Predicting ICU Delirium with a Regularized Logistic Regression Model: Device-Support Burden as an Informative...

Background/Objectives: Delirium is prevalent among intensive care unit (ICU) patients and is associated with prolonged ventilation, longer ICU stays, increased mortality, and post-discharge cognitive impairment. Because most current prediction tools were developed using European cohorts, their transportability to other ICU settings remains unclear. The device support burden has received limited attention, and preprocessing data leakage remains a secondary methodological concern. Methods: This prospective study enrolled consecutive adults from the Internal Medicine ICU of Ankara Etlik City Hospital. Fifty predictors from electronic health records were documented at ICU admission or within the first 24 h. Device support variables were assessed as a distinct domain using ablation analysis. We developed an elastic net regularized logistic regression model and validated it using 5 × 4-fold nested cross-validation. Preprocessing was restricted to the training folds, and gradient boosting was used as a comparison method. We also evaluated a fixed exploratory 11-predictor reduced model. Results: Delirium was detected in 50.8% of patients. The full regularized model achieved an area under the receiver operating characteristic curve (AUROC) of 0.696 (95% CI: 0.613-0.767). The fixed exploratory reduced model had a higher apparent AUROC of 0.756 (95% CI: 0.684-0.819) than the other models. This apparent advantage was absent when feature selection was repeated within each outer training fold (nested reduced model AUROC 0.700). The full-model calibration slope was 0.741 (95% CI: 0.345-1.265). This confidence interval was wide and included the value of 1.0. The fixed reduced-model slope was 0.235 (95% CI: 0.068-0.886), indicating that the predicted probabilities were too extreme. Ablation analysis showed that device support variables added predictive value to clinical and laboratory variables, whereas environmental variables alone did not significantly discriminate between the two groups. Conclusions: In this single-center Turkish ICU cohort, regularized logistic regression showed moderate performance in predicting delirium. Nested feature selection did not reproduce the fixed reduced model's apparent discrimination advantage. Device support variables are best interpreted as markers of care complexity rather than direct causal factors. External validation and recalibration are required before clinical use. Because delirium onset and device support start times were unavailable, the model estimated delirium risk across the ICU stay from predictors documented at ICU admission or within the first 24 h and was not time-anchored.