Abstract WP141: Prediction of Subclinical Language Deficit Using Machine Learning Based on Post-stroke Functional Connectivity Derived From Low Frequency Oscillations

Abstract

Introduction: Post-stroke neuropsychological evaluation can take a long time to assess impairments in subjects with no overt clinical deficits. We focus on language network based on a five-minute non-invasive resting state functional MRI (rs-fMRI) to identify subclinical language deficit (SLD) using a machine learning classifier. Such a system could supplement neuropsychological test to detect SLD from functional connectivity (FC). Predictive ability of FC derived from slow-4 (0.027-0.073 Hz) and slow-5 (0.01-0.027 Hz) bands are compared to that derived from all low frequency oscillations (LFO; 0.01-0.1 Hz). Methods: Sixty clinically non-aphasic, right-handed subjects were categorized into three age and gender matched groups based on stroke status, and normed verbal fluency score (VFS): 20 ischemic stroke subjects at higher risk of SLD (LD+; mean VFS = -1.77, mean age = 63.1 years) and 20 ischemic stroke subjects at lower risk of SLD (LD-; mean VFS = -0.05, mean age = 58.35 years), 20 healthy controls (HC; mean VFS = 0.21, mean age = 57.4 years). Anatomical T1-weighted and rs-fMRI scans were acquired within 14 days of stroke onset. A mask of 23 regions of interest (ROIs) was selected from the language network and blood-oxygen-level-dependent (BOLD) signal was extracted from each ROI corresponding to LFO, slow-4 and slow-5 bands. FC was computed by correlating time series from pairs of ROIs for all three frequency bands. A binary support vector machine (SVM) used 253 correlation indices to classify between groups: (a) LD+ vs LD-, (b) LD- vs HC (c) HC vs LD+. The classifier performance was assessed using leave-one-out cross-validation (LOOCV) with a univariate filter based feature selection. Results: LOOCV accuracies upon appropriate feature selection are reported by frequency band: Slow-4 band: 82.5% for (a), 90% for (b), 90% for (c). Slow-5 band: 85% for (a), 90 % for (b), 92.5% for (c). LFO: 82.5% for (a), 87.5% for (b), 90% for (c). Conclusion: Subjects at risk of SLD can be differentiated from healthy controls based on rs-fMRI with a high accuracy classifier, which otherwise could take longer to identify through neuropsychological assessment. Additionally, BOLD signal in the slow-5 band better characterized FC by providing a higher classification accuracy.