REVIEW 5 major objections 6 minor 11 references
Large Language Models and Non-Negative Matrix Factorization for Bioacoustic Signal Decomposition
T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read A hybrid of unsupervised non-negative matrix factorization and a pretrained large language model can separate overlapping heart and lung sounds and map each component to plausible clinical conditions without labels.
desk verdict Workshop-level demonstration of an already-published LLM-NMF pipeline, with the central separation claim resting on a single unvalidated figure. 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 load-bearing object is the standard NMF decomposition $\mathbf{V} \approx \mathbf{W}\mathbf{H}$ of a non-negative time-frequency matrix, solved by minimizing generalized Kullback-Leibler divergence with multiplicative updates. The rows of $\mathbf{H}$ are time-resolved source estimates, the columns of $\mathbf{W}$ encode each source's contribution, and the interpretation stage converts each row $\mathbf{h}_k$ into a feature vector $\mathbf{f}_k$, formats it as a prompt through a deterministic function $\phi$, and lets an LLM $\mathcal{L}$ map the prompt to a diagnostic label. The NMF step does the actual separation; the LLM step supplies the semantic mapping from waveform shape to medical terms.
What would settle it
One decisive test would record the heart and lung sources separately on the manikin, mix them with known relative gains, run the full pipeline, and compare each recovered component with the true source waveforms; the central claim fails if the component labeled 'heart' correlates more strongly with the recorded lung sound than with the recorded heart sound, or if the LLM's label disagrees with the source identity.
Extended reading notes
Core claim
The paper's central claim is that combining blind source separation with an LLM interpretation layer yields clinically meaningful labels for overlapping heart and lung sounds without ground-truth annotations or source priors. After computing a short-time Fourier transform of the mixed signal, NMF factorizes the magnitude spectrogram into a mixing matrix and a source matrix; each separated source row is converted into a feature vector and then into a text prompt, and the LLM maps the prompt to diagnostic terms. In the reported example, the separation stage isolates a lung component with low-frequency bursts and a heart component with irregular rhythm, and the LLM associates these with possible wheezing or airway obstruction and possible atrial fibrillation or other rhythm disorders. The demonstration is qualitative, relying on visual inspection of the waveforms and spectrograms.
Load-bearing premise
The load-bearing premise is that the NMF component visually classified as 'heart' or 'lung' truly corresponds to that physiological source; because that source identity is assigned by inspection before the LLM interprets it, a labeling error would invalidate the clinical readings, and manikin acoustics are only a proxy for real patient recordings.
Editorial extensions
If this is right
- Because NMF is unsupervised and the LLM is pretrained, the pipeline can be applied to new mixed recordings without labeled data or source-specific priors.
- The method converts separated waveforms into textual clinical suggestions, giving a human-readable link between signal shape and possible pathology.
- Future work identified in the paper, such as domain-specific LLM training and richer input features, would directly improve the accuracy of the interpretation layer.
- The reported results show the approach is at least feasible on controlled manikin recordings, the intended first step before clinical validation.
Reading between the lines
- The decisive test the paper leaves undone is to record each source separately on the manikin, mix them artificially, and check that each recovered component corresponds to the right source; until that is done, the 'heart' and 'lung' labels rest on visual similarity rather than ground truth.
- A natural quantitative extension would measure separation quality, for example correlation between recovered and true source signals, and label accuracy against clinician annotations, turning the qualitative example into a measurable claim.
- Because the LLM is a pretrained text model, it can produce plausible-sounding but unverified diagnoses; a production version would need a confidence score or a restricted label set to avoid over-claiming.
- If the separation stage generalizes, the approach could be applied to continuous home monitoring with consumer stethoscopes, where expert labels are scarce and overlapping cardiopulmonary sounds are common.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a bioacoustic source-separation framework that combines non-negative matrix factorization (NMF) with a pretrained large language model (LLM). Audio recordings of cardiopulmonary sounds are collected from a clinical manikin with a digital stethoscope, transformed via STFT, decomposed with NMF, and the separated components are converted into feature vectors that are then fed to an LLM to produce clinical interpretations such as 'atrial fibrillation' or 'wheezing'. The authors report qualitative results on a single figure and claim the method is unsupervised and does not require labeled data or prior source knowledge.
