Machine‐learning based MRI neuro‐anatomical signatures associated with cardiovascular and metabolic risk factors. (20th December 2022)

• Record Type: Journal Article
• Authors: Govindarajan, Sindhuja Tirumalai; Mamourian, Elizabeth; Erus, Guray; Abdulkadir, Ahmed; Melhem, Randa; Doshi, Jimit; Pomponio, Raymond; Tosun, Duygu; An, Yang; Sotiras, Aristeidis; Marcus, Daniel S.; LaMontagne, Pamela J.; Espeland, Mark A.; Masters, Colin L.; Maruff, Paul; Launer, Lenore J.; Fripp, Jurgen; Johnson, Sterling C.; Morris, John C.; Albert, Marilyn S.; Bryan, R. Nick; Habes, Mohamad; Shou, Haochang; Wolk, David A.; Nasrallah, Ilya M.; Davatzikos, Christos
• Is Part Of: Alzheimer's & dementia. Volume 18(2022)Supplement 1
• 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.067709
• Languages: English
• ISSNs: 1552-5260

Abstract

Abstract: Background: Lower gray matter (GM) and higher white matter (WM) volumes have been observed in cognitively unimpaired people with cardiovascular and metabolic risk factors (CVMs) such as hypertension, hyperlipidemia, diabetes, obesity, and smoking [Erus 2015, Habes 2016]. However, these group‐level observations do not explain the heterogeneous accumulated effects of CVMs on individual neuroanatomy. Multivariate CVM‐related changes characterized at the individual level using machine learning (ML)‐derived summary indices may better capture their differential roles in brain aging and dementia. In this work, we derive ML‐based indices that summarize brain changes from MRI related to CVMs. Method: Cognitively normal volunteers that were pooled and harmonized for the iSTAGING cohort were included in this study (N=24, 902 from 8 independent studies, 54.5% female, average age=62.4, age range 45‐75 years). Linear support vector classifiers were trained to discriminate between people with a specific CVM from those without the CVM using nested cross‐validation for hyperparameter tuning and model selection. Input features (n=150) consisted of age, sex, intracranial volume, harmonized regional GM and WM volumes, and lobar volumes of WM hyperintensities. The resulting distance of the projection onto the vector perpendicular to the decision plane yields a continuous scalar representing 'spatial pattern of abnormality for recognition' (SPARE) of each CVM. SPARE‐CVMs were computed for each participant and correlated with previously established imaging markers for brain age [Habes 2021] and Alzheimer's disease (SPARE‐AD)[Davatzikos 2009]. Effect sizes (Cohen's d) were calculated to assess the utility of SPAREs with higher values indicating greater separability between the CVM+ and CVM‐ groups. Result: Figure 1 shows the distribution, effect sizes and associated brain regions for each SPARE‐CVM. SPARE‐diabetes and SPARE‐hypertension yielded the largest effect sizes. SPAREs for cooccurring CVMs showed higher correlation (Figure 2) and greater corresponding effect sizes (Figure 3). For example, hyperlipidemia was separable by SPARE‐hypertension, but not other SPAREs. By contrast, brain age gap (SPARE‐BA) and SPARE‐AD were unrelated to CVMs. Conclusion: We derived sensitive and partially non‐overlapping quantitative indices that characterize the degree of CVM‐related neuroanatomical differences in cognitively unimpaired participants. SPARE‐CVMs were correlated for cooccurring CVMs but dissociated from MR imaging biomarkers for typical aging and Alzheimer's disease.