P19. Predicting readmission following fusion for scoliosis in pediatric patients: A machine-learning approach

Abstract

BACKGROUND CONTEXT Spine fusion surgery is a common treatment for scoliosis in pediatric patients. Hospital readmission within 30-days of surgery is a costly event. While studies have been undertaken to predict readmission within 30-days for orthopedic procedures, such as total shoulder arthroplasty and adult spine fusion surgery, no research has been conducted on the utility of machine learning algorithms to predict postoperative complications for pediatric patients undergoing spinal procedures. PURPOSE The study presented here analyzes the predictive capacity of a set of machine learning algorithms to preoperatively predict 30-day readmission after posterior spine fusion surgery for pediatric patients with scoliosis. STUDY DESIGN/SETTING This study was an analysis of NSQIP pediatric data. PATIENT SAMPLE The NSQIP pediatric database was queried between 2012 and 2018 to select patients who had undergone posterior arthrodesis surgery for scoliosis treatment. OUTCOME MEASURES A 30-day readmission after pediatric posterior arthrodesis surgery. METHODS The NSQIP pediatric database was queried to select patients who had undergone posterior arthrodesis surgery for scoliosis treatment. Predictive variables of interest were quantitative variables such as height, weight, and age, as well as categorical variables such as race and the presence or absence of comorbidities such as asthma, lung disease, ASA Class, cardiac risk factors, cognitive function, seizure presence, cerebral palsy, ACQ abnormality, and neuromuscular disorder. Patients who had missing data for one or more variables of interest were excluded from the analysis. Python s Sci-Kit Learn package was utilized to run five machine learning algorithms: logistic regression (LR), decision tree (DT), random forest (RF), gradient boosting (GB) and neural network (NN). Patients were randomly split into two groups, where 70% of patients were used to train the neural network, while the remaining 30% were used to test the validity of the neural network. The area under the curve (AUC) and prediction accuracy were used to determine the capacity of the algorithm in predicting 30-day readmission. A multivariable logistic regression was performed in R to identify risk factors for readmission. The significance value was set at 0.05. RESULTS Initially, 24,271 patients met the inclusion criteria. However, after excluding patients who had missing variables, 17,873 patients were analyzed, 825 of which were readmitted within 30 days of surgery (4.62%). The machine-learning algorithms exhibited AUC values between 0.513 and 0.736, with the GB algorithm performing the best and the DT algorithm performing the worst, and prediction accuracies between 92.8% (DTC) and 96.5% (GB). The logistic regression identified obesity (Odds Ratio (OR): 2.51, 95% Confidence Interval (CI):2.04-3.10, p-value CONCLUSIONS Machine learning algorithms are potentially valuable tools for predicting readmission after pediatric spine fusion surgery. These algorithms may assist clinicians and patients in determining the best treatment care plans to optimize outcomes and minimize re-admission risk. Examination of the external validity of these paradigms on other state, regional, and national databases to determine the applicability and accuracy of the algorithms may assist in someday bringing these predictive tools to clinical utilization. FDA DEVICE/DRUG STATUS This abstract does not discuss or include any applicable devices or drugs.