Pith. sign in

REVIEW 4 major objections 5 minor 4 cited by

Large Language Model-based Nonnegative Matrix Factorization For Cardiorespiratory Sound Separation

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read The paper presents LingoNMF, the first NMF method augmented by LLM feedback, and reports better heart and lung separation than standard NMF, alpha-NMF, and PL-NMF on two sound datasets.

desk verdict A new LLM-NMF combination that ships code and data, but the core penalty mechanism is not actually optimized and the evaluation is tuned on test data. read the letter →

arxiv 2502.05757 v1 pith:44G7EA47 submitted 2025-02-09 cs.SD eess.ASeess.SP

classification cs.SDeess.ASeess.SP
keywords largelanguagemodelsnonnegativematrixfactorizationblindsourceseparationheartsoundlungfundamentalfrequencypenaltyalpha-divergencedigitalstethoscope
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

This paper introduces LingoNMF, described as the first integration of a large language model with nonnegative matrix factorization (NMF) for separating heart and lung sounds. The central claim is that a LLaMA 2 model in a feedback loop improves separation by dynamically updating a fundamental-frequency penalty in the NMF cost function. The evidence comes from two datasets: 100 synthesized mixtures of real heart and lung sounds, and 210 clinical-manikin recordings of normal and abnormal sounds. A sympathetic reading of the paper takes the claimed separation gains as the contribution, with the LLM acting both as a tuner of the penalty term and as a provider of diagnostic text about the separated sounds.

What carries the argument

The central object is the augmented cost function $D' = D(\mathbf{Y}\|\mathbf{A}\mathbf{X}) + \lambda_f \|\mathbf{f}_{\mathrm{est}} - \mathbf{f}\|^2$, where $\mathbf{f}_{\mathrm{est}}$ is the vector of fundamental frequencies estimated as the power-spectral-density argmax for each current separated estimate. The stated mechanism is the feedback loop: the LLM receives spectro-temporal features of the current estimates, adjusts $\mathbf{f}$ toward values consistent with normal or abnormal cardiorespiratory sounds, and the factorization is re-run. This sits on top of PL-NMF, a multilayer, periodicity-aware version of $\alpha$-NMF that already uses autocorrelation to separate heart and lung components by their different period ranges.

What would settle it

Compute the gradient of the penalty term $\lambda_f \|\mathbf{f}_{\mathrm{est}}(\mathbf{A},\mathbf{X}) - \mathbf{f}\|^2$ with respect to $\mathbf{A}$ and $\mathbf{X}$ and check whether the unmodified multiplicative updates (3)-(4) form a descent step for it; if the penalty can increase, LingoNMF is not minimizing the stated objective. A paired experiment replacing the LLM's frequency choices with a grid search or random schedule on the same datasets would then settle whether the LLM feedback, rather than the penalty itself, produces the reported SDR gains.

Watch

Extended reading notes

Core claim

The authors propose that a large language model can participate in NMF iteration as a feedback controller rather than only as a post-processing annotator. In LingoNMF, after each factorization update, features of the current separated estimates are sent to LLaMA 2, which returns updated values for the vector of source fundamental frequencies; these values enter the cost function as a penalty on the deviation between estimated and target fundamental frequencies. The paper reports that this loop raises source-to-distortion ratio on both datasets for lung sounds and, on the 100-mixture dataset, for heart sounds, with the largest gains appearing in the LingoNMF columns of its comparison table.

Load-bearing premise

The paper assumes that the standard alpha-NMF multiplicative update rules minimize the modified cost with the added fundamental-frequency penalty, even though no update rule or gradient for that penalty is derived.

Editorial extensions

If this is right

  • If the reported gains hold, LLM feedback can replace manual or grid-search tuning of NMF penalty parameters during the factorization itself.
  • The same loop could be applied to other periodic biomedical recordings, such as bowel or fetal sounds, wherever a fundamental-frequency prior is useful.
  • The reported optimal range of five to seven input mixtures would guide the design of stethoscope arrays for practical monitoring.
  • The algorithm separates sounds using only the mixture, so it can be deployed where pure reference sounds are unavailable.
  • The LLM's text output also yields a human-readable abnormality analysis, coupling separation with diagnostic suggestions in one pipeline.

Reading between the lines

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

  • Not stated by the paper: on the manikin dataset the heart-sound SDR for LingoNMF is lower than for $\alpha$-NMF, so the consistency claim is stronger for lung sounds and for the synthesized dataset; a per-condition comparison with error bars would clarify where the method actually helps.
  • An ablation that replaces the LLM with a simple numerical rule or a grid search for the fundamental-frequency vector would reveal whether the language model itself, rather than the penalty mechanism, drives the reported improvement.
  • Because $f_0$ is estimated as the PSD argmax, abnormalities that alter harmonic structure can shift the estimate; the penalty update may need a robustness check against such shifts before clinical use.
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.

