A multimodal machine learning model for predicting dementia conversion in Alzheimer's disease.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Imaging

Background:

• Alzheimer's disease (AD) represents a significant portion of dementia cases, with Mild Cognitive Impairment (MCI) serving as a critical intermediate stage.
• Approximately 10-15% of individuals with MCI annually progress to AD, highlighting the need for accurate predictive tools.

Purpose of the Study:

• To identify the most robust machine learning model and imaging modality combination for predicting conversion from MCI to AD.
• To leverage multi-modal imaging features and demographic data for enhanced predictive accuracy.

Main Methods:

• Six machine learning models were evaluated using data from 196 subjects from multiple hospitals and the Alzheimer's Disease Neuroimaging Initiative.
• Features analyzed included regional volumes, white matter hyperintensity, Standardized Uptake Value Ratio (SUVR) from T1, T2-FLAIR MRIs, and amyloid PET (αPET).
• Hippocampal occupancy scores (HOC) and Fazekas scales were also incorporated.

Main Results:

• The Gradient Boosting Machine (GBM) model demonstrated the highest stability among the tested models.
• Model performance for predicting AD conversion was enhanced when T2-FLAIR MRI features were excluded from the combination.
• The study successfully predicted the probability of AD conversion in MCI patients over a four-year follow-up.

Conclusions:

• The GBM model, utilizing specific multi-modal imaging data, offers a robust approach for predicting AD progression in MCI patients.
• Excluding T2-FLAIR MRI data improved predictive performance, suggesting a refined feature set for future models.
• These findings can aid clinicians in the early diagnosis and treatment planning for individuals at risk of developing Alzheimer's disease.