Machine learning models identify predictive features of patient mortality across dementia types

Area of Science:

• Neuroscience
• Gerontology
• Medical Informatics

Background:

• Dementia care presents challenges due to varied disease progression and outcomes.
• Predictive models are crucial for identifying patients at high risk of mortality.
• Understanding mortality risk factors across different dementia types is essential.

Purpose of the Study:

• To develop machine-learning models for predicting dementia patient mortality at various survival thresholds.
• To identify key predictors of mortality in dementia patients.
• To conduct dementia-type-specific analyses for personalized risk assessment.

Main Methods:

• Utilized a large dataset (45,275 participants, 163,782 visits) from the U.S. National Alzheimer's Coordinating Center (NACC).
• Developed multi-factorial XGBoost models to predict mortality at 1-, 3-, 5-, and 10-year thresholds.
• Performed stratified analyses using dementia-type-specific models.

Main Results:

• Models achieved an AUC-ROC over 0.82 with nine parsimonious features.
• Key predictors included dementia-specific neuropsychological tests, with minimal influence from general age-related conditions.
• Stratified analyses revealed shared and distinct mortality predictors across eight dementia types, with specific clustering observed (e.g., vascular dementia with depression).

Conclusions:

• Machine-learning models can effectively predict dementia patient mortality using limited clinical features.
• Dementia-type-specific models enhance the ability to manage heterogeneous patient populations.
• This approach supports personalized clinical management by flagging at-risk individuals.