Research report: Beyond functional disability: longitudinal joint trajectories of muscle strength and walking speed and their association with new-onset diabetes risk among older adults without ADL impairment

Older adults whose grip strength and walking speed follow persistently unfavorable paths over time face a markedly higher risk of developing diabetes, even before any impairment in daily activities appears, according to a new longitudinal analysis published in Frontiers in Endocrinology. The study, which drew on four large aging cohorts from the United States, England, continental Europe, and China, found that people in the least favorable joint trajectory groups had two-and-a-half to more than four times the hazard of new-onset diabetes compared with participants whose functional levels remained higher.

The research team, led by Xiangyang Chen, Xuan Wu, and Sizhe Feng of the First Affiliated Hospital of Wenzhou Medical University, combined data from the Health and Retirement Study (HRS), the English Longitudinal Study of Ageing (ELSA), the Survey of Health, Ageing and Retirement in Europe (SHARE), and the China Health and Retirement Longitudinal Study (CHARLS). From the initial pooled samples, stringent exclusions left 2,600 HRS participants, 3,043 ELSA participants, 1,533 SHARE participants, and 3,550 CHARLS participants. Everyone included was free of diabetes and free of impairment in activities of daily living—dressing, bathing, eating, transferring, toileting, and indoor walking—throughout the first three waves used to build the trajectories.

Walking speed was measured with standardized timed walks: an 8-foot test at usual pace in HRS, ELSA, and SHARE, and a 2.5-meter test in CHARLS, with the mean of two trials used in each case. Muscle strength came from handgrip dynamometry, with the maximum value from each of both hands summed into a total grip strength score. The investigators then applied group-based trajectory modeling, a finite mixture approach that sorts individuals into latent subgroups sharing similar longitudinal patterns, fitting models of one to four classes and selecting the optimal solution using the Bayesian Information Criterion, minimum subgroup proportions of five percent, and mean posterior probabilities of at least 0.70.

The resulting trajectory structures differed by cohort. In HRS, ELSA, and SHARE, two groups emerged: a reference group labeled high-level decline with low-level stability, and an unfavorable group labeled low-level decline with low-level rise. In CHARLS, the optimal solution paired a high-level decline with low-level decline reference group against a low-level decline with high-level decline group. The authors stress that these classes should be interpreted within each cohort rather than treated as directly equivalent categories across populations, since age distributions, measurement protocols, and population characteristics varied.

During median follow-up periods of 10, 12, 6, and 8 years in HRS, ELSA, SHARE, and CHARLS respectively, 341, 265, 170, and 448 new cases of physician-diagnosed diabetes were recorded. In fully adjusted Cox models that also applied inverse probability of treatment weighting to balance baseline differences in age, sex, race, education, wealth, body mass index, hypertension, drinking and smoking status, cancer, cardiovascular disease, hip fracture, physical activity, cognitive function, and depressive symptoms, the unfavorable trajectory carried an adjusted hazard ratio of 4.17 (95% confidence interval 3.10–5.61) in HRS, 3.85 (2.55–5.79) in ELSA, 3.38 (2.25–5.06) in SHARE, and 2.60 (1.87–3.61) in CHARLS, all with p-values below 0.001.

Because death can preclude a diabetes diagnosis and distort incidence estimates, the team repeated the analysis with Fine–Gray subdistribution hazard models treating mortality as a competing risk. The estimates barely moved: 3.88 in HRS, 3.37 in ELSA, 3.18 in SHARE, and 2.47 in CHARLS, each statistically significant. Discrete-time logistic regression using follow-up waves as the time unit produced similarly elevated odds ratios of 4.30, 4.20, 4.93, and 4.88 across the four cohorts. Stratified analyses by age, sex, body mass index, hypertension status, and physical activity level showed generally consistent associations, although in SHARE the risk estimate among participants aged 70 or younger (hazard ratio 2.22, confidence interval 0.82–6.00) did not reach significance, a pattern the authors attribute partly to limited statistical power in that older cohort.

Notably, single-domain trajectory models built on muscle strength alone or walking speed alone showed limited and inconsistent associations across cohorts, whereas the joint trajectories were consistently linked to diabetes risk. The authors argue this supports treating grip strength and gait speed as complementary dimensions of physical function: strength reflects neuromuscular capacity and physiological reserve, while walking speed integrates lower-extremity performance, balance, cardiorespiratory fitness, and neuromotor coordination. Biologically, skeletal muscle is the primary site of glucose disposal, and age-related strength loss often coincides with type II fiber atrophy, mitochondrial dysfunction, impaired insulin signaling, chronic low-grade inflammation, and ectopic lipid accumulation in muscle—mechanisms that could plausibly connect functional decline to insulin resistance.

As a complementary analysis, the researchers built 129 candidate machine learning models per cohort—nine individual algorithms including random forest, XGBoost, LightGBM, CatBoost, AdaBoost, support vector machines, and others, plus 120 stacking combinations—with feature selection via LASSO and the Boruta algorithm performed strictly within training data, and class imbalance handled with SMOTE and undersampling. Stacking ensembles outperformed single models, and SHAP analysis indicated that the joint trajectory features contributed to predictions across cohorts. In a cross-cohort validation, models trained on HRS and ELSA, validated on SHARE, and tested externally on CHARLS retained discriminative performance, with AdaBoost performing best on the external test set. The authors caution that clinical utility requires prospective validation.

The study has important limitations. As an observational analysis, it cannot establish causation, and residual confounding remains possible. Diabetes was identified through self-reported physician diagnosis rather than laboratory biomarkers such as fasting glucose or HbA1c, so undiagnosed cases were likely missed—a nondifferential misclassification that would tend to bias the observed associations toward the null. Differences in walking protocols, grip strength devices, healthcare systems, and diabetes detection practices across countries could introduce heterogeneity, and participants excluded for missing measurements differed systematically from those included on several characteristics, so selection bias cannot be fully excluded. The authors also note that effect estimates such as the higher hazard ratio in HRS compared with CHARLS reflect differences in population risk profiles, event rates, and follow-up duration rather than any preferable effect size; the key finding is the consistent direction of association. Within those constraints, the work suggests that repeatedly measuring grip strength and usual walking speed may offer a simple adjunct to conventional diabetes risk assessment in older adults, complementing rather than replacing standard metabolic screening.

Subject of Research: Longitudinal joint trajectories of muscle strength and walking speed and their association with new-onset diabetes risk in older adults

Article Title: Beyond functional disability: longitudinal joint trajectories of muscle strength and walking speed and their association with new-onset diabetes risk among older adults without ADL impairment

Article References: Chen, X., Wu, X., & Feng, S. (2026). Beyond functional disability: longitudinal joint trajectories of muscle strength and walking speed and their association with new-onset diabetes risk among older adults without ADL impairment. Frontiers in Endocrinology, 17, 1935223. https://doi.org/10.3389/fendo.2026.1935223

Image Credits: AI Generated

DOI: 10.3389/fendo.2026.1935223

Keywords: diabetes risk, muscle strength, walking speed, aging cohorts, group-based trajectory modeling, grip strength, physical function, HRS, ELSA, SHARE, CHARLS, machine learning