A machine learning model using all patient variables predicted 30-day readmission following lumbar laminectomy with 95.3% accuracy (AUC 0.8059), compared to 79.6% using only predischarge variables.
Cohort (n=26,869)
Yes
Can machine learning models accurately predict 30-day hospital readmission in patients undergoing lumbar laminectomy?
Machine learning models using predischarge variables can predict 30-day hospital readmission following lumbar laminectomy with moderate accuracy (AUC 0.69), potentially aiding in targeted interventions.
Effect estimate: AUC 0.8059
OBJECTIVE Hospital readmission contributes substantial costs to the healthcare system. The purpose of this investigation was to create a predictive machine learning model to identify lumbar laminectomy patients at risk for postoperative hospital readmission. METHODS Patients who had undergone a lumbar laminectomy procedure in the period from 2011 to 2014 were isolated from the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) database. Demographic characteristics and clinical factors, including complications, comorbidities, length of stay, age, and body mass index, were analyzed in relation to whether or not the patients had been readmitted to the hospital within 30 days after their procedure by utilizing independent-samples t-tests. Supervised gradient boosting machine learning was then used to create two models to predict readmission—one with all collected patient variables and one with only the variables known prior to hospital discharge. RESULTS A total of 26,869 patients were evaluated, 5.59% (1501 patients) of whom had an unplanned readmission to the hospital within 30 days of their procedure. Readmitted patients were older and had a greater number of complications and comorbidities, longer operative time, longer hospital stay, higher BMI, and higher work relative value unit (RVU) operation score (p < 0.01). They also had a worse health status prior to surgery (p < 0.01) and were more likely to be sent to a skilled discharge destination postoperatively (p < 0.01). The model with all patient variables accurately identified 49.6% of readmissions with an overall accuracy of 95.33% (area under the curve AUC = 0.8059), with postdischarge complications and comorbidities as the most important predictors. The predictive model built with only clinical information known predischarge identified 40.5% of readmitted patients with an accuracy of 79.55% (AUC = 0.6901), with discharge destination, comorbidities, and American Society of Anesthesiologists (ASA) classification as the most influential factors in identifying readmitted patients. CONCLUSIONS In this study, the authors analyzed hospital readmissions following laminectomy and developed predictive models to identify readmitted patients with an accuracy of over 95% using all variables and over 79% when using only predischarge variables. Using only the variables available predischarge, the authors created a model capable of predicting 40% of the readmitted patients. This study provides data that will assist in the development of predictive models for readmission and the creation of interventions to prevent readmission in high-risk patients.
Kalagara et al. (Wed,) conducted a cohort in Lumbar laminectomy (n=26,869). Machine learning predictive modeling was evaluated on Unplanned hospital readmission within 30 days (AUC 0.8059). A machine learning model using all patient variables predicted 30-day readmission following lumbar laminectomy with 95.3% accuracy (AUC 0.8059), compared to 79.6% using only predischarge variables.