Predicting dementia with routine care EMR data.

Area of Science:

• Neurology
• Medical Informatics
• Artificial Intelligence

Background:

• Dementia diagnosis often occurs late, limiting timely intervention.
• Predicting dementia risk early is crucial for patient management.
• Routine healthcare data offers a valuable, underutilized resource for predictive modeling.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting dementia onset.
• To enable early, non-invasive, and cost-effective pre-screening of at-risk individuals.
• To utilize existing Electronic Medical Record (EMR) data for dementia risk prediction.

Main Methods:

• Training ML models on structured and unstructured data from EMRs (diagnoses, prescriptions, medical notes).
• Developing individual models for each data type and a combined model.
• Employing human-interpretable ML techniques for clinical adoption.

Main Results:

• The combined ML model demonstrated generalizability across multiple healthcare institutions.
• The model achieved nearly 80% accuracy in predicting dementia onset within one year.
• Identified key predictors of dementia, including known factors and novel insights from medical notes.

Conclusions:

• Routine EMR data can be effectively used to build accurate, generalizable dementia prediction models.
• The developed ML model offers a scalable solution for early dementia pre-screening.
• Further clinical investigation of novel predictors from unstructured data is warranted.