Comparison of Statistical Methods for Brain Age Prediction Using Neuroimaging Data.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomarker Discovery

Background:

• The aging global population necessitates reliable brain aging biomarkers.
• Brain age estimation using machine learning on neuroimaging data is a promising approach for assessing brain health and informing interventions.

Purpose of the Study:

• To evaluate and compare the performance of various kernel-based and ensemble machine learning models for brain age prediction.
• To assess the generalizability of these models on an independent dataset and investigate the impact of data harmonization.

Main Methods:

• Trained 25 kernel-based (SVM, RVM, GPR) and ensemble-based (Random Forest, XGBoost) models on T1-weighted MRI anatomical features from the Cam-CAN dataset.
• Evaluated model performance using mean absolute error (MAE) and prediction R² on the internal dataset and the external HCP-Aging dataset.
• Applied the ComBat pipeline for data harmonization on the HCP-Aging dataset to assess its effect on model performance.

Main Results:

• Non-linear models, particularly SVM with an RBF kernel, outperformed linear models on the internal dataset (MAE: 5.89 years, Prediction R²: 0.84).
• Extreme Gradient Boosting (XGB) showed robustness on non-harmonized external data (MAE: 7.45 years, Prediction R²: 0.64).
• After data harmonization, SVM with an RBF kernel achieved the highest accuracy on the external dataset (MAE: 7.05 years, Prediction R²: 0.63), highlighting the importance of harmonization for kernel-based models.

Conclusions:

• Combining non-linear machine learning models with data harmonization techniques significantly enhances the accuracy and generalizability of brain age prediction.
• These improved brain age prediction tools offer more reliable assessments of neurological health across diverse datasets, crucial for addressing age-related neurological diseases.