Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Explaining Recovery Trajectories of Older Adults Post Lower-Limb Fracture Using Modality-wise Multiview Clustering and Large Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A pipeline that clusters each sensor modality separately and asks a large language model to name the clusters produces labels that line up with clinical recovery scores for most modalities.

desk verdict New public dataset and a sensible workflow, but the statistical validation is undermined by day-level tests on forward-filled scores from 10 participants. read the letter →

arxiv 2506.12156 v1 pith:D2BZT6BK submitted 2025-06-13 cs.LG cs.AIcs.CV

Shehroz S. Khan , Ali Abedi , Charlene H. Chu This is my paper · ORCID
classification cs.LGcs.AIcs.CV
keywords multiviewclusteringlargelanguagemodelscontext-awarepromptingmultimodalsensordataolderadultslower-limbfractureclinicalvalidationunsupervisedlearning
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 claims that a fully unsupervised pipeline can turn daily multimodal sensor data into clinically meaningful descriptions of how older adults recover from lower-limb fractures at home. The pipeline splits the data by sensor modality, clusters each modality separately with K-means, and then asks a large language model, given context-aware prompts, to name each cluster. The authors check those names against biweekly clinical scores for social isolation, hip and knee function, and mobility, and report that for motion, heart rate, sleep, and step modalities the clusters differ significantly on most scores, while acceleration does not. If the claim holds, clinicians could flag at-risk patients from sensor streams alone, without manual chart review or self-reports.

What carries the argument

The load-bearing mechanism is the pairing of modality-wise K-means clustering with context-aware LLM prompting. Each sensor view is clustered separately, cluster centers are converted back to the original feature scale, and a prompt containing the feature names, their units, the patient context, and the cluster centers is given to the LLM, GPT-4o, to produce a short label and description. The labels are then validated statistically by testing whether the clinical-score distributions differ across clusters, using Shapiro-Wilk tests to choose between t-test or ANOVA and Mann-Whitney U or Kruskal-Wallis tests.

What would settle it

Re-run the same modality-wise clustering and clinical-score comparisons with participant as a random effect, or with one value per participant per assessment window. If most p-values no longer reach 0.05, the reported associations are an artifact of repeated days rather than evidence that the clusters carry clinical information; if they survive, the claim is supported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that modality-wise multiview clustering followed by LLM-generated cluster labels yields labels that are statistically associated with clinical outcomes for most sensor modalities. Two clusters are found for position, motion, step, and acceleration; five for heart rate; four for sleep. Using Mann-Whitney U and Kruskal-Wallis tests, the authors find significant differences between clusters on SIS, OHS, OKS, and TUG for heart rate, sleep, and step; for motion, SIS, OKS, and TUG are significant; for position, only SIS and TUG are significant, and the two clusters are extremely unbalanced, with 558 days in one cluster and 2 in the other. The acceleration view shows no significant association with any clinical score, which the authors attribute to device sensitivity and inconsistent wear. The paper concludes that most inferred cluster labels support the validity of the method for interpreting sensor data.

Load-bearing premise

The load-bearing premise is that the 560 day-level rows can be treated as independent samples in the statistical tests, even though they come from only 10 participants and each biweekly clinical score is assigned to 14 consecutive days.

Editorial extensions

If this is right

  • If the method works as claimed, clinicians could monitor recovery trajectories from wearable and home sensor data without waiting for scheduled clinical assessments or self-reports.
  • The modality-wise design shows which sensor streams carry clinical signal: motion, heart rate, sleep, and step data separate recovery groups, while acceleration in this dataset does not.
  • The failure of the acceleration view points to feature extraction and data quality as the bottleneck, motivating better sensors or representations before LLM labeling is useful.
  • The paper's proposed next step of combining modality-wise clusterings into a final overall clustering could support patient-level risk stratification rather than day-level labels.

Reading between the lines

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

  • In my reading, the headline claim depends on treating each of the 560 day rows as an independent observation; a reanalysis that accounts for repeated days per participant would be needed to know how much of the reported association is real.
  • The same pipeline could be tested by asking clinicians to name clusters from the same cluster centers and comparing their labels with the LLM's, which would separate the value of the clustering from the value of the language model.
  • Because the position view's significance rests on a single two-day cluster, any practical use of that view would need a stricter test or a larger cohort.
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

3 major / 5 minor

Summary. The paper proposes an unsupervised pipeline that clusters 560 day-level multimodal sensor records from 10 older adults separately by sensor modality using K-means, uses GPT-4o with context-aware prompts to label the clusters, and validates the clusters against bi-weekly clinical scores (SIS, OHS, OKS, TUG) through Mann–Whitney U and Kruskal–Wallis tests. The central claim is that most modality-specific cluster labels are statistically associated with clinical outcomes, confirming the approach's efficacy. The report finds this claim unsupported: the statistical tests pseudo-replicate 10 participants into 560 dependent rows, the position-view evidence rests on a two-day cluster, and the tests validate clusters rather than the LLM labels.

