Predicting future dementia from routine clinical MRI and linked healthcare data

Area of Science:

• Neuroimaging
• Machine Learning
• Public Health

Background:

• Early dementia detection is crucial for intervention and prevention.
• Current biomarkers are often invasive, costly, or not scalable for public healthcare.
• Routine brain MRI scans offer a widely available resource for dementia risk stratification.

Purpose of the Study:

• To assess the feasibility of using routine brain MRI scans for early dementia prediction.
• To develop and validate a machine learning model for dementia risk stratification.
• To ensure model reliability, interpretability, and safety for clinical application.

Main Methods:

• Retrospective case-control study using routine T1-weighted brain MRI and electronic health records.
• Support-vector-machine classifier with nested cross-validation for structural brain feature analysis.
• Distance-from-hyperplane (DFH) calibration to quantify prediction confidence.

Main Results:

• The model predicted dementia up to five years pre-diagnosis with an AUC of 0.71.
• Prediction accuracy improved closer to the time of diagnosis.
• Confidence-based stratification identified a high-confidence subgroup (35% of scans) with ~80% prediction accuracy.
• Model performance was robust across diverse NHS scanners and protocols.

Conclusions:

• Routine brain MRI data can predict future dementia years before clinical diagnosis.
• Confidence calibration enhances clinical interpretability and safety of machine learning models.
• This approach facilitates scalable early detection, risk stratification, and clinical trial recruitment.