Brain age prediction in a multiethnic Asian population: A comparison of machine learning algorithms and their application for early-stage cognitive impairment diagnosis.

An ensemble model combining linear regression, lasso, and SVR achieved the best brain age prediction performance among nine models tested.

• The ensemble model was trained on 17 features: 11 subcortical volumes and 6 lobe-level cortical thickness measures.
• Overall bias-corrected MAE was 4.04 years and R2 was 0.59.
• Nine brain age prediction models were developed and compared in total.
• The training cohort consisted of 406 cognitively normal individuals aged 45-86 years from two population-based studies.

Thalamic volume, lateral ventricle volume, accumbens area volume, and gray matter volume were identified as the most important features for brain age prediction.

• Feature importance was assessed using SHapley Additive exPlanations (SHAP) analysis based on the best performing ensemble model.
• Features included both subcortical volumes and lobe-level cortical thickness measures.
• A total of 17 structural MRI features were used in the final model.

The brain age model showed significant differences in BrainAGE across cognitive groups ranging from no cognitive impairment to dementia.

• The model was applied to an independent cohort including groups with no cognitive impairment (NCI), mild and moderate cognitive impairment no dementia (CIND), and dementia.
• Differences in BrainAGE across cognitive groups were examined using an ANOVA test.
• The results support BrainAGE as a biomarker capable of detecting early cognitive decline.

The brain age models were developed and validated in a multiethnic Asian population, addressing a gap in existing literature.

• The study used structural MRI data from two population-based studies.
• The authors note that methodological variation in ML algorithms and scarce evidence from various ethnic populations limit clinical translation of brain age biomarkers.
• The cohort consisted of individuals aged 45-86 years.

Structural MRI features at subcortical and lobe-level cortical thickness granularity were sufficient to build an interpretable and robust brain age model.

• The final feature set comprised 11 subcortical volumes and 6 lobe-level cortical thickness measures.
• The model achieved a bias-corrected MAE of 4.04 years.
• The authors describe the model as 'interpretable,' facilitated by SHAP analysis.
• The approach was designed to balance prediction accuracy with clinical interpretability.