Researchers challenge five-class tongue profile link to pain in type 2 diabetes

A methodological letter says uncertain class selection, possible misclassification and inconsistent multiple-testing correction weaken claims that tongue patterns represent clinically meaningful pain profiles.

Researchers have urged caution over a study that linked tongue characteristics with musculoskeletal pain in people with type 2 diabetes, arguing that the statistical profiles at the centre of the analysis may be less certain than the paper suggests. In a letter to the Journal of Pain Research, the authors said the original study by Zhang and colleagues offered an interesting exploratory look at tongue-related patterns, but that the choice of a five-class latent profile solution needed fuller justification before the groups could be treated as clinically meaningful. According to the letter, the issue is not whether five classes are impossible or wrong, but whether the evidence presented shows why that solution should be preferred over alternatives.

The methodological concern begins with model selection. Zhang and colleagues examined solutions from one to six classes, and the five-class model showed stronger fit than simpler versions, but the six-class model produced lower AIC and BIC values, while the Lo-Mendell-Rubin test remained significant across the range of solutions. The letter’s authors argue that these statistics do not converge on a single answer and that no single index, including entropy, should be used to decide class number on its own. They note that the decline in information criteria slowed after five classes, but say this kind of elbow judgement is informal and cannot substitute for a fuller account of why an added class should be rejected. They also point to simulation research showing that BIC, sample-size-adjusted BIC and likelihood-ratio tests such as the adjusted Lo-Mendell-Rubin and bootstrap likelihood-ratio test are among the more useful tools for class enumeration, while AIC and entropy are weaker guides.

The letter also questions how much confidence should be placed in the class assignments themselves. Latent profile analysis assigns people to groups on the basis of estimated probabilities, and researchers say later analyses can overstate certainty if they treat those assignments as fixed categories. The authors argue that average posterior probabilities and misclassification rates would help readers judge how stable the profiles really were. They also say the reported association between one profile and musculoskeletal pain may depend on how the classes were defined, since the meaning of any given profile can change if a different solution is chosen. In that context, they call for sensitivity analyses using plausible alternative class structures, especially the six-class model, and for methods that carry classification uncertainty forward into downstream testing.

A separate concern involves multiple testing. The original paper reportedly applied Benjamini-Hochberg correction to the fully adjusted model, but the profile-comparison results were presented with unadjusted values, leaving unclear whether the same correction framework was used for the key contrast linking one profile to pain. The letter says that if the same adjustment were applied across those comparisons, the result would no longer meet the usual threshold for statistical significance. The authors stop short of dismissing the finding, but say the paper should explain why one set of analyses was adjusted and another was not. They conclude that the profiles should be regarded as exploratory statistical patterns rather than established clinical phenotypes unless they can be reproduced in another sample and shown to add value beyond standard clinical features.