Significance. If the claims were valid, the pipeline would be a useful, low-cost way to make unsupervised sensor clustering interpretable for clinicians. The paper has strengths: a public dataset, publicly available code, a fixed random seed for reproducibility, and a clearly described prompt template. The validation avoids circularity because neither clustering nor LLM prompting uses clinical scores. However, the load-bearing statistical evidence is invalid due to non-independence, and the label-level claim is not tested, so the current paper does not establish the stated contribution.

major comments (3)
  1. [§4.3, Table 3; §3.2] The Mann–Whitney U and Kruskal–Wallis tests in Table 3 treat 560 day-level rows as independent observations, but Section 3.2 states that clinical assessments were collected bi-weekly and forward-filled to the preceding 14 days, and the data come from only 10 participants. The same clinical score is therefore duplicated on 14 consecutive rows per participant, and within-participant correlation is ignored, inflating the effective sample size by roughly an order of magnitude and making the reported p-values (e.g., p < 0.0001) untrustworthy. The statement in Section 4.3 that 'the number of samples per cluster was sufficient' does not address independence, which is a requirement of both tests. A valid analysis would use participant or assessment occasion as the unit, or a mixed-effects model with participant random effects; the paper reports no such analysis.
  2. [§4.3, Table 3, Position row] The position view reports statistically significant SIS (p = 0.0360) and TUG (p = 0.0499) results, yet the comparison is between clusters containing 558 and 2 of the 560 days, and the table itself marks Cluster 2 as NaN due to insufficient data. A Mann–Whitney U test with two observations in one group cannot yield reliable inference, so these p-values should not be counted as evidence for the method. The same row also illustrates why the position view's clinical differences (Figure 2a) are driven by a single two-day cluster.
  3. [Abstract; §4.1–4.3] The statistical tests compare clinical score distributions between clusters, but they do not test whether the LLM-generated cluster labels (e.g., 'Low Mobility with Extended Outdoor Time') are semantically correct descriptions of those clusters. The abstract's claim that the results demonstrate 'statistical significance of most modality-specific cluster labels generated by the large language model with respect to clinical scores' overstates what was measured. At most, the analysis validates the clustering, not the labels; validating the labels would require an evaluation of label semantics against clinical scores or expert judgment.
minor comments (5)
  1. [§4.3, Table 3] The table caption says that columns Cluster1–5 show the number of clusters per modality, but the entries are normality-test outcomes (Yes/No/NaN), not counts; the caption should describe the actual content.
  2. [§4.1] The manuscript says 'the best clusters were recorded' after varying K from 2 to 15, but no criterion for selecting the best K is given; specify the index used (e.g., silhouette) and report its value for each view.
  3. [§3.2] The paper removes the 'participant' column, which prevents any participant-stratified analysis and is part of the pseudo-replication problem; retaining it would allow mixed-effects or cluster-bootstrap analyses.
  4. [References] References [20] and [21] contain '[Insert Date You Accessed]' placeholders, and references [6] and [15] are the same paper (Brown et al., 2020) cited as different entries with different first authors.
  5. [Figure 2] The caption says 'mean distribution of various clinical scales,' which is ambiguous; clarify whether the panels show boxplots, means with error bars, or something else, and label the axes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the cluster labels are derived from sensor data and validated against clinical scores that are not inputs to clustering or LLM prompting.

