Machine Learning for Predictive Modeling of 90-day Readmission, Major Medical Complication, and Discharge to a Facility in Patients Undergoing Long Segment Posterior Lumbar Spine Fusion

Study Design.Retrospective case control study.Objective.To develop predictive models for postoperative outcomes after long segment lumbar posterior spine fusion (LSLPSF).Summary of Background Data.Surgery for adult spinal deformity is effective for treating spine-related disability; however, it has high complication and readmission rates.Methods.Patients who underwent LSLPSF (three or more levels) were identified in State Inpatient Database. Data was queried for discharge-to-facility (DTF), 90-day readmission, and 90-day major medical complications, and demographic, comorbid, and surgical data. Data was partitioned into training and testing sets. Multivariate logistic regression, random forest, and elastic net regression were performed on the training set. Models were applied to the testing set to generate AUCs. AUCs between models were compared using the method by DeLong et al.Results.37,852 patients were analyzed. The DTF, 90-day readmission, and 90-day major medical complication rates were 35.4%, 19.0%, and 13.0% respectively. For DTF, the logistic regression AUC was 0.77 versus 0.75 for random forest and 0.76 for elastic net (P<0.05 for all comparisons). For 90-day readmission, the logistic regression AUC was 0.65, versus 0.63 for both random forest and elastic net (P<0.05 for all comparisons). For 90-day major medical complications, the logistic regression AUC was 0.70, versus 0.69 for random forest and 0.68 for elastic net (P<0.05 for all comparisons).Conclusion.This study created comprehensive models to predict discharge to facility, 90-day readmissions, and 90-day major medical complications after LSLPSF. This information can be used to guide decision making between the surgeon and patient, as well as inform value-based payment models.Level of Evidence: 3.