185. A predictive model for C5 palsy after instrumented cervical fusion

Abstract

<h3>BACKGROUND CONTEXT</h3> C5 palsy is a common postoperative complication of cervical fusion, occurring in 5-24% of patients. Although their prognosis is typically favorable, patients who develop C5 palsy are often unable to perform basic activities of daily living, resulting in increased health care costs and decreased quality of life. Accurate prediction of C5 palsy is thus of clinical utility. Predictive models assessing C5 palsy remain rare in the literature. <h3>PURPOSE</h3> We aim to build an ensemble ML algorithm for prediction of C5 palsy after instrumented cervical fusion. Furthermore, we aim to compare the performance of this model with logistic regression and other standard ML models. We hypothesize that the ML model will identify novel predictive features. <h3>STUDY DESIGN/SETTING</h3> Retrospective, cohort study. <h3>PATIENT SAMPLE</h3> Adult patients who underwent instrumented cervical fusion at a tertiary-care academic medical center between 2013-2020 were included. <h3>OUTCOME MEASURES</h3> The primary outcome was C5 palsy within index admission or follow-up encounters. <h3>METHODS</h3> We developed an ensemble ML model predicting C5 palsy risk using AutoPrognosis, an automated ML framework that configures an ensemble of ML-based models and traditional statistics. We compared this model with logistic regression and four ML models (XGBoost, gradient boosting, AdaBoost, random forest). Discrimination was assessed using area under the receiver operating characteristic curve (AUROC). Calibration was assessed using calibration slope and calibration intercept. We additionally ranked importance of each included feature for model performance utilizing a partial dependence function. <h3>RESULTS</h3> A total of 1,024 patients were included in this study, with 52 cases of C5 palsy (5.1%). The ensemble model demonstrated the highest discrimination (AUROC: 0.733) compared to regression (AUROC: 0.710). This model was well-calibrated with a calibration slope of 0.773 and a calibration intercept of -0.025. The following features were the most important for ensemble model performance: preoperative nonmotor neurologic symptoms, spinal trauma, corpectomy, spinal malignancy, posterior fusion, presence of paresthesias, psychiatric comorbidity, fixation for spinal fracture, degenerative disc disease, interbody fusion. Five of the 10 most important features for the ensemble model were markedly less important for regression. <h3>CONCLUSIONS</h3> We report a well-calibrated ensemble ML model that predicts C5 palsy after instrumented cervical fusion with high discrimination. This represents the first ML algorithm for prediction of C5 palsy after cervical fusion to our knowledge. Individualized risk prediction for patients may facilitate improved preoperative patient counseling and risk stratification. This tool may also aid with addressing potentially modifiable risk factors such as psychiatric comorbidity, helping to decrease likelihood of C5 palsy. <h3>FDA DEVICE/DRUG STATUS</h3> This abstract does not discuss or include any applicable devices or drugs.