AI Sifts Through Brain Scans to Detect Alzheimer's Disease with Incredibly High Accuracy

Biomarkers for Alzheimer's Disease (AD) are quite effective for diagnosis and treatment plans in clinical settings. Patients may experience a suite of cognitive tests, Magnetic Resonance (MR) imaging, and have fluid samples drawn (blood, cerebral spinal fluid) to determine levels of biomarkers that help to categorize neurobiological status. As technology advances, so does the precision with which early diagnosis and prediction trajectories can be made.

"Pure AD" cases are actually rare. With age, additional neuropathological changes tend to materialize and may contribute to the pathology of several neurodegenerative diseases, including AD, which may cause mis-diagnosis or may lower efficacy of particular clinical trials that are targeted to a specific condition. Common culprits of non-AD neuropathological changes include aggregates of α-synuclein seen in Lewy body disease (LBD), transactive response DNA-binding protein (TDP-43) inclusions, and amyloid-β (Aβ) in the form of cerebral amyloid angiopathy (CAA). The combination of these co-pathologies, along with known AD biomarkers Aβ and tau, all contribute to neurodegeneration, ultimately leading to cognitive and clinical decline. Levels of each vary among individuals who may qualify for a clinical trial, but making predictions for longitudinal outcomes is difficult to determine.

On November 10th, in the News section of Nature, research by Duygu Tosun and colleagues was highlighted because of their development of a promising AI tool that can be used to impute the presence of these co-pathologies in patients. They identified MR imaging signatures of each co-pathology and used these to determine to what extent each contributes to variance in cognitive decline.

Antemortem MRI volumetric signatures associated with the presence of (A) TDP-43, (B) LBD, and (C) CAA pathology at autopsy. A detailed list of anatomical regions is provided in the Supplementary Material. Color map illustrates the coefficient estimates of the statistically significant (based on 95% confidence intervals) regional volumetric variables in predicting the presence of the corresponding comorbid non-AD pathology.

This tool was applied to three large AD study cohorts, and showed accuracies of 81% for LBD, 84% for TDP-43, and 81-93% for CAA. Using information from these signatures in conjunction with demographic, genetic, and AD variables, clinical trial models can be more precision-based and trial sizes can be reduced while maintaining power of model prediction.

The research article was published in Alzheimer's & Dementia.