Performance of machine and deep learning models for predicting delirium in adult ICU patients: A systematic review

Purpose: To summarize and evaluate the methodological quality of primary studies focusing on the use of machine or deep learning- based prediction models for delirium in ICU patients. Methods: A systematic review was conducted to identify relevant studies using six electronic databases. The Prediction Model Risk of Bias Assessment Tool (PROBAST) was used for appraising the quality of included studies. Results: Eleven studies published between 2018 and 2024 met the inclusion criteria, with most published after 2021. The majority used retrospective designs and sample sizes varied widely (140–48451 participants). Logistic regression and random forest models consistently demonstrated strong predictive performance (AUROC 0.83–0.91). Advanced techniques such as recurrent neural networks and gradient boosting also showed promise, particularly in handling complex time-series ICU data. Methodological quality was generally robust; however, common limitations included overfitting, lack of external validation, and incomplete handling of missing data. Conclusion: Machine learning models, especially logistic regression and random forest, offer promising, interpretable approaches for delirium prediction in ICU settings. Advanced deep learning models may further enhance predictive accuracy. Nonetheless, methodological limitations − mainly retrospective designs and limited external validation- limit their clinical generalizability. Future studies should focus on prospective validation and improving methodological rigor to support clinical implementation.