Stroke sarcopenia risk models show promise but lack validation

Risk prediction models for sarcopenia after stroke are proliferating rapidly, but a new scoping review finds that almost none of them have been tested outside the data that created them. The analysis, published in Frontiers in Neurology, pulled together every study published up to October 2025 that attempted to forecast which stroke patients will develop the muscle-wasting condition known as stroke-associated sarcopenia. The verdict: the field is young, promising, and riddled with methodological gaps that limit how much clinicians can trust the numbers on offer.

The review team, led by Ying Xia, Ling Huang, Xin Li, and Xiaolu Zhong of China Three Gorges University, searched nine English and Chinese databases and screened 1,379 records down to 12 eligible studies published between 2022 and 2025. Together those studies produced 15 prediction models, drawing on samples that ranged from just 80 to 1,226 participants. Reported prevalence of stroke-associated sarcopenia varied enormously across the studies, from 15.8 percent to 65.8 percent, a spread the authors attribute largely to differences in how sarcopenia was defined and measured rather than to true differences in disease burden. Nine of the 12 studies came from China, with the remainder from Turkey and Japan.

Stroke-associated sarcopenia refers to reductions in skeletal muscle mass, strength, or physical function that emerge after stroke-related brain injury. It matters clinically because stroke survivors with sarcopenia face progressive muscle wasting and malnutrition driven by restricted mobility and disease-related catabolism, a cycle that raises infection risk, prolongs hospital stays, and worsens overall prognosis. Previous studies have estimated that between 32.1 and 68.4 percent of stroke patients develop sarcopenia, making early identification of high-risk patients a pressing clinical goal. Risk prediction models are meant to serve that goal by converting population-level data into individualized probabilities that can guide preventive intervention.

Methodologically, the review followed the Arksey and O’Malley five-stage scoping framework, with data extraction guided by the Joanna Briggs Institute manual and model classification based on the TRIPOD reporting statement. The included studies split into two camps. Eight developed multivariable diagnostic prediction models, while four evaluated only single or few biomarkers without building a full model. Study designs were dominated by case–control and cross-sectional approaches, with two cohort studies and two prospective observational studies also represented. Most models relied on parametric methods, chiefly logistic regression, which assumes a predefined, typically linear relationship between predictors and outcome and yields interpretable coefficients that can be converted into odds ratios or nomograms.

Two studies ventured beyond regression into machine learning. One employed both logistic regression and a decision tree, and another compared logistic regression against random forest and XGBoost. Non-parametric methods of this kind do not assume a fixed functional form, allowing them to capture complex non-linear relationships and interactions. But the review’s authors caution that such models fit training data more closely without necessarily generalizing better, and they are more susceptible to overfitting in small samples. Machine learning models, in particular, therefore demand rigorous internal and external validation before any clinical deployment.

Performance figures tell a mixed story. Seven multivariable diagnostic models reported area under the curve values ranging from 0.743 to 0.970, while the four biomarker-only studies reported AUCs between 0.681 and 0.830. An AUC above 0.7 is generally considered acceptable discrimination, and the review notes that the multivariable models generally showed good diagnostic discrimination for prevalent sarcopenia in stroke patients. Yet discrimination alone does not guarantee clinical usefulness, and the review found calibration metrics inconsistently reported across the board. Three studies reported only the Hosmer-Lemeshow test, three reported only calibration plots, two reported both, and four reported no calibration metrics at all. Decision curve analysis, which quantifies net benefit across threshold probabilities, was performed in only six studies.

The validation picture is the review’s starkest finding. Of the 12 studies, only one performed true external validation, and that study, by Wang and colleagues, saw its AUC drop from 0.891 in the development set to 0.743 during geographical external validation, a decline the authors interpret as likely evidence of overfitting. Seven studies performed model development only, using internal resampling or random splits, and four did not develop a multivariable model at all. When the review team applied the PROBAST risk-of-bias tool to the eight eligible model studies, only one, by Yan and colleagues, earned an overall low risk of bias rating; the remaining seven were rated high risk. Bias concerns clustered in the outcome and analysis domains, stemming from inconsistent sarcopenia definitions, missing blinding information, small samples, too few outcome events, unreported missing-data handling, inadequate validation, and overfitting.

The predictors these models drew on spanned six dimensions: demographic and lifestyle factors, muscle morphology and body composition, medical history and comorbidities, neurological and cognitive-psychological factors, nutritional and feeding-related factors, and laboratory indicators. Age appeared in five studies, body mass index in four, smoking history in three, and NIHSS stroke severity scores in four. Tube feeding status and nutritional status each appeared in three studies, nutritional status being measured with different instruments and cut-offs across studies, including the MNA-SF, SGA, and MQSGA scales. Laboratory markers included C-reactive protein, albumin, hemoglobin, the C-reactive protein-to-albumin ratio, high mobility group box 1, creatine kinase, and the neutrophil-to-lymphocyte ratio.

Several of these laboratory markers stand out as the review’s most novel contribution. The C-reactive protein-to-albumin ratio, which integrates systemic inflammation and nutritional status, achieved an AUC of 0.814 for probable sarcopenia in acute ischemic stroke patients in one study, positioning it as a potentially convenient bedside screening tool. The neutrophil-to-lymphocyte ratio reflects post-stroke neuroimmune activation, in which neutrophil recruitment and T lymphocyte decline signal a disrupted inflammatory balance that promotes muscle protein degradation through the NF-κB pathway and muscle atrophy-related E3 ubiquitin ligases. High mobility group box 1, a damage-associated molecular pattern released by necrotic cells during post-stroke neuroinflammation, binds RAGE and TLR4 receptors and activates inflammatory cascades that impair skeletal muscle satellite cell function. The review’s authors stress, however, that each of these markers is supported by only one small study and should be treated as promising predictors rather than validated ones.

Muscle morphology measures also surfaced as unexplored territory in prior reviews. One Turkish study used masseter muscle cross-sectional area as both a predictor and part of the outcome definition, a design overlap the review’s authors flag as likely inflating the reported AUC and demanding careful interpretation. Another study used temporal muscle thickness, and a third included right calf circumference. The review concludes that future research should conduct multicenter prospective studies using standardized sarcopenia diagnostic criteria, adhere strictly to the TRIPOD statement, construct multidimensional models combining inflammatory-immune, musculoskeletal metabolic, and simple morphological indicators, and pursue external validation before any of these tools enter routine clinical practice. Until then, the models on offer remain research artifacts rather than bedside instruments.

Subject of Research: Risk prediction models for stroke-associated sarcopenia

Article Title: Prediction models for stroke-associated sarcopenia: a scoping review

Article References: Xia, Y., Huang, L., Li, X., & Zhong, X. (2026). Prediction models for stroke-associated sarcopenia: a scoping review. Frontiers in Neurology, 17, Article 1934093. https://doi.org/10.3389/fneur.2026.1934093

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1934093

Keywords: stroke-associated sarcopenia, risk prediction models, scoping review, sarcopenia, stroke, machine learning, logistic regression, external validation, PROBAST, biomarkers, C-reactive protein-to-albumin ratio, TRIPOD