1500. Can Machine Learning Algorithms Accurately Predict Discharge to Rehabilitation and Early Unplanned Readmissions Following Spinal Fusion? Analysis of a National Surgical Registry

Authors: Panagiotis Kerezoudis, MD; Anshit Goyal, MBBS; Che Ngufor, PhD; Brandon McCutcheon, MD; Curt Storlie, PhD; Mohamad Bydon, MD (Rochester, MN)

Introduction: Predictive models for discharge to rehabilitation and unplanned readmissions following spine surgery are sparse in the literature. We sought to utilize different machine learning algorithms to predict these outcomes in patients receiving spinal fusion.

Methods: We queried 2012-2013 ACS-NSQIP data for patients undergoing cervical or lumbar spinal fusion. Outcomes assessed included discharge to rehabilitation and unplanned readmission within 30 days after surgery. A total of 7 classification algorithms were assessed: Generalized Linear Model (logistic regression), elastic net, penalized discriminant analysis, naive Bayes, Artificial Neural Networks, Random Forest and Gradient Boosting Machines. Model performance was evaluated using overall accuracy, area-under-receiver operating characteristic curve (AUC), as well as sensitivity, specificity and positive and negative predictive values.

Results: Among 59, 145 cases of spinal fusion, incidence of discharge to rehabilitation/skilled nursing facility and 30-day unplanned readmission was 12.6% and 4.5% respectively. All classification algorithms showed excellent discrimination (AUC>0.85) for discharge to rehabilitation/skilled nursing facility with marginally higher sensitivity noted for Neural
networks (80%). Logistic regression showed comparable performance to other machine learning algorithms. By comparison, all models showed poorer predictive performance for unplanned readmission with AUC:0.63-0.66. Neural networks showed highest accuracy (71%) and specificity (72%) for predicting unplanned readmission compared to other algorithms. In general, better predictive performance was noted with models using imputed data.

Conclusion: In analysis of data from a multi-insitutional surgical registry, multiple machine learning algorithms were found to reliably predict discharge to rehabilitation/skilled nursing facility. Unplanned readmissions remained more challenging to predict. Logistic regression achieved equivalent predictive performance to more complex machine learning approaches. Future research should further validate these findings by utilizing larger datasets with a wider array of baseline variables, such as patient reported outcomes in order to generate possibly superior predictive models for each of the above outcomes.