Explainable Machine Learning with Pairwise Interactions for Predicting Conversion from Mild Cognitive Impairment to...

Area of Science:

• Neuroscience
• Medical Informatics
• Machine Learning

Background:

• Predicting cognitive decline in mild cognitive impairment (MCI) is crucial for timely intervention in potential Alzheimer's Disease (AD) progression.
• Explainable machine learning (XAI) models offer a promising approach to enhance predictive accuracy while maintaining interpretability for MCI-to-AD conversion.
• This study leverages the Explainable Boosting Machine (EBM) model with multimodal features to predict MCI to AD conversion with interpretability.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for predicting Alzheimer's Disease (AD) conversion from mild cognitive impairment (MCI).
• To identify key multimodal features (MRI, cognitive, clinical) that contribute to the prediction of MCI to AD conversion over various follow-up periods.
• To compare the predictive performance of different feature combinations using the Explainable Boosting Machine (EBM) model.

Main Methods:

• A retrospective case-control study utilizing data from 1042 MCI patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (2006-2022).
• The Explainable Boosting Machine (EBM) model was employed to predict AD conversion from MCI using combinations of MRI biomarkers, cognitive scores, demographics, and clinical features.
• Model performance was evaluated based on accuracy, sensitivity, specificity, and AUC-ROC across different follow-up periods, with consideration for interaction effects and feature interpretability.

Main Results:

• The EBM model achieved 85% prediction accuracy (AUC = 0.92) for five-year MCI to AD conversion using cognitive scores and MRI markers.
• Feature importance varied by follow-up period: MRI markers were critical in early stages, while cognitive scores became more important in the mid-term.
• Specific predictors identified include right inferior temporal volume (1-year), left inferior temporal thickness (2-year), FAQ scores (3-year), APOE4 status (4-year), and right entorhinal volume (5-year).

Conclusions:

• Explainable Boosting Machines (EBMs) with multimodal features provide a robust and interpretable method for predicting Alzheimer's Disease (AD) conversion from mild cognitive impairment (MCI).
• The study identified significant multimodal features that are crucial for predicting MCI to AD conversion at different time points.
• The developed prediction tool holds potential for improving clinical management strategies for patients at risk of AD.