Abstracts

Rationale: It is estimated that by 2020 almost 50% of all newly diagnosed with epilepsy will be aged 65 years or older, and with an aging population, rates of neurodegenerative disorders such Parkinson's Disease (PD) are likewise estimated to increase substantially in the next decades. Accurate diagnosis of seizures in the elderly and patients with neurodegenerative diseases is often delayed, potentially life-saving interventions may be prolonged, and evaluations are frequently extensive and costly. Over the past 15 years researchers from many disciplines have demonstrated the ability of machine learning algorithms to accurately detect subtle differences between patients to make disease classifications. Radiomics uses high-throughput extraction of quantitative features from medical images and then employs machine learning algorithms, to uncover disease characteristics that cannot be seen by the naked eye. These machine learning algorithms perform particularly well on neuroimaging applications and have been applied to detect multiple sclerosis, and seizures. Methods: 1.5T MRimaging was obtained from 162 subjects from five groups (Parkinson’s Disease [PD]; Temporal Lobe Epilepsy [TLE]; Young controls and Older controls). Using FreeSurfer, five regions of interest (ROI) were parcellated (Hippocampus, Amygdala, Cingulate gyrus, Entorhinal cortex & Thalamus). These ROIs are implicated in TLE, PD & undergo changes with aging. Using ITK Toolbox, 53 radiomic features were extracted from each ROI & analysed using a random forest machine learning algorithm. 4-fold cross validation was performed. Results: The random forest algorithm successfully classified each MRI into the correct group. The Receiver Operating Curve (ROC) for the classification of the subjects for each Fold is shown in Figure 1, & had an Area Under Curve (AUC) of 0.93. Conclusions: We developed radiomic biomarkers using readily acquired, non-invasive standard MRI scans as a way to ultimately yield relatively quick and cost-effective ways to identify at high-risk patients and improve diagnostic capabilities. Radiomic biomarkers, as used in oncology, have the potential to utilize current diagnostic standards of care. Our machine learning algorithm accurately identified TLE from PD and from older and younger controls, demonstrating that radiomic methodology can accurately classify different neurological disorders and different aged controls with high sensitivity and specificity. This proof of concept study highlights the potential for this novel methodology to be further developed into radiomic signatures for various neurological disorders that could be used as diagnostic, non-invasive and cost-effective biomarkers. Funding: No funding