Predicting Continuous Cognitive Decline: The Generalizability of a Multimodal Machine Learning Approach Including...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Gerontology

Background:

• Cognitive decline in aging shows significant individual variability.
• Predicting cognitive decline continuum aids personalized interventions.
• Machine learning models integrating diverse data show promise for predicting cognitive decline.

Purpose of the Study:

• To replicate and validate machine learning models for predicting continuous cognitive decline.
• To assess the generalizability of these models across independent datasets and acquisition sites.
• To evaluate the impact of structural magnetic resonance imaging (MRI) data on model performance.

Main Methods:

• Multi-target random forest regression models were used.
• Models predicted annual decline rates of Clinical Dementia Rating Scale Sum of Boxes (CDR-SOB) and Mini-Mental State Examination (MMSE).
• Model performance was evaluated within and across datasets (ADNI, OASIS-3).

Main Results:

• Integrating structural MRI data improved model performance in the ADNI cohort, consistent with prior OASIS-3 findings.
• Models exhibited statistically significant performance degradation when tested on unseen datasets.
• Models trained on key features performed comparably to those using all features in external validation, indicating feature redundancy.

Conclusions:

• Multimodal machine learning models for continuous cognitive decline prediction demonstrate partial generalizability.
• Performance degradation across datasets highlights challenges in real-world application.
• Unimodal models using only structural MRI features did not generalize well externally.