Referee Report

4 major / 5 minor

Summary. The paper proposes LingoNMF, an extension of the periodic-layer NMF (PL-NMF) algorithm for separating heart and lung sounds, in which a LLaMA-2 language model analyzes features of the currently separated signals and suggests updated fundamental-frequency targets that enter a quadratic penalty term added to the α-NMF cost. The method is evaluated on two datasets (100 synthesized mixtures of real heart/lung sounds and 210 manikin recordings) and compared against standard NMF, α-NMF, and PL-NMF using SDR, SIR, and SAR metrics. The abstract claims that the approach 'consistently outperformed existing methods.' My reading of the manuscript is that this central claim is not supported: the modified cost function in Eq. (7) is not minimized by the stated multiplicative updates, the main hyperparameters were selected on the evaluation datasets, and Table 2 itself contains a clear counterexample to 'consistently outperforms.'

Significance. If the technical issues were resolved, the idea of using an LLM to adjust a penalty term in NMF for cardiorespiratory sound separation would be a novel and potentially interesting contribution, particularly because the authors release their code and use publicly available data. The paper also usefully extends PL-NMF with an adaptive fundamental-frequency feedback loop. However, as presented, the contribution is not established: the missing update derivation means the LLM feedback cannot be verified to affect the factorization, the hyperparameter tuning on test data undermines the comparison, and the reported results contradict the headline claim. These are load-bearing problems, not presentation issues, and they warrant rejection of the manuscript in its current form.

major comments (4)
  1. [2.3, Eq. (7) and Eq. (8)] The LingoNMF objective in Eq. (7) includes the penalty term λ_f ||f_est − f||^2, where f_est is defined in Eq. (8) as the argument that maximizes the PSD of the current row estimates. However, the algorithm described in Section 2.3 and Algorithm 1 uses the standard α-NMF multiplicative updates (3) and (4), which are derived for D(Y||AX) alone. The paper gives no gradient, auxiliary function, or update rule for the penalty term. Because an argmax is piecewise constant in the factorization variables, the derivative of f_est with respect to the NMF factors is zero almost everywhere, so the penalty term cannot steer the updates as written. Consequently, the algorithm does not minimize the stated cost function, and the separation gains reported in Table 2 cannot be attributed to the LLM-driven fundamental-frequency penalty.
  2. [3.2, Table 1, Fig. 3] The hyperparameters α, the number of layers L, and λ_f were selected by scanning values and choosing the configuration that yielded the highest SNR on the evaluation datasets (Section 3.2; Table 1; Fig. 3). This is a form of test-set selection: the reported advantage of LingoNMF over the baselines may reflect that these parameters were tuned to the evaluation data, whereas no equivalent tuning is reported for standard NMF, α-NMF, or PL-NMF. The comparison is therefore not a fair test of the proposed LLM feedback mechanism.
  3. [Table 2] The abstract's claim that LingoNMF 'consistently outperformed existing methods' is contradicted by Table 2. On Dataset Two heart sounds, LingoNMF achieves SDR 7.9 dB, while PL-NMF achieves 9.4 dB and α-NMF achieves 13.7 dB. This is a direct reversal of the claimed advantage, and it means the paper's own evaluation does not support the unqualified claim of consistent superiority. The claim should be restricted to the metrics and datasets where LingoNMF is actually better, or the experimental design and conclusions should be revised.
  4. [3.3, Table 2] No statistical uncertainty is reported for any of the SDR, SIR, or SAR values in Table 2. There are no standard deviations, confidence intervals, or numbers of independent runs. Without such information, it is impossible to judge whether small differences (e.g., lung SDR 8.9 vs. 8.4 dB in Dataset One) are significant, and the reverse result for heart sounds in Dataset Two cannot be assessed as a real failure versus noise. This is especially important because the paper's main claim depends on the quantitative comparisons in this table.
minor comments (5)
  1. [Eq. (2)] The α-divergence formula is garbled in the manuscript, with missing parentheses and unclear subscript/superscript ranges; it should be typeset carefully so the cost function is unambiguous.
  2. [Algorithm 1] The notation Y ← λ'.Y + λ. is not explained: it appears to be an affine normalization but the parameters λ' and λ. are not defined in the text, and the superscripts in the update equations are difficult to parse.
  3. [Section 2.4] The sentence 'we applied the BSS_EVAL toolbox to measure performance' should specify the version and the exact global or local criteria used, since BSS_EVAL can report several variants of SDR/SIR/SAR.
  4. [Section 2.3 / Fig. 5] The claim that the LLM provides 'detailed insights for disease prediction' is presented as a contribution but is not evaluated quantitatively anywhere in the paper; if it remains a claimed benefit, it needs a separate assessment beyond a single illustrative example.
  5. [References] References [3] and [4] (fever detection with infrared thermography and bipedal walking) appear unrelated to the topic of cardiorespiratory sound separation and should be removed or replaced with relevant prior work.

