Qifan Xiao,1 Shirui Zhou,2 Bin Tang,1 Yuqing Zhu1 

1General Practice, International Department, China-Japan Friendship Hospital, Beijing, People’s Republic of China; 2Graduate School, Beijing University of Chinese Medicine, Beijing, People’s Republic of China

Correspondence: Yuqing Zhu, International Department of China-Japan Friendship Hospital, No. 2 Yinghua East Street, Chaoyang District, Beijing, 100029, People’s Republic of China, Email circe_77@126.com

Background: Delirium superimposed on dementia (DSD) is a severe complication in older adults with dementia, marked by fluctuating cognition, inattention, and altered consciousness. Detection is challenging due to symptom overlap, yet it contributes to cognitive decline, prolonged hospitalization, and increased mortality. Identifying key risk factors and developing an accurate prediction model is crucial for timely intervention. This study aimed to establish a machine learning-based model to predict delirium risk, focusing on significant predictors to aid clinical decision-making.
Methods: We prospectively collected clinical data from 636 older dementia patients. Five machine learning algorithms—Extreme Gradient Boosting (XGB), Random Forest (RF), Multilayer Perceptron (MLP), Categorical Boosting (CB), and Logistic Regression (LR)—were used to construct prediction models. Feature importance was analyzed using SHapley Additive exPlanations (SHAP) to identify key risk factors. Data included demographic information, biochemical parameters, comorbidities, medication history, and Visual Analogue Scale (VAS) scores.
Results: The final analysis included 636 older dementia patients, with a mean age of 78.2 ±  6.3 years, of whom 187 (29.4%) developed delirium during hospitalization. The XGB model demonstrated the best performance, achieving the highest area under the receiver operating characteristic curve (0.930), accuracy (0.870), F1 score (0.892), and area under the precision-recall curve (0.989). The Brier score for the XGB model was 0.08. The SHAP method identified cerebrovascular disease, sedative drug use, hemoglobin levels, VAS score ≥ 4, superoxide dismutase, diabetes, hsCRP, hypertension, family presence, and hyperlipidemia as the most significant risk factors for delirium. The top 10 variables were used to construct a compact XGB model, which also exhibited good predictive performance.
Conclusion: This study developed a machine learning-based prediction model for delirium risk in older dementia patients, with the XGB model demonstrating the best performance. The identified key risk factors provide insights for early intervention, potentially improving delirium management in clinical practice.

Keywords: delirium, machine learning, dementia, older, prediction, risk factors