claims or add comparisons.
Circularity Check
No significant circularity: the paper is an empirical VAE demonstration whose central claim does not reduce by construction to its inputs, and the self-citations are peripheral.
full rationale
The paper contains no derivation chain that reduces to its own inputs. The only equation given is the standard VAE ELBO in Methods, and the reconstruction term is applied to the mixed signal, not to the separated outputs; no fitted parameter is renamed as a prediction. The separation step, 'The separated signals are obtained by decoding distinct regions in the latent space associated with different acoustic sources,' is underspecified and not entailed by the ELBO, but that is an empirical validation gap rather than circularity: the claim is asserted from t-SNE and spectrogram inspection, not derived from a definition that already includes the conclusion. Self-citations [1,2,3,17] provide prior NMF work and the dataset description, but they are not load-bearing for the central claim, are not invoked as a uniqueness theorem, and do not forbid alternative methods. The possible use of ground truth to label clusters would weaken the unsupervised claim, yet the paper does not exhibit the specific equation-level reduction or fitted-parameter renaming required to classify that as circularity. Therefore no circular step is established, and the score is 0.
Assumptions & free parameters
free parameters (3)
- Latent dimension k
- VAE architecture and training hyperparameters
- Latent cluster-to-source assignment
assumptions (4)
- standard math Standard VAE ELBO and reparameterization: q_phi(z|x), p_theta(x|z), prior p(z)=N(0,I).
- domain assumption Heart and lung sounds are generated by independent latent factors that remain factorized in the VAE latent space.
- domain assumption Recordings from one clinical manikin are representative of clinical heart and lung sound mixtures.
- ad hoc to paper Unlabeled t-SNE clusters correspond one-to-one to the heart and lung sources and the assignment is known.
Cite this review
Pith. "Pith review of Blind Source Separation in Biomedical Signals Using Variational Methods." pith.science (2026). https://pith.science/paper/GY6IOZAN
@misc{pith2026250618281,
author = {Pith},
title = {Pith review of: Blind Source Separation in Biomedical Signals Using Variational Methods},
year = {2026},
howpublished = {\url{https://pith.science/paper/GY6IOZAN}},
note = {Machine review of arXiv:2506.18281}
}
read the original abstract
This study introduces a novel unsupervised approach for separating overlapping heart and lung sounds using variational autoencoders (VAEs). In clinical settings, these sounds often interfere with each other, making manual separation difficult and error-prone. The proposed model learns to encode mixed signals into a structured latent space and reconstructs the individual components using a probabilistic decoder, all without requiring labeled data or prior knowledge of source characteristics. We apply this method to real recordings obtained from a clinical manikin using a digital stethoscope. Results demonstrate distinct latent clusters corresponding to heart and lung sources, as well as accurate reconstructions that preserve key spectral features of the original signals. The approach offers a robust and interpretable solution for blind source separation and has potential applications in portable diagnostic tools and intelligent stethoscope systems.
Figures
Reference graph
Works this paper leans on
-
[17]
Y. Torabi, S. Shirani and J. P. Reilly, "Descriptor: Heart and Lung Sounds Dataset Recorded From a Clinical Manikin Using Digital Stethoscope (HLS-CMDS)," in IEEE Data Descriptions, vol. 2, pp. 133-140, 2025, doi: 10.1109/IEEEDATA.2025.3566012
arXiv 2025
-
[1]
Torabi, Y., Shirani, S., & Reilly, J. P. (2025). Large language model-based nonnegative matrix factorization for cardiorespiratory sound separation. arXiv preprint arXiv:2502.05757. https://doi.org/10.48550/arXiv.2502.05757
-
[3]
Torabi, Y., Shirani, S., Reilly, J. P., & Gauvreau, G. M. (2023). A new non-negative matrix factorization approach for blind source separation of cardiovascular and respiratory sound based on the periodicity of heart and lung function. arXiv preprint arXiv:2305.01889
arXiv 2023
-
[2]
Torabi, Y., Shirani, S., Reilly, J. P., & Gauvreau, G. M. (2024). MEMS and ECM sensor technologies for cardiorespiratory sound monitoring—A comprehensive review. Sensors (Basel, Switzerland), 24(21),
work page 2024
-
[4]
Baraeinejad, B., Fallah Shayan, et al. (2022). Design and implementation of an ultralow-power ECG patch and smart cloud-based platform. IEEE Transactions on Instrumentation and Measurement, 71, 1–11
work page 2022
-
[5]
Bhatnagar, P., Zaferani, et al. (2023). Advancing personalized healthcare and entertainment: Progress in energy harvesting materials and techniques of self-powered wearable devices. Progress in Materials Science, 139, 101184
2023
-
[6]
Baraeinejad, B., et al. (2023). "Clinical IoT in Practice: A Novel Design and Implementation of a Multi-functional Digital Stethoscope for Remote Health Monitoring," TechRxiv Preprints, Nov. 7, 2023, doi: 10.36227/techrxiv.24459988
-
[7]
Ahmadi Ranjbar, M., Ghaleh, et al. (2025). Beyond subjective measures: Systematic review of deep learning in chronic pain—Modalities, methods, and applications
work page 2025
Show all 17 references
-
[8]
Razmara, P., Khezresmaeilzadeh, T., & Jenkins, B. K. (2024). Fever detection with infrared thermography: Enhancing accuracy through machine learning techniques. 2024 IEEE EMBS International Conference on Biomedical and Health Informatics (BHI)
2024
-
[9]
Ashrafi, N., Liu, et al. (2024). Deep learning model utilization for mortality prediction in mechanically ventilated ICU patients. Informatics in Medicine Unlocked, 49, 101562
2024
-
[10]
Ashrafi, N., Abdollahi, et al. (2025). Enhanced prediction of ventilator-associated pneumonia in patients with traumatic brain injury using advanced machine learning techniques. Scientific Reports, 15(1), 11363
2025
-
[11]
Nejat, H., Sherfey, J., & Bastos, A. M. (2025). Predictive routing emerges from self-supervised stochastic neural plasticity. bioRxiv, 2024.12.31.630823
2025
-
[12]
Nejat, H., Sherfey, J., & Bastos, A. M. (2024). Replicating spectro-temporal dynamics in neurobiologically realistic neural networks via a self-supervised approach. Cognitive Computational Neuroscience Conference Abstract
2024
-
[13]
Azizi, S., Nazemi, et al. (2024). Automated optimization of deep neural networks: Dynamic bit-width and layer-width selection via cluster-based Parzen estimation. 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE), 1–6
2024
-
[14]
Movahhedrad, M., Chen, Z., & Kuo, C.-C. J. (2024). A green learning approach to efficient image demosaicking. 2024 IEEE International Conference on Big Data (BigData), 1067–1074
2024
-
[15]
Baraeinejad, B., et al. (2024). Design and implementation of an IoT-based respiratory motion sensor. arXiv preprint arXiv:2412.05405
2024 arXiv
-
[16]
H., Dolamic, L., & Frossard, P
Sadrizadeh, S., Dabiri Aghdam, A. H., Dolamic, L., & Frossard, P. (2023). Targeted adversarial attacks against neural machine translation. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1–5
2023
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.