Mining healthcare data for markers of the multiple sclerosis prodrome

Background: Previous studies suggest the existence of a prodromal period in multiple sclerosis, but little is known about the phenotypic characteristics. This study aims to characterize the multiple sclerosis (MS) prodrome using data mining analytics in the healthcare setting. Methods: We identified people with MS and matched general population controls using health administrative data in two Canadian provinces (British Columbia and Saskatchewan). Using a training dataset (66.6% of British Columbia's cohort), Ll penalized logistic regression models were fitted to predict MS from physician and hospital encounters (via International Classification of Diseases ROD] codes) and prescriptions filled (as drug classes) during the five years before the MS case's first demyelinating event. Internal and external validation of identified predictors was performed using logistic regression on the remaining British Columbia (33.4%) and Saskatchewan data. Adjusted odds ratios (aORs) and Area under the Curve (AUC) metrics for the models' predictive performance were reported. Results: We identified 8,669 MS cases and 40,867 controls. Good predictive performance was observed for physician data (internal/external validation AUC = 0.81/ 0.79). Physician-generated ICD codes that were associated with MS and validated in both provinces included disorders of the central and peripheral nervous system, disorders of the eye, and cerebrovascular disease (aOR = 1.3-7.0). Overall, hospital and prescription data showed very poor and poor predictive performance (internal/external validation AUCs = 0.54/0.55 and 0.66/0.61, respectively). However, hospitalizations related to the urinary system or spinal cord diseases, or prescriptions for urinary antispasmodics or anti-vertigo preparations, were associated with 2 to 3-fold higher odds of MS (aOR = 2.3-3.3). Conclusions: Findings provide insight into the clinical characteristics of the MS prodrome. Diagnostic codes from physician encounters were capable of differentiating between MS cases and controls.