Predicting postoperative complications, adverse 90-day readmissions, and 2-year reoperations is critical for improving surgical decision-making, prognostication, and planning because of the prevalence of anterior cervical discectomy and fusion (ACDF). There have been attempts to use machine learning to foretell ACDF-related postoperative problems. However, they have been constrained by small sample sizes and/or inappropriate models. These investigations were deemed successful at an area under the curve (AUC) of less than or equal to 0.70. Other methods, not confined to ACDF, zeroed in on particular kinds of complications and achieved an AUC of 0.70–0.76. Adult patients who received an ACDF surgery from 2007 to 2016 (N=176,816) were identified by querying the IBM MarketScan Commercial Claims and Encounters 

Database and Medicare Supplement. For the purpose of predicting complications within 90 days of surgery, readmission within 90 days of surgery, and reoperation within 2 years, deep neural networks were compared against logistic regression and support vector machines. To further approximately reach an upper bound, researchers created random deep-learning model architectures and trained them on the 90-day complication task. Finally, they used deep learning to examine the significance of each input variable in predicting ACDF-related problems in the first 90 days after surgery. Models built using deep neural networks had an AUC of 0.832, 0.713, and 0.671 when predicting complications within 90 days, readmissions within 90 days, and repeat surgeries within 2 years, respectively. AUCs of 0.820, 0.712, and 0.671 were found using logistic regression. 

The results of support vector machine methods were much lower. Estimated maximum performance of 0.832 was found for deep learning. The biggest predictors of surgical problems within the first 90 days were myelopathy, age, human immunodeficiency virus, prior myocardial infarctions, obesity, and documentary weakness. Upon completion of multi-center validation, the deep neural network may be utilized to forecast complications for clinical applications. The findings imply that little to no new information is to be gained from the interactions between the input variables. Novel variables should be identified in future research to improve prediction ability.