Circularity Check

2 steps flagged · score 6.0 of 10

LingoNMF's reported gains are partly circular: SNR-maximizing hyperparameters are selected on the test datasets, and the LLM penalty term is not connected to the multiplicative updates it is claimed to optimize.

  1. fitted input called prediction [Section 3.2 (Implementation and Setup); Table 1; Fig. 3; Table 2]
    "We tested various 𝛼, number of layers, and 𝜆3, selecting the configuration that yielded the highest source-to-noise ratio (SNR) values. ... We tested various values of 𝜆3, and selected 𝜆3=0.01 for optimal heart and lung sound separation (Fig. 3)."

    The same datasets and the same SNR/SDR metric used to report Table 2 were used to choose 𝛼, the number of layers, and 𝜆𝑓. LingoNMF's Table 2 scores are therefore the maximum of a parameter scan over the evaluation data, not independent predictions. The claim that LingoNMF consistently outperforms other methods is forced by the selection procedure: the configuration was picked to maximize the very quantity later reported as evidence. This is the fitted-input-called-prediction pattern.

  2. other [Section 2.3, Eqs. (7)-(8) vs. Eqs. (3)-(4); Table 2]
    "The modified cost function 𝐷′ is defined as (7), where 𝐷(𝒀||𝑨𝑿) represents the original temporal cost function, 𝜆𝑓 is scaling factor ... 𝐷′=𝐷(𝒀||𝑨𝑿)+𝜆𝑓‖𝒇est−𝒇‖². (7) ... It is estimated as the frequency that maximizes the power spectral density (PSD), calculated by (8) ... 𝑓est=𝑎𝑟𝑔𝑚𝑎𝑥... (8) The update rules of the Alpha NMF algorithm given by (3) and (4) [20,21]."

    The objective in (7) adds the penalty λ_f‖f_est−f‖², but the implemented updates (3)-(4) are the standard α-NMF rules for D(Y||AX) alone. No gradient, auxiliary function, or update for the penalty is derived, and f_est is an argmax over the PSD of the current estimates. The equations the algorithm actually runs therefore do not minimize the stated D′ by construction. Consequently the LLM feedback is mathematically inert: it cannot steer A or X, and the separation improvements in Table 2 cannot be attributed to the LLM penalty; they reduce to the unmodified PL-NMF updates under the test-selected hyperparameters of Section 3.2.

full rationale

The central NMF derivation is not a self-citation chain: PL-NMF is the authors' earlier published algorithm, LLaMA-2 is external, and no uniqueness theorem is imported to force the choice. The load-bearing circularity is empirical. First, hyperparameters (α, layers, λ_f) are selected on the same datasets and same SNR metric that Table 2 reports, so the 'consistent outperformance' is a best-of-scan result rather than an out-of-sample prediction. Second, the stated LLM-modified cost function (7) is not connected to the multiplicative updates (3)-(4) that the algorithm executes; the f_est argmax penalty has no derived gradient or update rule, so the feedback loop cannot change the factorization. Thus the claimed LLM-driven improvement reduces, by construction, to the unmodified NMF updates plus test-time tuning. This is partial circularity, not full definitional equivalence, because the factorization itself is real unsupervised processing. Note also that the abstract's 'consistently outperformed' is factually contradicted by Table 2 on Dataset Two heart sounds (LingoNMF SDR 7.9 vs. α-NMF 13.7); that is a correctness issue, not a circularity issue.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central comparison rests on domain priors about heart and lung periodicities, hand-selected hyperparameters, and an unproven equivalence between the modified cost function and the original update rules. There are no new physical entities.

