Alzheimer's disease (AD) is a degenerative neurological disorder marked by cognitive decline and functional disability. Despite the extensive use of magnetic resonance imaging (MRI) in machine learning (ML)-based AD studies, the relative and combined contributions of MRI-derived morphometric (MO), microstructural (MS), and graph-theoretical (GT) features are still not well explored in a unified, comparative framework. It remains unclear whether adding multimodal MRI-derived features consistently improves the predictive performance of ML-based approaches for AD diagnosis and cognitive decline. Addressing this gap, this study systematically analyzed the individual (MO, MS, GT) and combined (MO+MS, MO+GT, MS+GT, MO+MS+GT) utility of MRI-based feature sets. We developed an ensemble-based ML framework with a nested cross-validation module for two key tasks: (i) Alzheimer's disease cognitive stage classification (DSC) and (ii) longitudinal cognitive decline prediction (LCDP) in terms of mini-mental state examination (MMSE) score. In this study, we conducted feature ablation and statistical analysis to evaluate performance improvements resulting from the incremental addition of feature sets. The results of the study indicated that the proposed ensemble-based ML approach achieved the best predictive performance (balanced accuracy [BACC]: 0.898 ± 0.051) using a combination of MO and MS feature sets for cognitively normal (CN) vs. AD dementia (CN–ADD). In contrast, the best results for mild cognitive impairment (MCI) vs. ADD (MCI–ADD) and CN–MCI were achieved using the MO feature set alone, with BACC of 0.769 ± 0.116 and 0.652 ± 0.044, respectively. Likewise, for the LCDP task, the MO-based ensemble learner achieved an R2 of 0.212 ± 0.177. These results demonstrate that MO features capture the most robust disease-related information, while multimodal integration offers task-specific and limited benefits. In addition, these findings demonstrate the potential of integrated MRI-derived features in ML frameworks for enhancing ADD diagnosis and cognitive decline prediction and underscore the importance of feature selection based on task complexity.
A systematic study on the integration of MRI connectivity metrics for Alzheimer's diagnosis, staging, and cognitive decline prediction
Kreshpa, Wendy;Rosso, Nicola;Piana, Michele;Roccatagliata, Luca;Cirone, Alessio;Luigi, Lorenzini;Campi, Cristina;Pardini, Matteo;Garbarino, Sara
2026-01-01
Abstract
Alzheimer's disease (AD) is a degenerative neurological disorder marked by cognitive decline and functional disability. Despite the extensive use of magnetic resonance imaging (MRI) in machine learning (ML)-based AD studies, the relative and combined contributions of MRI-derived morphometric (MO), microstructural (MS), and graph-theoretical (GT) features are still not well explored in a unified, comparative framework. It remains unclear whether adding multimodal MRI-derived features consistently improves the predictive performance of ML-based approaches for AD diagnosis and cognitive decline. Addressing this gap, this study systematically analyzed the individual (MO, MS, GT) and combined (MO+MS, MO+GT, MS+GT, MO+MS+GT) utility of MRI-based feature sets. We developed an ensemble-based ML framework with a nested cross-validation module for two key tasks: (i) Alzheimer's disease cognitive stage classification (DSC) and (ii) longitudinal cognitive decline prediction (LCDP) in terms of mini-mental state examination (MMSE) score. In this study, we conducted feature ablation and statistical analysis to evaluate performance improvements resulting from the incremental addition of feature sets. The results of the study indicated that the proposed ensemble-based ML approach achieved the best predictive performance (balanced accuracy [BACC]: 0.898 ± 0.051) using a combination of MO and MS feature sets for cognitively normal (CN) vs. AD dementia (CN–ADD). In contrast, the best results for mild cognitive impairment (MCI) vs. ADD (MCI–ADD) and CN–MCI were achieved using the MO feature set alone, with BACC of 0.769 ± 0.116 and 0.652 ± 0.044, respectively. Likewise, for the LCDP task, the MO-based ensemble learner achieved an R2 of 0.212 ± 0.177. These results demonstrate that MO features capture the most robust disease-related information, while multimodal integration offers task-specific and limited benefits. In addition, these findings demonstrate the potential of integrated MRI-derived features in ML frameworks for enhancing ADD diagnosis and cognitive decline prediction and underscore the importance of feature selection based on task complexity.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
