Wednesday, September 26, 2018 7:35 AM–9:00 AM ePosters: P42. Preliminary validation of a machine-learning MRI-based algorithm to identify patients with functional impairment associated with lumbar stenosis

Abstract

BACKGROUND CONTEXT The anatomic changes characterized by magnetic resonance imaging (MRI) in patients with lumbar spinal stenosis weakly correlate with pain intensity and functional limitation in patients who endorse the classic evoked symptom pattern of neurogenic claudication (NC). PURPOSE To identify MRI features that correlate with the evoked symptom pattern of neurogenic claudication using a novel machine learning algorithm. STUDY DESIGN/SETTING Subjects with chronic low back pain and a lumbar MRI study were recruited from a single academic neurosurgery department and evaluated using a standardized protocol of validated, disease specific questionnaires and a treadmill-based functional assessment. PATIENT SAMPLE A total of 159 subjects with lumbar spinal stenosis were categorized into NC and nonNC cohorts based on clinical assessment. An independent radiologist blinded to clinical cohort analyzed the MRI data. Subjects with nonstandard MRI imaging data or studies compromised by artifact (eg lack of fat-suppression or presence of obscuring artifact) were excluded. 61 subjects (28 females) with 33 classified as NC+ were included. The mean age was 68.4±11. OUTCOME MEASURES Classification of NC based on imaging variables. METHODS The T2 weighted sagittal MRI FS images were segmented into posterior fat, subcutaneous fat, muscle, vertebral bodies, spinal canal and paravertebral space, then further segmented into seven sub-regions corresponding to levels T12 through S1. Image data was quantified for topography of each ROI or sub-ROI (area, compactness, and mass eccentricity); MRI signal (mean, standard deviation, skewness, kurtosis, entropy, and energy) and texture (gray-level co-occurrence matrix) (angular second moment, contrast, mutual information, correlation, dissimilarity, entropy, homogeneity, marginal entropy, and variance). Together the regions and measurements generated 1,512 features. Left and right mean and absolute difference were calculated to generate features representing asymmetry. All measurements were adjusted for gender and age differences; calibrated for gain differences using the characteristics of the background and normalized using the inverse rank transform. The normalized features were used to construct a diagnostic model based on logistic models. The model was constructed using a feature selection algorithm under a 10-fold cross-validation scheme repeated 5times. RESULTS Differences texture in the vertebral body and the spinal canal were strongly associated with the clinical diagnosis. The model validation predicts that MRI-based logistic models can separate NC with a sensitivity of 0.75 (95% CI 0.55–0.89) with a specificity of 0.97 (95% CI 0.84–1.00) and a diagnostic odds ratio of 96. CONCLUSIONS A machine learning model of MRI features can determine the likelihood of the clinical presentation of NC. Differences in vertebral body signal and the spinal canal show the strongest association to the clinical diagnosis. Further validation of the model may help facilitate pain management strategies.