Feasibility of Machine Learning in the Prediction of Short-Term Outcomes Following Anterior Cervical Discectomy and Fusion

Abstract

Increased emphasis is being placed on efficiency and resource utilization when performing anterior cervical discectomy and fusion (ACDF), and accurate prediction of complications is increasingly important to optimize care. This study aimed to compare predictive models for postoperative complications following ACDF using machine learning (ML) models based on traditional comorbidity indices. In this retrospective case series, the American College of Surgeons National Surgical Quality Improvement Program database was queried between 2011 and 2017 for all elective, primary ACDF cases. Levels of surgery, use of interbody implants, and graft selection were calculated by procedural codes. Six ML algorithms were constructed using available preoperative and intraoperative features. The overall dataset was randomly split into training (80%) and validation (20%) subsets wherein the training subset optimized the model, and the validation subset was evaluated for accuracy. ML models were compared with models constructed from American Society of Anesthesiologists classification and frailty index alone. There were 42,194 ACDF cases eligible for inclusion. Mean age was 47.7 ± 11.6 years, body mass index was 30.4 ± 6.7, and levels of operation were 1.6 ± 0.7. ML algorithms uniformly outperformed comorbidity indices in predicting complications. Logistic regression ML algorithm was the best performing for predicting any adverse event (area under the curve [AUC] 0.73), transfusion (AUC 0.90), surgical site infection (AUC 0.63), and pneumonia (AUC 0.80). Gradient boosting trees ML algorithm was the best performing for predicting extended length of stay (AUC 0.73). ML algorithms modeled the development of postoperative adverse events with superior accuracy to that of comorbidity indices and may guide preoperative clinical decision making before ACDF.