Biomarkers.

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Alzheimer's disease (AD) diagnosis is shifting towards physiological testing, including neuroimaging like Positron Emission Tomography (PET) and fluid biomarkers.
• Current diagnostic processes are lengthy, averaging two years, highlighting the need for more efficient methods.
• Artificial intelligence (AI) offers potential for analyzing PET scans to identify patients with AD.

Purpose of the Study:

• To investigate the potential of explainable AI in diagnosing Alzheimer's disease using amyloid PET imaging.
• To evaluate AI model performance in classifying patients based on amyloid load in specific brain regions.

Main Methods:

• Utilized medical data and PET scans from 541 patients in the Alzheimer's Disease Neuroimaging Initiative (ADNI) study.
• Trained AI models on amyloid load data from brain regions of interest to classify patients as normal or impaired.
• Tested AI models using amyloid-beta, tau, and phosphorylated tau biomarker thresholds, with machine learning algorithms applied to up to 10 brain regions.

Main Results:

• The phosphorylated tau (p-tau) model achieved the highest test accuracy at 72.8%.
• The amyloid-beta (Aβ) biomarker model demonstrated superior overall classification performance with an Area Under the Curve (AUC) of 0.80.
• Shapley importance values identified the precuneus and left lingual gyrus as significant regions for the Aβ model.

Conclusions:

• Explainable AI shows potential for AD diagnosis from amyloid PET scans, achieving moderate accuracy with relevant brain regions.
• Larger datasets and techniques like stratified cross-validation are necessary to enhance accuracy and address dataset imbalances.
• AI-driven analysis of neuroimaging biomarkers holds promise for improving the efficiency and accuracy of Alzheimer's disease diagnosis.