Reconsidering Latent Profile Selection and Clinical Interpretation in Tongue-Imaging Research on Musculoskeletal Pain [Letter]

Dear editor

Zhang et al examined the relationship between tongue characteristics and musculoskeletal pain among patients with type 2 diabetes mellitus using principal component analysis (PCA) and latent profile analysis (LPA). Their study provides an exploratory approach to characterizing tongue-related patterns in this population. However, the interpretation of data-derived profiles as clinically meaningful subgroups depends critically on the transparency and robustness of the class-enumeration process. We would like to highlight several methodological considerations regarding the selection of the five-class solution and the interpretation of its association with musculoskeletal pain.¹

First, the Rationale for Selecting Five Classes Warrants Fuller Explanation

The authors evaluated models containing one to six classes. The reported five-class model had an Akaike information criterion (AIC) of 5473.894, a Bayesian information criterion (BIC) of 5596.347, an entropy of 0.836, and a smallest class comprising 15.5% of the sample. The six-class model had lower AIC (5356.156) and BIC (5500.476), while its entropy was slightly lower (0.833) and its smallest class accounted for 10.4% of participants. The Lo–Mendell–Rubin (LMR) likelihood-ratio test was reported as significant for every solution from two to six classes.¹

These findings do not establish that the six-class solution is preferable or that the five-class solution is invalid; rather, the reported criteria do not all converge on a single solution. Information criteria, likelihood-ratio tests, classification quality, class size, and substantive interpretability provide different types of evidence and should be considered together.^2,3 Entropy, in particular, describes the separation of assigned classes and should not independently determine the number of classes.³ A significant LMR result for the six-class model indicates that six classes fit better than five by that test, so the likelihood-ratio test cannot by itself identify where enumeration should stop.

The authors explain that the decrease in AIC and BIC slowed after five classes and that entropy was high with every class exceeding 15%. This rationale could be developed further, for three reasons. First, a slowing decrement is an informal elbow judgement; the five-to-six decrement remains substantial in absolute terms (ΔAIC 117.7; ΔBIC 95.9), and no formal test of the elbow is available. Second, the sample-size-adjusted BIC and the bootstrap likelihood ratio test, which simulation work suggests are comparatively reliable for class enumeration,² are not reported. Third, a requirement that every class exceed 15% is not standard guidance, and the smallest class in the six-class solution (10.4%, approximately 61 participants) would not ordinarily be considered too small to interpret.

We therefore do not suggest that the five-class solution is incorrect, and we agree that enumeration should not rest on a single index. A clearer account of the substantive considerations would nonetheless help readers evaluate the decision: did the additional class represent a substantively distinct tongue-characteristic pattern, or did it primarily subdivide an existing profile? Were there concerns about sparse classes, unstable estimates, or limited clinical interpretability?

Second, the Robustness of Profile Assignment and Downstream Associations Deserves Attention

LPA assigns individuals to profiles according to estimated probabilities, and treating assigned membership as an observed categorical variable in later analyses may not fully reflect classification uncertainty.^2,3 Although the reported entropy suggests reasonably differentiated profiles, entropy alone cannot establish that individual assignments are sufficiently certain for inference. Reporting average posterior probabilities and misclassification rates by profile would help readers judge the extent of classification uncertainty.

This issue is relevant because the authors used the fifth profile as the reference category and reported a statistically significant association between profile 4 and musculoskeletal pain (adjusted odds ratio (OR) 2.048; 95% confidence interval (CI) 1.148–3.655), whereas the other reported profile comparisons were not statistically significant.¹ The interpretation of this contrast depends partly on the stability and distinctiveness of the profiles. If the number or composition of classes changes, the meaning of “profile 4” and “profile 5” may also change, potentially affecting the estimated association.

We therefore encourage the authors to clarify whether the pain association was examined under alternative plausible class solutions, particularly the six-class model. Where feasible, sensitivity analyses could compare profile composition and the pattern of adjusted associations across candidate models, and approaches that carry classification uncertainty forward would further strengthen the inferential basis. The aim would be to establish whether the substantive interpretation is robust to reasonable modeling choices.

A related question concerns the treatment of multiplicity. The Methods state that results from the fully adjusted model were corrected for multiple testing using the Benjamini–Hochberg procedure, and the table of individual tongue parameters accordingly reports adjusted Q values. The tables reporting composite scores and profile comparisons, however, present unadjusted P values, and their notes do not mention correction. This matters because the profile comparison is central to the study’s conclusion: only the contrast for profile 4 was significant among the four reported comparisons (P = 0.015). If the same Benjamini–Hochberg procedure were applied across these four comparisons, the corresponding adjusted value would be approximately 0.060. We raise this not to suggest that the association is spurious, but because readers cannot tell whether this result is reported under the same multiplicity framework as the parameter-level analyses, or why it would be exempt. Stating the approach explicitly, and reporting adjusted values where applicable, would resolve the ambiguity.

Third, the Clinical Meaning of the Identified Profiles Should Remain Appropriately Exploratory

The study combines PCA-derived tongue-characteristic information with LPA to identify patterns in the observed sample.¹ Such data-driven profiles can generate hypotheses, but they should not automatically be treated as established clinical phenotypes. Their reproducibility and clinical utility require evidence beyond separation within the derivation sample. Methodological work on latent class and profile analysis emphasizes the importance of considering model adequacy and the substantive meaning of the resulting classes.^2,3

Accordingly, the five profiles would be more cautiously described as exploratory statistical patterns unless their stability is supported by further analyses or independent data. External validation in a separate sample would help establish whether similar profiles can be recovered and whether their relationship with musculoskeletal pain persists. It would also be valuable to know whether the profiles add information beyond conventional clinical characteristics.

Finally, the outcome definition merits consideration when translating the findings into clinical interpretation. Musculoskeletal pain can differ by anatomical location, distribution, and clinical presentation. A binary outcome that pools pain reports may be useful for an initial association analysis, but it cannot establish whether tongue-characteristic profiles relate similarly to different pain presentations. This does not negate the reported association but limits the specificity of the clinical conclusions that can be drawn from it. Since limb pain was the most frequently reported site and nearly a third of affected participants reported two or more sites,¹ analyses by pain location could help clarify the scope of the observed relationship if the data permit.

In summary, Zhang et al present an exploratory investigation of tongue-characteristic profiles and musculoskeletal pain in patients with type 2 diabetes mellitus.¹ A clearer justification for retaining five classes, together with sensitivity analyses addressing profile stability, classification uncertainty, the robustness of the pain association, and the treatment of multiplicity, would strengthen the methodological transparency and clinical interpretation of the findings. These clarifications would help distinguish sample-specific statistical patterns from reproducible profiles with potential value for future pain research.

Funding Statement

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Disclosure

The authors declare that there are no known competing financial interests or personal relationships that could appear to influence the work reported in this communication.