External validity of four risk scores predicting 30-day mortality after surgery

BACKGROUND: Surgical risk prediction tools can facilitate shared decision-making and efficient allocation of perioperative resources. Such tools should be externally validated in target populations before implementation. METHODS: Predicted risk of 30-day mortality was retrospectively derived for surgical patients at Royal Perth Hospital from 2014 to 2021 using the Surgical Outcome Risk Tool (SORT) and the related NZRISK (n=44 031, 53 395 operations). In a sub-population (n=31 153), the Physiology and Operative Severity Score for the enumeration of Mortality (POSSUM) and the Portsmouth variant of this (P-POSSUM) were matched from the Copeland Risk Adjusted Barometer (C2-Ai, Cambridge, UK). The primary outcome was risk score discrimination of 30-day mortality as evaluated by area-under-receiver operator characteristic curve (AUROC) statistics. Calibration plots and outcomes according to risk decile and time were also explored. RESULTS: All four risk scores showed high discrimination (AUROC) for 30-day mortality (SORT=0.922, NZRISK=0.909, P-POSSUM=0.893; POSSUM=0.881) but consistently over-predicted risk. SORT exhibited the best discrimination and calibration. Thresholds to denote the highest and second-highest deciles of SORT risk (>3.92% and 1.52–3.92%) captured the majority of deaths (76% and 13%, respectively) and hospital-acquired complications. Year-on-year SORT calibration performance drifted towards over-prediction, reflecting a decrease in 30-day mortality over time despite an increase in the surgical population risk. CONCLUSIONS: SORT was the best performing risk score in predicting 30-day mortality after surgery. Categorising patients based on SORT into low, medium (80–90th percentile), and high risk (90–100th percentile) might guide future allocation of perioperative resources. No tools were sufficiently calibrated to support shared decision-making based on absolute predictions of risk.