Automated Imaging Differentiation for Dementia: Including Alzheimer Disease Dementia and Dementia with Lewy Bodies.
Authors
Affiliations (6)
Affiliations (6)
- J. Crayton Pruitt Family Department of Biomedical Engineering, University of Florida.
- Department of Neurology, Mayo Clinic, Scottsdale, AZ.
- Department of Applied Physiology and Kinesiology, University of Florida.
- Digital Worlds Institute, University of Florida.
- Department of Neurology, University of Pennsylvania.
- Department of Radiology, Mayo Clinic, Rochester, MN.
Abstract
Differentiation of Alzheimer's disease dementia (ADD) and dementia with Lewy bodies (DLB) remains a challenge. Free-water imaging has been investigated in neurodegenerative diseases and was found to be associated with neurodegeneration and neuroinflammation. This retrospective cohort study tested whether Automated Imaging Differentiation for Dementia (AIDD), combining diffusion free-water imaging (FWI) and support vector machine, predicts ADD vs DLB with high accuracy. Diffusion MRI data was rendered from ADNI, NACC, and PDBP. Free-water and free-water corrected fractional anisotropy were calculated for each participant using a bi-tensor model. Diffusion metrics were randomly assigned to training and testing sets. The primary outcome was the area under the curve (AUC) in the test set. AIDD was paired with antemortem MRI to predict postmortem pathology. A total of 519 diffusion scans were processed with 258 ADD (mean age 73.7 (8.8), 50% male), 129 DLB (mean age 69.3, 88% male), and 132 controls (mean age 73.6 (6.8), 40% male). The machine learning sample included 387 scans,129 ADD with a mean age of 72.8 (8.7), 52.7% male; 129 DLB with a mean age of 69.3 (8.1), 87.6% male; and 129 controls with a mean age of 73.7 (6.8), 39.5% male). AIDD showed high training AUC for ADD vs DLB = 0.995 (95% CI, 0.985-1.000), ADD vs controls = 0.992 (95% CI, 0.982-1.000), DLB vs controls = 0.991 (95% CI, 0.983-0.999), and controls vs ADD/DLB = 0.990 (95% CI, 0.979-1.000). The testing AUCs were similar: ADD vs DLB = 0.995, ADD vs controls = 0.958, DLB vs controls = 0.939, controls vs ADD/DLB = 0.903. AIDD predictions were confirmed pathologically in a cohort of 13 patients. This study demonstrates that machine learning in combination with free-water imaging can differentiate ADD, DLB, and normal aging with high clinical and pathological accuracy. Advancement in early detection of dementia can lead to more appropriate treatment plans, especially for DLB, and improved disease stratification that have hindered drug development trials. This study provides Class II evidence that Automated Imaging Differentiation for Dementia, combining diffusion free-water imaging and machine learning accurately distinguishes Alzheimer's Disease from Dementia with Lewy bodies.