Artificial intelligence models for predicting ASA physical status classification: a systematic review and...

Review Approach involved a systematic search of five major databases including PubMed and Scopus from inception through October 2025. Investigators identified studies comparing automated predictions against human-assigned classifications in real or simulated datasets. The team applied the PROBAST+AI tool to evaluate the quality of every included publication. Certainty of evidence was determined using the GRADE framework to ensure rigorous synthesis. Statistical analysis utilized random-effects models to pool the primary outcome of quadratic weighted kappa. Meta-regression was conducted to explore factors influencing model performance across the diverse studies. The researchers calculated secondary outcomes including sensitivity, specificity, and overall accuracy for all models. This structured methodology allowed for a comprehensive quantitative assessment of current diagnostic performance.

Key Findings From the Literature show a pooled quadratic weighted kappa of 0.69, reflecting moderate-to-substantial agreement between models and clinicians. The analysis of 402,336 cases across thirteen studies provided the basis for these results. Pooled accuracy for the systems reached 0.66, while sensitivity was 0.51 and specificity was 0.78. Subgroup analysis revealed that large language models achieved significantly higher agreement than traditional machine learning models. This performance difference was statistically significant with a P-value of 0.04. Meta-regression indicated that combining adult and pediatric data resulted in lower predictive performance. The data demonstrate that digital systems can effectively replicate human risk stratification to a meaningful degree. These results highlight the varying capabilities of different computational architectures in clinical settings.

Synthesis and Implications indicate that artificial intelligence systems achieve moderate-to-substantial concordance with human clinicians when assigning risk classifications. The authors propose that these tools could function as reliable adjuncts for preoperative stratification. Large language models demonstrated superior performance compared to traditional machine learning approaches in this meta-analysis. The researchers suggest that combining diverse patient populations, such as adults and children, may negatively impact model performance. These findings support the integration of digital models into clinical workflows to assist with risk assessment tasks. The evidence suggests that automated systems provide a consistent alternative to subjective human evaluation. Future implementation should focus on the specific benefits of advanced language-based models in perioperative settings. This review provides a foundation for understanding the diagnostic capabilities of current computational risk assessment tools.