Wednesday, September 26, 2018 10:35 AM – 12:00 PM Cervical Spine Trauma: 22. A simplified clinical prediction rule for prognosticating independent walking after spinal cord injury: a prospective study from a Canadian multicenter spinal cord injury registry

Abstract

BACKGROUND CONTEXT Models for predicting recovery in patients affected with traumatic spinal cord injury (tSCI) have been developed in recent years to optimize patient-centered care. Several databases and analytical models have been validated in their ability to predict tSCI recovery, yet recent findings have queried the accuracy of these models in patients whose prognoses are least predictable, ie, AIS B and C patients. PURPOSE To examine the predictive accuracy of van Middendorp's five-variable and Hicks’ three-variable LR models in the subset of AIS B and C patients. Secondly, to assess the predictive accuracy of a novel, seven-variable logistic regression model in this cohort. Thirdly, to investigate the utility of advanced computing models based on artificial intelligence in the context of AIS B and C prognostication. STUDY DESIGN/SETTING Prospective cohort study using the Rick Hansen Spinal Cord Injury Registry, a pan-Canadian prospective observational registry. PATIENT SAMPLE Patients who sustained tSCI between 2004 and 2016. OUTCOME MEASURES Neurological severity (AIS) at time of admission and level of injury was assessed within 15 days of injury according to the International Standards for Neurological Classification of Spinal Cord Injury (ISNCSCI). Independent walking ability was assessed by the FIM at ≥12 months from time of injury, by which time neurologic and functional outcomes are known to plateau (16-20). METHODS Three logistic regression models and two artificial intelligence models were evaluated for their ability to predict walking ability one-year postinjury. Comparison of prognostic performance between models was calculated by area under the receiver operating characteristic curve (AUC). RESULTS A total of 675 tSCI patients within the RHSCIR were identified for our study. From 160 AIS B and C patients, our modified LR model demonstrated an AUC of 0.854 (95% CI 0.808–0.900). For the same cohort, van Middendorp et al.’s LR model generated an AUC of 0.833 (95% CI 0.771–0.895), while the artificial intelligence model presenting the highest degree of prognostic ability (neural networking) generated an AUC of 0.872. There were no significant differences in predictive accuracy between models (p>.05). The AUC for 515 AIS A and D patients utilizing van Middendorp et al.’s LR model was 0.954. CONCLUSIONS Application of our novel, seven-variable LR model, two previously tested LR models, and two artificial intelligence models demonstrated limited predictive accuracy for AIS B and C patients. These findings suggest that former models achieved strong prognostic accuracy through a high proportion of AIS A and D patients in their analysis cohort; furthermore, present models cannot effectively predict ambulatory outcomes in AIS B and C patients, regardless of methodology. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs.