P126. Development of a novel ensemble machine learning algorithm for prediction of complications and readmission after posterior cervical spinal fusion

Abstract

<h3>BACKGROUND CONTEXT</h3> Cervical spinal fusion is one of the most commonly performed surgeries in the United States and is rapidly growing in prevalence. The average age and comorbidity burden of patients undergoing cervical fusion has also increased, elevating the risk of perioperative complications and unplanned readmissions. Given the cost and morbidity associated with these poor outcomes, accurate preoperative risk stratification is of great utility. Advanced machine learning (ML) methods have become increasingly employed in spinal surgery due to their ability to recognize complex, non-linear relationships between variables. ML models for preoperative risk stratification of patients undergoing posterior cervical fusion remain limited. <h3>PURPOSE</h3> We aim to develop a novel ensemble ML algorithm for prediction of major perioperative complication and unplanned readmission after posterior cervical fusion. Our secondary aim is to compare its performance against benchmark ML models and logistic regression (LR). <h3>STUDY DESIGN/SETTING</h3> Retrospective, cohort study. <h3>PATIENT SAMPLE</h3> Patients 18 years or older at a non-federal California hospital who underwent posterior cervical spinal fusion. <h3>OUTCOME MEASURES</h3> Readmission within 30 days, major perioperative complications (venous thromboembolism within 30 days, myocardial infarction within 7 days, pneumonia within 7 days, systemic infection within 7 days, surgical site bleeding within 90 days, and wound complications within 90 days). <h3>METHODS</h3> This is a retrospective cohort study of adult patients who underwent instrumented or non-instrumented posterior cervical spinal fusion at any California hospital between 2015-2017. We developed a ML-based model predicting complication risk using AutoPrognosis, an automated ML framework that configures the optimally performing ensemble of ML-based prognostic models. We compared our model with LR and four standard ML models (XGBoost, Gradient Boosting, AdaBoost, Random Forest). Predictive performances were assessed using area under the receiver operating characteristic curve (AUROC). We ranked the contribution of the included variables to model performance. <h3>RESULTS</h3> Of the 6,822 patients who met inclusion criteria, there were 1,279 cases of at least one major complication or readmission (18.8%). The ensemble AutoPrognosis model had superior risk prediction (AUROC: 0.679 + 0.011) compared to LR (0.651 + 0.014). Furthermore, this model outperformed the four benchmark ML algorithms. The variables most important for AutoPrognosis include history of malignancy, number of comorbidities, and hospital volume. Five of the 10 most important features for AutoPrognosis were markedly less important for LR: teaching hospital status, implant complications, pneumonia, stroke and chronic kidney disease. <h3>CONCLUSIONS</h3> We report the use of a novel ensemble ML algorithm for prediction of major perioperative complications after posterior cervical fusion. Our novel AutoPrognosis algorithm is well-calibrated and outperforms LR and four standard ML algorithms. By automating the choice of appropriate model as well as tuning of model hyperparameters, AutoPrognosis can be used by clinicians who may not possess an in-depth knowledge of ML methods – facilitating the use of advanced ML methods in mainstream clinical research. Notably, the predictors most important for AutoPrognosis are different from those for LR. This suggests that the superior discriminative capability of AutoPrognosis results from not just its ability to select among different ML models but also by capturing complex non-linear relationships between variables that logistic regression is unable to capture. <h3>FDA DEVICE/DRUG STATUS</h3> This abstract does not discuss or include any applicable devices or drugs.