Machine-learning models for Alzheimer's disease diagnosis using neuroimaging data: survey, reproducibility, and...

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Clinical Alzheimer's disease (AD) diagnosis often occurs late, limiting treatment efficacy.
• Existing methods fail to differentiate stable mild cognitive impairment (sMCI) from progressive mild cognitive impairment (pMCI).
• Early AD diagnosis offers significant benefits across diverse populations, yet clinical translation of advanced techniques remains limited.

Purpose of the Study:

• To survey preprocessing, data management, and machine learning (ML) techniques for AD diagnosis using neuroimaging.
• To evaluate the reproducibility and generalizability of open-source ML models for AD detection.
• To identify challenges hindering the clinical application of ML in AD diagnosis and biomarker discovery.

Main Methods:

• Systematic review of ML and deep learning (DL) methods applied to structural MRI (sMRI), fMRI, and PET data for AD diagnosis.
• Empirical assessment of open-source ML model performance using varying data cohorts.
• Comparative analysis of model generalizability across different datasets.

Main Results:

• Existing ML models demonstrate limited generalizability when applied to different neuroimaging data cohorts.
• Reproducibility and cross-cohort performance vary significantly among evaluated open-source models.
• Preprocessing and data management strategies impact model performance and generalizability.

Conclusions:

• Current ML models for AD diagnosis require further development to improve generalizability and clinical utility.
• Addressing challenges in data heterogeneity and model validation is critical for reliable AD biomarker discovery.
• Future research should focus on robust, generalizable ML solutions for early and accurate AD detection.