full rationale

The paper's derivation chain is self-contained with respect to its central claim. K-means clustering is applied separately to sensor-derived feature views, and the resulting cluster centers are converted back to original scale and given to GPT-4o through context-aware prompts that contain no clinical-score information (Section 3.3, Section 4.1). The validation step then compares the distribution of clinical scores (SIS, OHS, OKS, TUG) across the resulting clusters (Section 4.2, Section 4.3). Because the clinical scores are explicitly excluded from the clustering features and from the LLM prompts, the significant p-values in Table 3 are not forced by construction; they are an external check on the cluster assignments. The self-citations to the MAISON-LLF dataset [14,25] and to earlier clustering work [1] are references to data and standard methods, not load-bearing arguments that reduce the claimed result to an unverified prior conclusion. The acknowledged limitations, such as fixed random seed, LLM label variability, and the non-significant acceleration view, are stated as limitations rather than hidden inputs. A concern about pseudo-replication from forward-filled clinical scores is a statistical validity issue, not a circularity issue, because the forward filling does not make the clinical scores a function of the cluster assignments or of the sensor features used for clustering. Therefore no step in the paper's derivation is equivalent to its inputs by construction, and no circularity score above zero is warranted.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on routine statistical tests whose validity depends on the independence assumption, which is violated by the study design, and on an unspecified cluster-count selection. The analysis introduces no new entities and fits only the number of clusters per view plus a hidden random seed.

free parameters (2)
  • Number of clusters K per modality view = position=2, motion=2, heart rate=5, sleep=4, step=2, acceleration=2
    Selected by running K=2..15 and recording the 'best' clusters; no criterion (e.g., silhouette) is stated in Section 4.1, so the choice is not justified.
  • Random seed for K-means = unspecified fixed value
    Section 3.3 fixes the random state for reproducibility, but the value is not reported, limiting exact reproduction.
assumptions (4)
  • domain assumption Bi-weekly clinical scores are forward-filled to all 14 preceding days and treated as daily ground truth
    Section 3.2: 'all 14 preceding days were assigned the same values for clinical assessments.' This creates 14 copies of the same score and inflates the effective sample size.
  • domain assumption Day-level observations are independent samples for hypothesis testing
    Section 4.3 applies Mann-Whitney U and Kruskal-Wallis tests to 560 day-level observations from 10 participants without accounting for within-participant correlation.
  • ad hoc to paper The 'best' number of clusters for each view is identifiable
    Section 4.1 says K varied from 2 to 15 and 'the best clusters were recorded' but does not state the selection metric.
  • standard math Standard statistical test assumptions (Shapiro-Wilk, Mann-Whitney U, Kruskal-Wallis)
    Invoked in Section 4.3; normality checks were performed via Shapiro-Wilk and the suitability of each test is assumed from the stated conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explaining Recovery Trajectories of Older Adults Post Lower-Limb Fracture Using Modality-wise Multiview Clustering and Large Language Models." pith.science (2026). https://pith.science/paper/D2BZT6BK

@misc{pith2026250612156,
  author       = {Pith},
  title        = {Pith review of: Explaining Recovery Trajectories of Older Adults Post Lower-Limb Fracture Using Modality-wise Multiview Clustering and Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2BZT6BK}},
  note         = {Machine review of arXiv:2506.12156}
}
read the original abstract

Interpreting large volumes of high-dimensional, unlabeled data in a manner that is comprehensible to humans remains a significant challenge across various domains. In unsupervised healthcare data analysis, interpreting clustered data can offer meaningful insights into patients' health outcomes, which hold direct implications for healthcare providers. This paper addresses the problem of interpreting clustered sensor data collected from older adult patients recovering from lower-limb fractures in the community. A total of 560 days of multimodal sensor data, including acceleration, step count, ambient motion, GPS location, heart rate, and sleep, alongside clinical scores, were remotely collected from patients at home. Clustering was first carried out separately for each data modality to assess the impact of feature sets extracted from each modality on patients' recovery trajectories. Then, using context-aware prompting, a large language model was employed to infer meaningful cluster labels for the clusters derived from each modality. The quality of these clusters and their corresponding labels was validated through rigorous statistical testing and visualization against clinical scores collected alongside the multimodal sensor data. The results demonstrated the statistical significance of most modality-specific cluster labels generated by the large language model with respect to clinical scores, confirming the efficacy of the proposed method for interpreting sensor data in an unsupervised manner. This unsupervised data analysis approach, relying solely on sensor data, enables clinicians to identify at-risk patients and take timely measures to improve health outcomes.

Figures

Figures reproduced from arXiv: 2506.12156 by the authors.

