Machine Learning-Based Prediction Models for Delirium: A Systematic Review and Meta-Analysis

To critically appraise and quantify the performance studies by employing machine learning (ML) to predict delirium.

A systematic review and meta-analysis.

Articles reporting the use of ML to predict delirium in adult patients were included. Studies were excluded if (1) the primary goal was only the identification of various risk factors for delirium; (2) the full-text article was not found; and (3) the article was published in a language other than English/Chinese.

PubMed, Embase, Cochrane Library database, Web of Science, Grey literature, and other relevant databases for the related publications were searched (from inception to December 15, 2021). The data were extracted using a standard checklist, and the risk of bias was assessed through the prediction model risk of bias assessment tool. Meta-analysis with the area under the receiver operating characteristic curve, sensitivity, and specificity as effect measures, was performed with Metadisc software. Cochran Q and I² statistics were used to assess the heterogeneity. Meta-regression was performed to determine the potential effect of adjustment for the key covariates.

A total of 22 studies were included. Only 4 of 22 studies were quantitatively analyzed. The studies varied widely in reporting about the study participants, features and selection, handling of missing data, sample size calculations, and the intended clinical application of the model. For ML models, the overall pooled area under the receiver operating characteristic curve for predicting delirium was 0.89, sensitivity 0.85 (95% confidence interval 0.84‒0.85), and specificity 0.80 (95% confidence interval 0.81-0.80).

We found that the ML model showed excellent performance in predicting delirium. This review highlights the potential shortcomings of the current approaches, including low comparability and reproducibility. Finally, we present the various recommendations on how these challenges can be effectively addressed before deploying these models in prospective analyses.