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

The American Society of Anesthesiologists (ASA) physical status classification remains the cornerstone of preoperative risk assessment but is limited by its subjective nature and only moderate inter-rater reliability. Artificial intelligence (AI) models have recently been applied to automate ASA prediction; however, the magnitude and reliability of their agreement with that of clinicians remain unclear. This systematic review and meta-analysis aimed to quantitatively synthesize the current evidence regarding the diagnostic performance of AI-based ASA classification systems.

A comprehensive search of the PubMed, Embase, Scopus, Web of Science, and CENTRAL (from inception to October 2025) databases was conducted. Eligible studies compared AI-predicted and clinician-assigned ASA classifications in real or simulated preoperative datasets. Study quality was assessed using the PROBAST+AI tool. The certainty of the evidence was graded using the GRADE framework. The primary outcome was inter-rater agreement measured using the quadratic weighted kappa (QWK); secondary outcomes included accuracy, sensitivity, and specificity. Random-effects meta-analysis and meta-regression analyses were performed.

Thirteen studies involving 402,336 cases were analyzed. The pooled QWK was 0.69 (95% CI 0.62-0.76), representing moderate-to-substantial agreement with human assessors. Pooled accuracy was 0.66, sensitivity 0.51, and specificity 0.78. Subgroup analysis showed significantly higher agreement for LLMs than for traditional ML models (P=0.04). Meta-regression revealed lower performance in studies combining adult and pediatric data.

This meta-analysis demonstrated that AI systems, particularly LLMs, showed moderate-to-substantial concordance with clinicians in assigning ASA physical status classifications, suggesting their potential utility as reliable adjuncts for preoperative risk stratification.