JMIR Publications

ABSTRACT

Background:

Fatigue is increasingly recognized as a clinically important syndrome among patients with stroke. Although a growing number of prediction models have been developed for this population, evidence regarding their methodological rigor, predictive performance, and generalizability remains fragmented.

Objective:

This study aimed to evaluate and characterize existing models designed to detect or predict fatigue in patients with stroke.

Methods:

We systematically searched PubMed, Embase, Web of Science, the Cochrane Library, China National Knowledge Infrastructure, Wanfang Database, VIP Database, and the Chinese Biomedical Literature Database from inception to May 2026. Random-effects meta-analyses were performed using the Hartung–Knapp–Sidik–Jonkman method to synthesize model performance metrics, including the pooled area under the receiver operating characteristic curve. Subgroup and sensitivity analyses were conducted to explore sources of heterogeneity. The robustness of findings from small studies was assessed using funnel plots, Egger’s test, and Deeks’ funnel plot asymmetry test.

Results:

Twelve studies comprising 34 diagnostic models were included. In the training sets, the pooled area under the receiver operating characteristic curve was 0.84（95%CL 0.77-0.90）, with a sensitivity of 0.75（95%CL 0.63-0.84）and a specificity of 0.84（95%CL 0.76-0.89）. In the internal validation sets, the corresponding estimates were 0.83（95%CL 0.78-0.87）for the area under the curve, 0.78（95%CL 0.70-0.85）for sensitivity, and 0.80（95%CL 0.68-0.88） for specificity. In the external validation sets, the area under the curve was 0.82（95%CL 0.74-0.91）, with a sensitivity of 0.71（95%CL 0.64-0.78） and a specificity of 0.79（95%CL 0.73-0.83）. Subgroup analyses indicated that, in the training sets, models developed with sample sizes of 200 or more achieved a significantly higher area under the curve than those based on fewer than 200 participants (0.88 vs. 0.74; P < 0.001). With respect to validation strategy, models subjected to both internal and external validation also showed superior discrimination compared with those evaluated by internal validation alone (0.91 vs. 0.80; P < 0.001). No statistically significant differences were observed across modeling approaches, data sources, or study designs in the training, internal validation, or external validation sets (all P > 0.05).

Conclusions:

The available models showed generally favorable discriminatory performance. Their clinical applicability, however, remains constrained by a high risk of bias and the limited use of external validation. Future work should therefore prioritize rigorously designed, prospective, multicenter studies to develop and validate more robust prediction models.