Significance. If the central claims were fully validated, the approach could offer an interpretable, label-free pipeline for bioacoustic monitoring in which LLM post-processing translates separated components into clinically meaningful text. However, the manuscript does not currently provide quantitative evidence that the NMF components correspond to true heart and lung sources, nor any evaluation of the LLM's diagnostic suggestions, so the practical significance is not established. The controlled manikin recording setup is a useful element, but it does not substitute for ground-truth validation.
major comments (5)
- [Results, Figure 1] The central claim that the method 'effectively separates overlapping bioacoustic sources' is supported only by visual inspection of a single figure. There are no separation metrics, no comparison to ground-truth source signals, no signal-to-distortion ratios, and no baseline NMF or other BSS method. The absence of quantitative validation makes the central claim unverifiable as presented.
- [Results and Methods] The assignment of NMF rows to 'heart' and 'lung' components is made by the authors' visual pattern inspection before the LLM interpretation; this labeling is circular because the subsequent LLM 'recognition' of atrial fibrillation or wheezing merely confirms the prior visual identification. Without independent ground truth (e.g., simultaneously recorded reference heart/lung channels or known synthetic mixtures), the clinical interpretations are not tested.
- [Methods, Eq. (1)] The KL divergence formula is garbled: the displayed expression is not a valid divergence (subscripts and brackets are malformed), and the text does not specify the summation indices or the multiplicative update rules. Since the factorization objective is central to the algorithm, this needs correction and precise definition.
- [Introduction and reference [1]] The manuscript claims novelty for combining NMF and LLMs, but reference [1] by the same authors already introduces 'large language model-based nonnegative matrix factorization for cardiorespiratory sound separation.' The incremental contribution of this manuscript over [1] is not articulated; the new elements (e.g., the feature extraction and prompt template) are described only at a high level.
- [Discussion] The paper's own limitations statement concedes that the LLM 'operates purely on pattern-to-text mappings, which may limit its accuracy' and that interpretability is constrained by feature quality. However, no evidence is given about LLM accuracy, feature quality, or how the mapping is evaluated; this concession underscores that the reported clinical associations are unvalidated.
minor comments (6)
- [Methods] The feature vector f_k and prompt template phi are never specified; the reader cannot reproduce the LLM step.
- [Methods] The NMF rank r is not described; no rank-selection criterion or sensitivity analysis is provided.
- [Methods] STFT parameters (window, hop, FFT size) are not given.
- [Abstract and Results] The abstract says 'no labeled data or prior knowledge of source types,' but the visual labeling of components in Results implies prior knowledge of expected heart/lung morphology.
- [References] Reference [2] is an earlier periodicity-based NMF approach; the relation of the present work to that method is not discussed.
- [General] The manuscript contains formatting issues in the equations and notation, and the YouTube link under 'Audiovisual Material' is not cited in the text.
Circularity Check
The only validation of source separation is the authors' visual labeling of components by the same patterns the LLM is then asked to interpret, so the clinical findings restate the labeling criterion rather than test the method.
-
self definitional
[Results (Figure 1)]
"Following decomposition, the lung component was identified by its low-frequency bursts, which are consistent with known wheezing patterns. This pattern was further interpreted by the large language model (LLM), which suggested the possibility of a respiratory abnormality such as wheezing or airway obstruction. The heart component, in contrast, displayed regular rhythmic peaks disrupted by irregular intervals. The LLM recognized these anomalies as potential signs of atrial fibrillation or other rhythm disorders."
The only evidence that NMF separated true heart/lung sources is the authors' visual identification of components by their patterns. Those same patterns are then presented as the LLM's clinical interpretation: the component is called 'lung' because of low-frequency bursts, and that burst pattern justifies the wheezing suggestion; the component is called 'heart' because of rhythmic peaks with irregular intervals, and that irregularity justifies the atrial-fibrillation suggestion. The LLM thus restates the labeling criterion in medical vocabulary rather than validating it.
