Objective: With this work, we aim to automatically classify patients with Parkinson’s disease (PD) and controls, only on the basis of neuroimaging parameters.

Background: Dysfunctions in neural circuits, involving both cortical and subcortical regions, have been found in Parkinson’s disease (PD) thanks to the use of neuroimaging techniques. In particular, abnormalities in temporal, parietal and frontal areas have been identified, also related with disease progression [1]. Machine learning (ML) methods identify patterns in data, without the use of any a-priori knowledge [2].

Method: We used magnetic resonance imaging (MRI) structural T1 data from the Parkinson’s Progression Markers Initiative (PPMI) public dataset. We collected data from 106 patients with PD (males, xy ± y years old) with no cognitive deficit, 73 patients with PD with mild cognitive impairment (MCI), and 106 healthy controls (males, xy ± y years old). Cortical thickness (CT) and subcortical volumes were obtained using FreeSurfer [3]. The SVM with RBF kernel was trained using cross-validation with a leave-one-out procedure. SVM was tested for 20 values of the parameters C, exponentially spaced from 2^–7 to 2¹². Separate SVMs were trained for 35 regions of interest (ROIs) CT and 20 subcortical ROIs volume.

Results: Highest accuracy was reached between PD and HC for right postcentral gyrus CT (67%), followed by left postcentral gyrus (65%) and right caudal middle frontal CT (63%) For volumes, the best accuracy was reached by left amygdala (63%). It was not possible to discriminate between PD and MCI using CT, while HC vs MCI gave the best accuracy for left CT in anterior cingulate gyrus, frontal gyrus and superior parietal regions (all 64%).

Conclusion: Our results are in line with findings in literature [4,5]. Most affected regions are automatically identified to be in the motor cortex, temporal and parietal areas. This confirms the novel theory that PD is also a cortical disease, with a distinctive pattern. In fact, it is possible to distinguish patients from controls only on the basis of CT in these regions. Our results confirm, without the bias given by a-priori knowledge-driven analyses, that PD causes a disruption in cortical networks.

References: [1] doi: 10.1016/j.parkreldis.2019.05.012 [2] doi: 10.1016/j.jad.2019.06.019 [3] https://surfer.nmr.mgh.harvard.edu/ [4] doi: 10.1002/mds.26590 [5] 10.3389/fnhum.2018.00469

« Back to MDS Virtual Congress 2020

MDS Abstracts - https://www.mdsabstracts.org/abstract/cortical-thickness-and-volume-to-automatically-identify-parkinsons-patients-using-svm/