An autopsy‐validated, easily deployable MRI predictor of Alzheimer's disease tau pathology. (20th December 2022)

• Record Type: Journal Article
• Authors: Vogel, Jacob W; Strandberg, Olof; Gaiteri, Chris; Cieslak, Matthew; Covitz, Sydney; Wolk, David A.; Davatzikos, Christos; Hansson, Oskar; Satterthwaite, Theodore
• Is Part Of: Alzheimer's & dementia. Volume 18(2022)Supplement 6
• Page Start: n/a
• Page End: n/a
• Publication Date: 2022-12-20
• Subjects: Alzheimer's disease -- Periodicals; Alzheimer Disease -- Periodicals; Dementia -- Periodicals; Démence; Maladie d'Alzheimer; Périodique électronique (Descripteur de forme); Ressource Internet (Descripteur de forme); 616.83
• DOI: 10.1002/alz.063336
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Tau pathology is a hallmark of Alzheimer's disease, but measuring this pathology in vivo with PET is expensive. MRI represents a low‐cost alternative for measuring in vivo disease progression, but is non‐specific. We present an MR‐based marker of tau‐specific pathology using machine learning (ML). Method: 1851 pairs of RO948 tau‐PET and T1‐weighted MRI scans were aggregated from 1262 BioFINDER II study participants. The Braak I‐IV ROI from Cho et al., 2016 was used to represent tau‐pathology. A tau‐positive threshold (SUVR=1.41) was defined using Gaussian mixture modeling, and Braak I‐IV SUVR was z‐scored using the mean and SD of the tau‐negative group. A grid search of 80 ML pipelines with 100 rounds of nested 5‐fold cross‐validation was performed while varying estimator, dimension reduction strategy, preregression of intracranial volume (ICV), and one vs multiple scans‐per‐subject. For all models, 294 features from Freesurfer were used. The best‐performing model was then applied to 1225 participants from the ROSMAP cohort, yielding MR‐predicted tau. A subset (n=159) of these participants had postmortem AT8‐immunostained tau pathology, Braak staging, TDP‐43 and lewy body pathology. Result: The best pipeline involved random forest regression with FDR‐based feature selection (RMSE[95%CI]=3.67[3.6‐3.8]; R2[95%CI]=0.42[0.40‐0.45]; Figure1; Figure2A ). (Lack of) ICV pre‐regression had the biggest positive impact across pipelines. This model significantly predicted zTau in both tau‐postive and tau‐negative individuals (Figure2B), and showed fair weighted precision (0.84) and recall (0.80) in tau‐positivity classification (Figure2C) . Temporoparietal and medial temporal regions contributed most to predictions (Figure2D) . When applied to ROSMAP, antemoretem MR‐predicted tau predicted postmortem tau pathology (p=0.001) and Braak stage (p=0.005; Figure3A ) but not other pathologies (Figure3B), predicted clinical diagnosis and progression (Figure3C ), and outperformed a common neurodegeneration marker (Figure3D‐F) . Conclusion: We create an easily deployable tau‐specific MRI biomarker, which was validated using postmortem tau pathology in a separate dataset. MR‐predicted tau provides a more specific non‐invasive biomarker for researchers without access to PET.