National risk prediction model for perioperative mortality in non-cardiac surgery

Background Many multivariable models to calculate mortality risk after surgery are limited by insufficient sample size at development or by application to cohorts distinct from derivation populations. The aims of this study were to validate the Surgical Outcome Risk Tool (SORT) for a New Zealand population and to develop an extended NZRISK model to calculate 1-month, 1-year and 2-year mortality after non-cardiac surgery. Methods Data from the New Zealand National Minimum Data Set for patients having surgery between January 2013 and December 2014 were used to validate SORT. A random 75 per cent split of the data was used to develop the NZRISK model, which was validated in the other 25 per cent of the data set. Results External validation of SORT in the 360 140 patients who underwent surgery in the study period showed good discrimination (area under the receiver operating characteristic curve (AUROC) value of 0 center dot 906) but poor calibration (McFadden's pseudo-R-2 0 center dot 137, calibration slope 5 center dot 32), indicating it was invalid in this national surgical population. Internal validation of the NZRISK model, which incorporates sex and ethnicity in addition to the variables used in SORT for 1-month, 1-year and 2-year outcomes, demonstrated excellent discrimination with AUROC values of 0 center dot 921, 0 center dot 904 and 0 center dot 895 respectively, and excellent calibration (McFadden's pseudo-R-2 0 center dot 275, 0 center dot 308 and 0 center dot 312 respectively). Calibration slopes were 1 center dot 12, 1 center dot 02 and 1 center dot 02 respectively. Conclusion The SORT performed poorly in this national population. However, inclusion of sex and ethnicity in the NZRISK model improved performance. Calculation of mortality risk beyond 30 days after surgery adds to the utility of this tool for shared decision-making.