Figure 1
Figure 1. Generation of cluster labels on multimodal data using multiview clustering [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. The mean distribution of various clinical scales for each clusters obtained [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 26 canonical work pages

  1. [1]

    Cluster center initialization algorithm for k-means clustering,

    S. S. Khan and A. Ahmad, “Cluster center initialization algorithm for k-means clustering,” Pattern recognition letters, vol. 25, no. 11, pp. 1293–1302, 2004

    work page 2004

  2. [2]

    Survey of clustering algorithms,

    R. Xu and D. Wunsch, “Survey of clustering algorithms,”IEEE Transactions on neural networks, vol. 16, no. 3, pp. 645–678, 2005

    work page 2005

  3. [3]

    A comprehensive survey of clustering algorithms,

    D. Xu and Y. Tian, “A comprehensive survey of clustering algorithms,”Annals of data science, vol. 2, no. 2, pp. 165–193, 2015

    work page 2015

  4. [4]

    A survey on multiview clustering,

    G. Chao, S. Sun, and J. Bi, “A survey on multiview clustering,”IEEE transactions on artificial intelligence, vol. 2, no. 2, pp. 146–168, 2021

    2021

  5. [5]

    Multi-view clustering: A survey,

    Y. Yang and H. Wang, “Multi-view clustering: A survey,”Big data mining and analytics, vol. 1, no. 2, pp. 83–107, 2018

    work page 2018

  6. [7]

    Deepseek llm: Scaling open-source language models with longtermism,

    X. Bi, D. Chen, G. Chen, S. Chen, D. Dai, C. Deng, H. Ding, K. Dong, Q. Du, Z. Fu et al., “Deepseek llm: Scaling open-source language models with longtermism,” arXiv preprint arXiv:2401.02954, 2024

    arXiv 2024

  7. [8]

    Wear- able devices and explainable unsupervised learning for covid-19 detection and mon- itoring,

    A. Hasasneh, H. Hijazi, M. A. Talib, Y. Afadar, A. B. Nassif, and Q. Nasir, “Wear- able devices and explainable unsupervised learning for covid-19 detection and mon- itoring,” Diagnostics, vol. 13, no. 19, p. 3071, 2023

    work page 2023

  8. [9]

    Unsupervised human activity recognition via large language models and iterative evolution,

    J. Gao, Y. Zhang, Y. Chen, T. Zhang, B. Tang, and X. Wang, “Unsupervised human activity recognition via large language models and iterative evolution,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 91–95

    work page 2024

Show all 31 references
  1. [10]

    Evaluating large language models as virtual annotators for time-series physical sensing data,

    A. Hota, S. Chatterjee, and S. Chakraborty, “Evaluating large language models as virtual annotators for time-series physical sensing data,” ACM Trans. Intell. Syst. Technol. , Sep. 2024, just Accepted. [Online]. Available: https://doi.org/10.1145/3696461

    2024 doi

  2. [11]

    Enhanced interpretation of novel datasets by sum- marizing clustering results using deep-learning based linguistic models,

    S. Verma, D. Kumaret al., “Enhanced interpretation of novel datasets by sum- marizing clustering results using deep-learning based linguistic models,”Applied Intelligence, vol. 55, no. 5, pp. 1–23, 2025

    2025

  3. [12]

    A human-centric and llm-enhanced explainable framework for sens- ing data exploratory analysis,

    E. Paraschou, “A human-centric and llm-enhanced explainable framework for sens- ing data exploratory analysis,” Ph.D. dissertation, ARISTOTLE UNIVERSITY OF THESSALONIKI, 2024

    2024

  4. [13]

    Integrating vision- language semantic graphs in multi-view clustering,

    J. Ke, Z. Wen, Y. Yang, C. Cui, Y. Ren, X. Pu, and L. He, “Integrating vision- language semantic graphs in multi-view clustering,” inProceedings of the Thirty- Third International Joint Conference on Artificial Intelligence, 2024, pp. 4273– 4281

    2024

  5. [14]

    MAISON-LLF: Multimodal AI-Based Sensor Platform for Older iNdividuals–Lower Limb Fracture,

    A. Abedi, C. Chu, and S. Khan, “MAISON-LLF: Multimodal AI-Based Sensor Platform for Older iNdividuals–Lower Limb Fracture,” Dataset, 2025, available online: https://zenodo.org/records/14597613

    2025

  6. [15]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Nee- lakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,” arXiv preprint arXiv:2005.14165, 2020. Title Suppressed Due to Excessive Length 15

    2005 arXiv

  7. [16]

    Aras human activity datasets in multiple homes with multiple residents,

    H. Alemdar, H. Ertan, O. D. Incel, and C. Ersoy, “Aras human activity datasets in multiple homes with multiple residents,” in2013 7th International Conference on Pervasive Computing Technologies for Healthcare and Workshops. IEEE, 2013, pp. 232–235