free parameters (4)
  • lambda_f = 0.01
    Selected by scanning lambda_f values and picking the one with the highest heart and lung SNR (Section 3.2, Fig. 3).
  • alpha and layer count L = best SNR configuration from Table 1, e.g., alpha=0.5 with one layer for lung
    Authors tested alpha in {-1, 0, 0.5, 1, 2, 10} and layers 1 to 4, then selected the configuration with the highest SNR on the evaluation data.
  • initial fundamental frequency f0 = 50 Hz for both heart and lung
    Set by hand for both sources and dynamically updated by the LLM during separation (Section 3.2).
  • LLM generation settings = temperature 0.6, top_p 0.9, max_tokens 512
    Hand-chosen text generation parameters for LLaMA-2-7b-chat; no ablation shows they do not affect results.
assumptions (4)
  • standard math Nonnegative matrix factorization with alpha-divergence converges under the multiplicative updates (3) and (4).
    The paper relies on this convergence as the base optimizer for PL-NMF and LingoNMF.
  • domain assumption Heart sounds have average period 0.8 to 1.2 seconds and lung sounds 2 to 5 seconds, with normal fundamental frequencies near 50 Hz.
    This prior is used to select which separated row is heart versus lung and to initialize the penalty target in Eq. (7).
  • ad hoc to paper The LLM's suggested fundamental-frequency targets are a reliable optimizer for the penalty term.
    No validation is provided that the LLM outputs improve the objective; this is the core new mechanism of LingoNMF.
  • ad hoc to paper The original alpha-NMF multiplicative updates minimize the modified cost D' in Eq. (7).
    No update rule for the added penalty term is derived, and f_est depends on an argmax over the PSD, which is not differentiable in the usual NMF updates.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Large Language Model-based Nonnegative Matrix Factorization For Cardiorespiratory Sound Separation." pith.science (2026). https://pith.science/paper/44G7EA47

@misc{pith2026250205757,
  author       = {Pith},
  title        = {Pith review of: Large Language Model-based Nonnegative Matrix Factorization For Cardiorespiratory Sound Separation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/44G7EA47}},
  note         = {Machine review of arXiv:2502.05757}
}
read the original abstract

This study represents the first integration of large language models (LLMs) with non-negative matrix factorization (NMF), marking a novel advancement in the source separation field. The LLM is employed in two unique ways: enhancing the separation results by providing detailed insights for disease prediction and operating in a feedback loop to optimize a fundamental frequency penalty added to the NMF cost function. We tested the algorithm on two datasets: 100 synthesized mixtures of real measurements, and 210 recordings of heart and lung sounds from a clinical manikin including both individual and mixed sounds, captured using a digital stethoscope. The approach consistently outperformed existing methods, demonstrating its potential to significantly enhance medical sound analysis for disease diagnostics.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. NeuroMambaLLM: Dynamic Graph Learning of fMRI Functional Connectivity in Autistic Brains Using Mamba and Language Model Reasoning

    eess.IV 2026-02 unverdicted novelty 7.0 of 10

    NeuroMambaLLM learns dynamic brain graphs from fMRI with a Mamba model and a frozen LLaMA-3-8B with LoRA, reaching 72.12% accuracy on ABIDE I.

  2. SimMark: A Robust Sentence-Level Similarity-Based Watermarking Algorithm for Large Language Models

    cs.CL 2025-02 conditional novelty 6.0 of 10

    SimMark watermarks LLM text by rejection-sampling sentence pairs so their embedding similarity falls in a secret interval, and detects it with a soft-count z-test that survives paraphrasing.

  3. Blind Source Separation in Biomedical Signals Using Variational Methods

    eess.AS 2025-06 conditional novelty 4.0 of 10

    A VAE trained on mixed manikin heart and lung sounds produces distinct latent clusters and visually matching source spectrograms.

  4. Large Language Models and Non-Negative Matrix Factorization for Bioacoustic Signal Decomposition

    eess.AS 2025-07 reject novelty 2.0 of 10

    An NMF-plus-LLM pipeline separates overlapping heart and lung sounds and generates tentative clinical labels, but the demonstration is qualitative and lacks validation.

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages · cited by 4 Pith papers

  1. [4]

    CONCLUSION In this paper, we proposed a non-negative matrix factorization algorithm for blind source separation of heart and lung sounds enhanced by large language models (LLMs). Moreover, The integration of a large language model enhances the separation results in two unique ways: by providing detailed insights for disease prediction and by dynamically o...

  2. [19]

    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,

    Xiong, W., Ma, Y., & Ma, J. (2024). Dual-Graph Regularized Sparse Robust Adaptive Concept Factorization. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5059976 [20] Cichocki, A., Lee, H., Kim, Y.-D., & Choi, S. (2008). Non-negative matrix factorization with α-divergence. Pattern Recognition Letters, 29(9), 1433-1440. [21] Lin CJ. Projected grad...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.