ARTS: A novel In-vivo classifier of arteriolosclerosis for the older adult brain

Abstract

Brain arteriolosclerosis, one of the main pathologies of cerebral small vessel disease, is common in older adults and has been linked to lower cognitive and motor function and higher odds of dementia. In spite of its frequency and associated morbidity, arteriolosclerosis can only be diagnosed at autopsy. Therefore, the purpose of this work was to develop an in-vivo classifier of arteriolosclerosis based on brain MRI. First, an ex-vivo classifier of arteriolosclerosis was developed based on features related to white matter hyperintensities, diffusion anisotropy and demographics by applying machine learning to ex-vivo MRI and pathology data from 119 participants of the Rush Memory and Aging Project (MAP) and Religious Orders Study (ROS), two longitudinal cohort studies of aging that recruit non-demented older adults. The ex-vivo classifier showed good performance in predicting the presence of arteriolosclerosis, with an average area under the receiver operating characteristic curve AUC = 0.78. The ex-vivo classifier was then translated to in-vivo based on available in-vivo and ex-vivo MRI data on the same participants. The in-vivo classifier was named ARTS (short for ARTerioloSclerosis), is fully automated, and provides a score linked to the likelihood a person suffers from arteriolosclerosis. The performance of ARTS in predicting the presence of arteriolosclerosis in-vivo was tested in a separate, 91% dementia-free group of 79 MAP/ROS participants and exhibited an AUC = 0.79 in persons with antemortem intervals shorter than 2.4 years. This level of performance in mostly non-demented older adults is notable considering that arteriolosclerosis can only be diagnosed at autopsy. The scan-rescan reproducibility of the ARTS score was excellent, with an intraclass correlation of 0.99, suggesting that application of ARTS in longitudinal studies may show high sensitivity in detecting small changes. Finally, higher ARTS scores in non-demented older adults were associated with greater decline in cognition two years after baseline MRI, especially in perceptual speed which has been linked to arteriolosclerosis and small vessel disease. This finding was shown in a separate group of 369 non-demented MAP/ROS participants and was validated in 72 non-demented Black participants of the Minority Aging Research Study (MARS) and also in 244 non-demented participants of the Alzheimer's Disease Neuroimaging Initiative 2 and 3. The results of this work suggest that ARTS may have broad implications in the advancement of diagnosis, prevention and treatment of arteriolosclerosis. ARTS is publicly available at https://www.nitrc.org/projects/arts/.

Keywords: Arteriolosclerosis; Brain; Cognition; MRI; Machine learning; Pathology.

Fig. 1

Venn diagrams describing the available data for the groups of MAP/ROS participants included in training, translating, and testing the classifier. (A) The training group was used for training and testing the ex-vivo classifier. (B) The translation group was used for translating the ex-vivo classifier to in-vivo. (C) The testing group was used for testing the performance of the in-vivo classifier, i.e. ARTS, in predicting the presence of arteriolosclerosis based on in-vivo brain MRI. (D) The cognitive decline testing group (or CDT group), excluding MAP/ROS participants with dementia, was used for testing the hypothesis that higher ARTS score in non-demented older adults is associated with a greater decline in cognition two years after baseline in-vivo MRI. There was no overlap between the training group and the testing or CDT groups.

Fig. 2

Median FA values were extracted from the four numbered regions and were used as features in the classifier. The regions are displayed in the space of the IIT Human Brain Atlas v.5. The mean FA template and white matter skeleton of the atlas are shown in grayscale and green color, respectively. Note that the regions are only defined on the skeleton and have been dilated in this figure to enhance visualization. According to the Regionconnect software (Qi and Arfanakis, 2021) the top five most likely connections through each of the four regions are: (Region 1) lateral occipital to middle temporal (13.5%), fusiform to lateral occipital (13.3%), inferior temporal to lateral occipital (10.1%), middle temporal to pericalcarine (9%), fusiform to lingual (6.2%); (Region 2) lateral occipital to superior temporal (9.9%), middle temporal to superior temporal (6.6%), superior parietal to superior temporal (6.2%), inferior parietal to superior temporal (6%), banks of the superior temporal sulcus to superior temporal (5.5%); (Region 3) precentral to thalamus (17.8%), paracentral to thalamus (10.7%), superior frontal to superior parietal (9.6%), precentral to caudate (6.1%), postcentral to precentral (4.3%); (Region 4) precuneus to superior parietal (12%), inferior parietal to superior parietal (6.7%), left superior parietal to right superior parietal (6.5%), isthmus cingulate to superior parietal (6.5%), cuneus to superior parietal (6.1%). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

ROC curves for ex-vivo classification of arteriolosclerosis using classifiers based on different combinations of features. The black line and gray cloud represent the mean and standard deviation of the best performing classifier that combines WMH, regional FA, and demographic features.

Fig. 4

AUC values for in-vivo prediction of arteriolosclerosis in participants of the testing group with antemortem intervals (AMI) less than the threshold value shown on the horizontal axis. The number of participants with AMI less than the threshold is indicated above the curve.

Fig. 5

In-vivo ARTS confidence scores at different levels of arteriolosclerosis severity assessed pathologically for participants of the testing group with AMI ≤ 2.4 years. Stages 2 and 3 were pooled because stage 3 included only one participant. Boxes represent ARTS scores from 25th-75th percentiles, the horizontal lines inside the boxes represent median scores, and the whiskers illustrate the range of variation.