    2013

  8. [17]

    Summarizing clustering results using sentence proto- type based language models,

    A. Khare and D. Kumar, “Summarizing clustering results using sentence proto- type based language models,” in2023 3rd International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME). IEEE, 2023, pp. 1–6

    2023

  9. [18]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of Machine Learning Research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  10. [19]

    A unified approach to interpreting model predic- tions,

    S. M. Lundberg and S.-I. Lee, “A unified approach to interpreting model predic- tions,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NeurIPS), 2017, pp. 4765–4774

    2017

  11. [20]

    Gpt-4 technical report,

    OpenAI, “Gpt-4 technical report,” https://openai.com/research/gpt-4, 2023, ac- cessed: [Insert Date You Accessed]

    2023

  12. [21]

    Mistral 7b,

    A. Q. Jiang, E. Anger, C. Dancette, S. Edunov, N. Goyal, J. Heek, G. Izacard, E. M. Mhamdi, A. Rame, G. Wenzeket al., “Mistral 7b,” https://mistral.ai/news/ announcing-mistral-7b/, 2023, accessed: [Insert Date You Accessed]

    2023

  13. [22]

    Lifesnaps, a 4-month multi-modal dataset capturing unobtrusive snapshots of our lives in the wild,

    S. Yfantidou, C. Karagianni, S. Efstathiou, A. Vakali, J. Palotti, D. P. Giakatos, T. Marchioro, A. Kazlouski, E. Ferrari, and Š. Girdzijauskas, “Lifesnaps, a 4-month multi-modal dataset capturing unobtrusive snapshots of our lives in the wild,” Scientific Data, vol. 9, no. 1,...

    2022

  14. [23]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sas- try, A. Askell, P. Mishkin, J. Clark, S. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” arXiv preprint arXiv:2103.00020, 2021

    2021 arXiv

  15. [24]

    Fellbaum,WordNet: An Electronic Lexical Database

    C. Fellbaum,WordNet: An Electronic Lexical Database. MIT Press, 1998

    1998

  16. [25]

    Multimodal sensor dataset for monitor- ing older adults post lower-limb fractures in community settings,

    A. Abedi, C. H. Chu, and S. S. Khan, “Multimodal sensor dataset for monitor- ing older adults post lower-limb fractures in community settings,”arXiv preprint arXiv:2501.13888, 2025

    2025 arXiv

  17. [26]

    Psychometric evaluation of the social isolation scale in older adults,

    N. R. Nicholson Jr, R. Feinn, E. Casey, and J. Dixon, “Psychometric evaluation of the social isolation scale in older adults,”The Gerontologist, vol. 60, no. 7, pp. e491–e501, 2020

    2020

  18. [27]

    The oxford hip score: the pa- tient’s perspective,

    V. Wylde, I. D. Learmonth, and V. J. Cavendish, “The oxford hip score: the pa- tient’s perspective,” Health and quality of life outcomes, vol. 3, pp. 1–8, 2005

    2005

  19. [28]

    The oxford knee score; problems and pitfalls,

    S. L. Whitehouse, A. W. Blom, A. H. Taylor, G. T. Pattison, and G. C. Bannister, “The oxford knee score; problems and pitfalls,”The Knee, vol. 12, no. 4, pp. 287– 291, 2005

    2005

  20. [29]

    The timed “up & go

    D. Podsiadlo and S. Richardson, “The timed “up & go”: a test of basic functional mobility for frail elderly persons,” Journal of the American geriatrics Society, vol. 39, no. 2, pp. 142–148, 1991

    1991

  21. [30]

    Normative physical fitness scores for community-dwelling older adults,

    H.-T. Chen, C.-H. Lin, and L.-H. Yu, “Normative physical fitness scores for community-dwelling older adults,” Journal of Nursing Research, vol. 17, no. 1, pp. 30–41, 2009

    2009

  22. [31]

    Mobility, disability, and social engagement in older adults,

    A. L. Rosso, J. A. Taylor, L. P. Tabb, and Y. L. Michael, “Mobility, disability, and social engagement in older adults,”Journal of aging and health, vol. 25, no. 4, pp. 617–637, 2013

    2013

  23. [32]

    Spatiotemporal clustering: a review,

    M.Y.Ansari,A.Ahmad,S.S.Khan,G.Bhushan,andMainuddin,“Spatiotemporal clustering: a review,”Artificial Intelligence Review, vol. 53, pp. 2381–2423, 2020

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.