(I-587) Factors Associated with 30-Day Readmissions Following Posterior Cervical Fusion: Comparison of Machine Learning and Traditional Statistical Models
Associate Professor of Neurosurgery Stanford University Stanford, California, United States
Introduction: Machine learning serves as a promising tool to identify and subsequently reduce readmissions following cervical spine surgery. This study aimed to identify factors associated with readmissions after posterior cervical fusion (PCF) using machine learning and logistic regression models.
Methods: The Optum Clinformatics database was used to identify patients undergoing posterior cervical fusion with instrumentation between 2004 and 2017. Four machine learning models (random forest; stochastic gradient boosting machine; extreme gradient boosting; penalized logistic regression chosen using elastic net variants of the least absolute shrinkage and selection operator) and a multivariable logistic regression model were generated and used to identify factors most influential in predicting unplanned 30-day readmission. These models were also compared to one another in terms of ability to predict unplanned 30-day readmissions (as measured by area under the receiver operating curve; AUC).
Results: A total of 3,814 patients were identified, of which 874 (22.9%) were readmitted within 30 days of initial surgery. Discharge status, depression diagnosis, and number of procedure codes were the most influential factors for the logistic regression model, while discharge status, admission cost, and length of stay visits had greatest relevance for the GBM model. The gradient boosting machine model yielded the highest mean AUC of all machine learning models, and significantly outperformed the logistic regression model in predicting unplanned 30-day readmission (mean AUC 0.864 vs. 0.829, p=0.0001).
Conclusion : Factors associated with readmission following posterior cervical fusion vary based on standard logistic regression and machine learning models used, highlighting the complementary roles these models have in predicting 30-day readmission. Gradient boosting machine model yielded greatest predictive ability in terms of readmission following posterior cervical fusion. These findings encourage further research, development, and use of machine learning models to accurately identify high-risk patients, guide resource allocation to reduce their likelihood of readmission and, in turn, decrease the utilization burden on our healthcare system.
