Systematic development and validation of a predictive model for major postoperative complications in the Peri-operative Quality Improvement Project (PQIP) dataset.

Complications are common following major surgery and are associated with increased use of healthcare resources, disability and mortality. Continued reliance on mortality estimates risks harming patients and health systems, but existing tools for predicting complications are unwieldy and inaccurate. We aimed to systematically construct an accurate pre-operative model for predicting major postoperative complications; compare its performance against existing tools; and identify sources of inaccuracy in predictive models more generally. Complete patient records from the UK Peri-operative Quality Improvement Programme dataset were analysed. Major complications were defined as Clavien-Dindo grade ≥ 2 for novel models. In a 75% train:25% test split cohort, we developed a pipeline of increasingly complex models, prioritising pre-operative predictors using Least Absolute Shrinkage and Selection Operators (LASSO). We defined the best model in the training cohort by the lowest Akaike's information criterion, balancing accuracy and simplicity. Of the 24,983 included cases, 6389 (25.6%) patients developed major complications. Potentially modifiable risk factors (pain, reduced mobility and smoking) were retained. The best-performing model was highly complex, specifying individual hospital complication rates and 11 patient covariates. This novel model showed substantially superior performance over generic and specific prediction models and scores. We have developed a novel complications model with good internal accuracy, re-prioritised predictor variables and identified hospital-level variation as an important, but overlooked, source of inaccuracy in existing tools. The complexity of the best-performing model does, however, highlight the need for a step-change in clinical risk prediction to automate the delivery of informative risk estimates in clinical systems.