Across CHARLS and ELSA, pain was the strongest predictor of incident frailty in MASLD, with depressive symptoms mediating 74.0% and 46.1% of the effect

In two independent prospective national cohorts, pain was the strongest predictor of incident frailty among middle-aged and older adults with MASLD, contributing more to prediction than age, grip strength, waist circumference, or glycated hemoglobin. Prior pain–frailty evidence came largely from general older populations or cross-sectional designs; this study places pain within the specific metabolic liver disease population and shows that all three feature selection methods retained pain among 31 candidates, with a mean absolute SHAP of 0.184 versus 0.102 for the next-ranked feature. Prospective design with 3,622 CHARLS and 2,059 ELSA participants followed 7–8 years, three cross-checked feature selection methods, and a model frozen before external validation; the authors state that SHAP importance reflects contribution to the fitted model's predictions and is not evidence of causal predominance.

Pain remained independently associated with incident frailty in the fully adjusted model including the baseline frailty index, and the association was strongest among those with the lowest baseline frailty burden. The crude RR of 1.921 (CHARLS) and 1.945 (ELSA) attenuated to 1.219 and 1.310 after adjusting for the baseline frailty index, indicating that part of the crude association reflected baseline proximity to the frailty threshold while a residual association persisted beyond standard metabolic risk factors; stratification showed a gradient with RRs of 1.794 and 1.583 at FI <0.10. A prespecified four-model modified Poisson framework (log link, HC1 robust standard errors) applied verbatim in both cohorts, supported by sensitivity analyses using complete cases, a TyG-WC alternative steatosis definition, and the FI-29 outcome, with direction and magnitude remaining stable.

Depressive symptoms were the principal mediating pathway from pain to frailty, accounting for 74.0% of the total effect in CHARLS and 46.1% in ELSA. Seven candidate mediators were screened in CHARLS; depressive symptoms had the largest indirect effect (0.237) while grip strength showed a smaller indirect effect (0.093) and the remaining five were not statistically significant; the mediation analysis used the FI-28 outcome excluding CES-D items and adjusted for baseline depressive symptoms, removing overlap between mediator and outcome definition. Structural equation modeling (lavaan, WLSMV) fitted in five imputed datasets and pooled with Rubin's rules, with proportions mediated reported across imputations (CHARLS 68.8%–80.1%; ELSA 43.0%–48.1%); the authors note that mediation estimates rely on the usual no-unmeasured-confounding assumptions.

A logistic regression model based on 14 routinely available variables achieved an external validation AUC of 0.753, outperforming more complex algorithms such as support vector machines, neural networks, and gradient boosting, and is paired with an online risk calculator. Feature selection, algorithm selection, hyperparameter tuning, and the probability threshold were all determined within CHARLS, the model was frozen before any ELSA application, and SMOTE was applied only inside the training portion of each cross-validation fold so that validation folds, the test set, and the external cohort never saw synthetic samples; adding pain yielded a ΔAUC of 0.033 and continuous NRIs of 0.397 and 0.409. Internal test AUC 0.740 (95% CI 0.706–0.774), external validation AUC 0.753 (0.728–0.779), calibration slope 1.127, Brier score 0.160, and decision-curve net benefit up to a threshold probability of 0.49; the authors state that the online tool's risk categories are research-use examples and have not been clinically validated.

The results apply to middle-aged and older adults with LAP-defined MASLD who were below the frailty threshold at baseline (FI <0.25) and had final follow-up data, namely the CHARLS participants (median 56 years, 70% female) and ELSA participants (median 64 years, 53% female); in this setting pain can serve as a low-cost entry point for frailty risk stratification, paired with the online calculator for individualized risk communication and shared decision-making, and it suggests screening for depressive symptoms among those reporting pain to identify the highest-risk individuals. The authors note that the attributable fractions describe statistical attribution under the stated assumptions rather than the proportion of cases that treating pain would prevent, so the work offers hypotheses and tools for future interventional trials, finer-grained pain measures, and imaging-based MASLD definitions rather than treatment recommendations.

Pain was measured with a single binary item (current bodily pain in CHARLS versus frequent problems with pain in ELSA), capturing neither intensity, duration, location, nor interference, and the items differed between cohorts, which may introduce measurement error and cross-cohort heterogeneity; MASLD was defined by LAP rather than liver biopsy or transient elastography, and the sex-specific thresholds were validated mainly in Asian populations, leaving their performance in the English population less certain; four of the 33 frailty items were not harmonized between cohorts and the CES-D scoring differed (0/0.5/1 versus 0/1), which the authors addressed with the FI-29 sensitivity analysis; additionally, unmeasured confounding by physical activity, diet quality, or analgesic use, the possible bidirectional relationship between pain and frailty, the no-unmeasured-confounding assumption of the mediation analysis, and the use of only the final follow-up wave without inverse-probability-of-censoring weighting are open questions a reader should keep in mind when interpreting causal meaning. The available text is an incomplete version that does not include Figures 1–7 or Supplementary Tables S1–S19, so judgments about figure-level detail rest on the main-text descriptions alone.