Colocalization of atrophy and tau improves AI classification of Alzheimer phenotypical variants: Tau imaging. (7th December 2020)

• Record Type: Journal Article
• Authors: Damasceno, Pablo F.; La Joie, Renaud; Maia, Pedro D.; Visani, Adrienne; Iaccarino, Leonardo; Strom, Amelia; Edwards, Lauren; Tempini, Maria Luisa Gorno; Jagust, William J.; Miller, Bruce L.; Rabinovici, Gil D.; Raj, Ashish
• Is Part Of: Alzheimer's & dementia. Volume 16(2020)Supplement 1
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-12-07
• 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.046258
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: The high cognitive and anatomical heterogeneity in phenotypical variants of Alzheimer's Disease imposes a formidable challenge for biomarker‐based discriminatory models. Motivated by recent findings (La Joie, 2020) suggesting a causal relationship between regional tau accumulation and future atrophy development, we used Machine Learning techniques to test whether colocalization of these two biomarkers could improve clinical differentiation of AD phenotypical variants. Method: We investigated 85 amyloid‐positive patients with a clinical diagnosis of AD (MCI or dementia stage), including 17 patients with PCA, 16 with lvPPA, and 52 typical cases (tAD, Table 1). Patients received 3‐tesla MRI and PET with [11C]PIB and [18F]Flortaucipir. Thickness measurements were turned into age and sex adjusted z‐scores (Potvin, 2017). Standardized Uptake Value Ratio (SUVR) images were created using reference regions. Mean SUVR or cortical thickness values were extracted from 72 FreeSurfer regions (Figure 1). Random Forest (RF) and Support Vector Machines (SVM) were used for supervised classification of patients into one of the three diagnosis above. Four distinct models were compared: based on amyloid, atrophy, tau, or on the product between atrophy and tau (hence capturing the colocalization of these two biomarkers). Selected ROIs were acquired using SVM feature importance and used in a second SVM method (SVM‐ROIs). Result: Figure 2a shows the accuracy of the classification models as a function of the used input features. As a general trend: i) amyloid‐based models had the poorest classification accuracy; ii) tau‐based SVM models outperformed atrophy‐based predictions; iii) the product of atrophy and tau significantly improved accuracy; and iv) in SVM models, selection of a subset of regions consistently improved classifications over those models using all brain regions. Figure 2b shows the confusion matrix for the most accurate model (SVM on selected ROIs). Conclusion: In the cohort used in this study, the product of atrophy and tau provides a better classification feature than each feature considered in isolation. Our results suggest that the colocalization of these two biomarkers represent a valuable instrument for clinical decision‐making in phenotypical variants of AD. References: La Joie, et al. Science Translational Medicine (2020). Potvin, et al. Neuroimage (2017).