Biomarkers.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Dementia is linked to 14 major risk factors, many affecting brain structure.
• Brain imaging measures can potentially identify dementia risk.
• Medical AI models often lack generalizability due to bias.

Purpose of the Study:

• To explore dementia risk factors using brain imaging and machine learning.
• To evaluate the performance of Automated Machine Learning (AutoML) against benchmark models.
• To assess model generalizability using an ethnically diverse dataset.

Main Methods:

• Utilized the Health and Aging Brain Study-Health Disparities (HABS-HD) dataset with over 3,000 participants.
• Investigated six risk factors (ApoE ε4, hypertension, obesity, smoking, diabetes, alcohol) and MMSE scores.
• Employed AutoML to automate pipeline selection and compared it with nine benchmark models and an ensemble using mean cortical thickness features.

Main Results:

• AutoML demonstrated superior performance in 5 out of 7 prediction tasks.
• AutoML achieved 57.9% balanced accuracy in Mini-Mental State Examination (MMSE) classification, outperforming benchmark models (56.7%) and ensembles (55.5%).
• Unique pipelines were selected by AutoML for 5 out of 7 tasks, indicating adaptability.

Conclusions:

• AutoML shows significant potential in Alzheimer's disease research by adapting to specific prediction tasks.
• The study highlights the lack of a universal model for predicting dementia risk factors from brain structure.
• AutoML reduces experimenter bias and improves model generalizability in medical AI applications.