full rationale
The algorithmic chain (STFT, standard NMF with KL divergence, feature extraction, pretrained LLM) is not circular: the NMF update rules are generic and the LLM is not trained or fine-tuned on the manikin data. The circularity enters at the evaluation step. In Results, the 'lung component' is selected because it shows low-frequency bursts 'consistent with known wheezing patterns,' and the 'heart component' because it shows rhythmic peaks with irregular intervals; these same patterns are then cited as the LLM's suggestions of wheezing and atrial fibrillation. This makes the clinical interpretation a restatement of the component-selection criterion, not an independent confirmation that NMF recovered the true physiological sources. The paper provides no ground-truth source signals, no rank-selection procedure, no quantitative separation metric, and no specification of the prompt template, so the central claim of 'effective separation' is supported only by this circular visual labeling. The self-citations ([1], [2], [11]) do not in themselves make the derivation circular: [1] describes the same LLM+NMF combination and [2] a periodicity-based NMF, but this paper uses standard NMF and a pretrained LLM, so the machinery is independent. The main deficiency is the circular validation of the source identities, not a load-bearing citation chain. This warrants a moderate score of 4: partial circularity in the evidence for the central claim, with independent algorithmic components.
Assumptions & free parameters
free parameters (2)
- NMF rank r
- Feature vector dimension d
assumptions (3)
- domain assumption NMF with KL divergence can separate overlapping heart and lung sounds into additive non-negative spectrogram components.
- domain assumption The clinical manikin recordings are representative of real cardiorespiratory acoustics and the separated components correspond to genuine heart and lung sources.
- domain assumption A pretrained LLM can map numerical acoustic features to clinically valid labels through a text prompt.
Cite this review
Pith. "Pith review of Large Language Models and Non-Negative Matrix Factorization for Bioacoustic Signal Decomposition." pith.science (2026). https://pith.science/paper/MR2D75K2
@misc{pith2026250709161,
author = {Pith},
title = {Pith review of: Large Language Models and Non-Negative Matrix Factorization for Bioacoustic Signal Decomposition},
year = {2026},
howpublished = {\url{https://pith.science/paper/MR2D75K2}},
note = {Machine review of arXiv:2507.09161}
}
read the original abstract
Large language models have shown a remarkable ability to extract meaning from unstructured data, offering new ways to interpret biomedical signals beyond traditional numerical methods. In this study, we present a matrix factorization framework for bioacoustic signal analysis which is enhanced by large language models. The focus is on separating bioacoustic signals that commonly overlap in clinical recordings, using matrix factorization to decompose the mixture into interpretable components. A large language model is then applied to the separated signals to associate distinct acoustic patterns with potential medical conditions such as cardiac rhythm disturbances or respiratory abnormalities. Recordings were obtained from a digital stethoscope applied to a clinical manikin to ensure a controlled and high-fidelity acquisition environment. This hybrid approach does not require labeled data or prior knowledge of source types, and it provides a more interpretable and accessible framework for clinical decision support. The method demonstrates promise for integration into future intelligent diagnostic tools.
Figures
Reference graph
Works this paper leans on
-
[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
-
[11]
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
-
[2]
Torabi, Y., Shirani, S., & Reilly, J. P. (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
-
[3]
Torabi, Y., Shirani, S., & Reilly, J. P. (2024). Exploring Sensing Devices for Heart and Lung Sound Monitoring, arXiv preprint, doi: 2406.12432,
arXiv 2024
-
[4]
Baraeinejad, B., et al. (2024). Design and implementation of an IoT-based respiratory motion sensor. arXiv preprint arXiv:2412.05405
work page Pith review arXiv 2024
-
[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
work page 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]
Vazifeh, A. R., & Fleischer, J. W. (2024). Manifold learning for personalized and label-free detection of cardiac arrhythmias. arXiv preprint arXiv:2506.16494. https://doi.org/10.48550/arXiv.2506.16494
Show all 11 references
-
[8]
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
-
[9]
Analysis of EEG based functional connectivity during mental arithmetic task,
H. Nejat and S. H. Sardouie, "Analysis of EEG based functional connectivity during mental arithmetic task," 2021 28th National and 6th International Iranian Conference on Biomedical Engineering (ICBME), Tehran, Iran, Islamic Republic of, 2021, pp. 118-124, doi: 10.1109/ICBME54...
2021
-
[10]
& Valero-Cuevas, F
Urbina-Meléndez, D., Azadjou, H. & Valero-Cuevas, F. J. Brainbody-task co-adaptation can improve autonomous learning and speed of bipedal walking. arXiv preprint arXiv:2402.02387 (2024)
2024 arXiv
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.