Technology and Dementia Preconference.

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Dementia Prediction

Background:

• AI models for dementia prediction often use curated research datasets, limiting real-world generalizability.
• A gap exists in evaluating AI model performance on diverse, real-world patient data.

Purpose of the Study:

• To assess the generalizability of AI models trained on large research datasets for dementia prediction using real-world UK National Health Service (NHS) data.
• To evaluate the clinical utility of AI-driven dementia risk prediction in a real-world setting.

Main Methods:

• 3D T1-weighted MRI scans and electronic health records from 1140 individuals at SLaM NHS Trust were analyzed.
• 3D ResNet models, trained on NACC and ADNI datasets, were used to classify dementia in the SLaM cohort.
• AI-derived probabilities of Alzheimer's disease dementia (p_AD) were integrated with clinical covariates in a Cox proportional hazards model to predict time-to-diagnosis.

Main Results:

• AI models achieved classification accuracies ranging from 65.3% to 68.6% on the SLaM cohort.
• The p_AD significantly predicted time-to-diagnosis, with hazard ratios ranging from 3.58 to 4.67 (p<0.0017).
• A 0.1 increase in p_AD correlated with a 13.6%-16.7% increased risk of dementia diagnosis within 8 years.

Conclusions:

• AI models trained on large research datasets effectively predict dementia progression in real-world NHS patients.
• The study demonstrates the clinical utility and generalizability of AI models for dementia risk assessment in routine healthcare.
• AI-driven prediction of dementia risk shows significant potential for early intervention and management.