Abstract
Early detection of pathological progression in Mild Cognitive Impairment (MCI) is a key challenge in Alzheimer's disease (AD) research. In this work, we focus on longitudinal modelling of disease progression using three publicly available datasets: ADNI, AIBL and OASIS-2. Cortical and subcortical features are extracted from T1-weighted MRI using FreeSurfer, and mixed-effects models are employed to characterise region-wise temporal trajectories. We propose an ensemble of linear models operating on trajectory-derived features to classify MCI-to-AD progression. In particular, we introduce Trajectory-based Apparent Brain Features (t-ABF), a set of interpretable biomarkers obtained by combining regression-based longitudinal modelling with classification. A longitudinal feature selection along with biased forward feature selection is used to identify the most informative region-based trajectories. Experimental results show that the trajectory-derived ABF model predicts pathological progression and capturing progression acceleration patterns. The proposed approach improves biomarker detection while preserving interpretability, providing a clinically meaningful alternative to less transparent deep learning methods.