Five-year dementia prediction and decision support system based on real-world data

Area of Science:

• Medical Informatics
• Machine Learning
• Neuroscience

Background:

• Dementia risk prediction models often use volunteer data, limiting real-world applicability.
• Electronic Health Records (EHR) offer realistic data but require extensive processing for clinical use.
• Alzheimer's disease and related dementias are complex, multifactorial neurodegenerative conditions.

Purpose of the Study:

• To develop and validate a machine learning (ML) based risk prediction model for dementia using real-world EHR data.
• To identify prognostic rulesets for dementia based on clinical characteristics from multimodal EHR data.
• To improve the clinical interoperability and accuracy of dementia prediction.

Main Methods:

• Utilized a ten-year export of multimodal EHR data from Johns Hopkins Health System.
• Developed an interpretable binary classification model for predicting dementia onset within a five-year period.
• Employed 5-fold cross-validation for model evaluation.

Main Results:

• The model achieved a mean test accuracy of 0.722 (95% CI: 0.722-0.723).
• The model attained an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.795 (95% CI: 0.794-0.795).
• Identified prognostic rulesets for dementia based on clinical characteristics.

Conclusions:

• Machine learning models utilizing real-world EHR data can accurately predict dementia onset.
• Multimodal data analysis and modeling of combined effects are crucial for understanding dementia risk pathways.
• The developed model demonstrates potential for improved clinical decision support in dementia care.