Validation of Random Forest Machine Learning Models to Predict Dementia-Related Neuropsychiatric Symptoms in...

Area of Science:

• Gerontology
• Psychiatry
• Machine Learning

Background:

• Neuropsychiatric symptoms (NPS) significantly contribute to the social burden of dementia.
• The role of NPS in dementia is often underestimated in clinical practice.

Purpose of the Study:

• To validate predictive models for identifying psychotic and depressive symptoms in dementia patients.
• To inform decision-makers by providing tools to estimate NPS prevalence in population databases.

Main Methods:

• Utilized electronic health records of 4,003 dementia patients.
• Applied random forest machine learning algorithms to build predictive models.
• Assessed model performance using calibration and discrimination metrics on a test set of 1,000 patients.

Main Results:

• Neuropsychiatric symptoms were present in 58% of patients.
• The psychotic symptom model achieved an area under the curve of 0.80; the depressive model achieved 0.74.
• Model accuracy decreased when symptom probability was below 25%; key predictors included medication use and sedation levels.

Conclusions:

• Validated predictive models demonstrate good performance for estimating NPS prevalence in population databases.
• These models offer a valuable tool for understanding the scope of neuropsychiatric symptoms in dementia.