Machine learning models to predict onset of dementia: A label learning approach

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Informatics
• Neurodegenerative Disease Research

Background:

• Developing predictive models for neurodegenerative diseases is crucial for early intervention.
• Administrative and electronic health record data offer a rich source for predictive modeling.

Purpose of the Study:

• To develop and validate a machine learning model for predicting incident mild cognitive impairment, Alzheimer's Disease, and related dementias.
• To utilize structured data from administrative and electronic health records for dementia prediction.

Main Methods:

• A large cohort (n=121,907 patients, n=5,307,045 controls) was established using data within 2 years of diagnosis.
• Gradient boosting machine (LightGBM) was employed for model training, with cohorts matched on key demographics and utilization.
• Prediction models were evaluated using data from 3-8 years post-index date.

Main Results:

• The 2-year incident model achieved an Area Under the Curve (AUC) of 87% with 47% sensitivity on a held-out test set.
• The 3-year prediction model achieved 24% sensitivity (AUC 71%), decreasing to 15% sensitivity (AUC 72%) by year 8.
• Model performance indicates a potential for early detection of dementia cases.

Conclusions:

• The developed machine learning model demonstrates capability in discriminating incident dementia cases.
• This predictive tool holds promise for screening patients for clinical trial recruitment.
• The model can serve as a valuable asset for patient management strategies in dementia care.