Comparison of Deep Learning and Classical Machine Learning Algorithms to Predict Post-operative Outcomes for Anterior Cervical Discectomy and Fusion Procedures with State-of-the-art Performance

Study Design. Retrospective cohort. Objective. Due to Anterior cervical discectomy and fusion (ACDF) popularity, it is important to predict post-operative complications, unfavorable 90-day readmissions, and 2-year re-operations to improve surgical decision making, prognostication and planning. Summary of Background Data. Machine learning has been applied to predict post–operative complications for ACDF; however, studies were limited by sample size and model type. These studies achieved 0.70 AUC. Further approaches, not limited to ACDF, focused on specific complication types, and resulted in AUC between 0.70–0.76. Methods. The IBM MarketScan Commercial Claims and Encounters Database and Medicare Supplement were queried from 2007-2016 to identify adult patients who underwent an ACDF procedure (N=176,816). Traditional machine learning algorithms, logistic regression, support vector machines, were compared with deep neural networks to predict: 90-day post-operative complications, 90-day readmission, and 2-year reoperation. We further generated random deep learning model architectures and trained them on the 90-day complication task to approximate an upper bound. Lastly, using deep learning, we investigated the importance of each input variable for the prediction of 90-day post-operative complications in ACDF. Results. For the prediction of 90-day complication, 90-day readmission, and 2-year reoperation, the deep neural network-based models achieved area under the curve (AUC) of 0.832, 0.713, and 0.671. Logistic regression achieved AUCs of 0.820, 0.712, and 0.671. SVM approaches were significantly lower. The upper bound of deep learning performance was approximated as 0.832. Myelopathy, age, HIV, previous myocardial infarctions, obesity, and documentary weakness were found to be the strongest variable to predict 90-day post-operative complications. Conclusions. The deep neural network may be used to predict complications for clinical applications after multi-center validation. The results suggest limited added knowledge exists in interactions between the input variables used for this task. Future work should identify novel variables to increase predictive power.