Predicting 5-Year Survival and Mortality in Dementia Patients: A Data-Driven Approach Using XGBoost for Enhanced Care and Resource Allocation.

Abstract

Objective: This study develops an eXtreme Gradient Boosting (XGBoost) regression model to identify key predictors of mortality and 5-year survival in dementia patients, highlighting the role of comorbidities. The findings highlight key risk factors that may facilitate targeted adjustments in clinical care and resource allocation for high-risk patients.

Methods: We used Taiwan's National Health Insurance dataset to develop and validate an XGBoost model predicting 5-year survival in dementia patients aged 65 years or older. The cohort (n=6,556) was split into 80% for training, 10% for validation, and 10% for testing. A total of 24 variables, including comorbidities and demographic factors, were selected as predictors. Hyperparameters were tuned to optimize performance, with a learning rate of 0.1, 1,000 estimators, and a maximum depth of 10. Regularization techniques were applied to prevent overfitting.

Results: The XGBoost model achieved 81.86% accuracy in predicting 5-year survival, with a receiver operating characteristic area under the curve of 0.81 and a log loss of 0.61. Of the 37 initial features, 24 were included, and the top 10 predictors were nasogastric tube insertion, chronic kidney disease, cancer, lung disease, urinary tract infection, fracture, peripheral vascular disease, antidepressant use, hypertension, and upper gastrointestinal issues.

Conclusion: The XGBoost model effectively predicts 5-year survival in dementia patients, identifying key predictors that can guide targeted care, preventive strategies, and healthcare resource planning.