Three‐stage classification of Alzheimer's disease in MRI in clinical records of the Mass General Brigham Health Care System

Abstract

AbstractBackgroundDiagnostic deep learning has frequently been applied to diagnosis of Alzheimer's and Mild Cognitive Impairment in brain MRIs collected in a controlled research setting, but due to the presence of confounding factors, similar methods have rarely been applied post‐hoc to clinical MRI. Data matching, or the matching of datasets across different classes with respect to a number of confounds, is a method of finding subsets of an overall dataset, that can be used to mitigate the confounding factor problem.Method467,464 multimodal clinical MRIs of the brain from 37,311 patients in the Mass General Brigham (MGB) Healthcare System were either excluded (based on presence of neoplasms or head trauma) or labeled as AD, MCI, or control, based on a previous prescription of rivastigmine, galantamine, or donepezil (MCI), or memantine (AD), or the lack of a history of central nervous system medications. A data matching algorithm was applied to 18 technical and demographic confounding factors to isolate an unconfounded matched dataset set to train an ensemble of five 3D ResNet‐50 deep learning models. Across the ensemble, the test set totaled 287,367 files. See Figure 1.ResultIn distinguishing controls from AD/MCI on a per‐patient basis, an AUROC of 0.82 was achieved, though with lower performance in distinguishing between AD and MCI. Variations in performance were seen with MRI modality, with MPRAGE having the highest performance overall (0.859 AUROC). See Figure 2.ConclusionWhen carefully accounting for confounds, the use of deep learning models is feasible in the analysis of highly‐confounded clinical MRI.