Development and external validation of an online interpretable machine-learning model for predicting delirium risk in...

ObjectiveDelirium is a frequent complication in acute heart failure and is associated with poor outcomes in the intensive care unit. Machine learning methods can leverage high-dimensional clinical data to support early risk stratification; however, models specifically designed to predict delirium in acute heart failure are limited, underscoring the need for reliable and clinically applicable tools.MethodsPatients with acute heart failure were identified using International Classification of Diseases codes, and delirium was assessed using the Confusion Assessment Method for the Intensive Care Unit. Baseline clinical data were subjected to feature selection using least absolute shrinkage and selection operator, Boruta and recursive feature elimination. Eight machine learning models were developed, and the optimal model was evaluated for discrimination, calibration and clinical utility, with SHapley Additive exPlanations used to interpret predictor contributions.ResultsTen key predictors were identified using three feature selection methods. The gradient boosting machine model demonstrated favourable discrimination and calibration in both internal and external validation cohorts, although the areas under the receiver operating characteristic curve of the neural network and Light Gradient Boosting Machine models were not significantly different from those of the gradient boosting model in the DeLong comparisons. SHapley Additive exPlanations identified the Glasgow Coma Scale score, Sequential Organ Failure Assessment score and sedative use as major contributors to risk prediction. The final model was implemented as an online risk calculator to provide individualised risk estimates.ConclusionsThe model showed acceptable performance for estimating delirium risk in patients with acute heart failure and may support early intensive care unit risk stratification. The web-based calculator enables individualised assessment; however, prospective multi-centre validation is needed before broad clinical implementation.