Prediction models for functional outcomes in prolonged disorders of consciousness: a systematic review and meta-analysis.

All included studies were judged at high risk of bias, and 50% had high concerns in applicability. Every included study had fewer than 10 outcome events per predictor variable, which raises the risk that the models overfit the data. Many models used the same Coma Recovery Scale-Revised scores as both predictors and outcomes, which can inflate apparent accuracy. External validation was rare, with only three studies performing it, so generalizability to other settings is unclear. Several models did not report key performance metrics, so only a subset could be pooled. Neurophysiological models often excluded patients with metallic implants, cranial defects, agitation, or unstable vital signs, so they may not apply to many patients with prolonged disorders of consciousness. Most studies came from China and Italy, with China contributing 42.1% and Italy 26.3%, which may limit transferability to other health systems.

Do not read the pooled 82% sensitivity and 85% specificity as evidence that these models are ready for patient care. All included studies were at high risk of bias, only three had external validation, and the review concluded that no current model is sufficiently reliable for direct clinical use.