Authors: James Edward Towner, MD; Viet-Duy Nguyen; Yan Li, PhD; Jiebo Luo, PhD; Yan Li, MD, PhD (Rochester, NY)

Introduction:

Accurate pre-surgical risk prediction and stratification potentially has value in informing perioperative decision making for physicians and patients. We sought to develop a spine surgery specific tool to examine risk factors associated with the following: days from operation to death, hospital admission > 30 days, any reoperation, unplanned reoperation, any hospital readmission, or unplanned hospital readmission.

Methods:

Multicenter, prospectively collected data from the American College of Surgeons National Surgical Quality Improvement Program database was used to examine fifty risk factors from over 2,000,000 patient records using machine learning techniques. Computational procedural terminology codes were used to define specific spine surgical interventions.  If any risk factors had 20% missing data, we excluded the patients with missing data. If missing data for any risk factor was higher than 80%, we excluded the entire risk factor from analysis. Fifty out of sixty-five risk factors satisfied this requirement and were used in the analysis. Management of imbalanced data was handled with upsampling (centroid cluster technique) and downsampling (synthetic minority over-sampling technique) methods. 10-fold cross-validation was utilized to evaluate the system with the dataset divided in an 8:1:1 ratio of training set: validation set: test set.

Results:

Gradient boosting techniques (XGBoost) and ensemble learning technique (random forest) performed better than regression methods (logistic regression technique). Our system achieved >90% accuracy for predicting every outcome which had sufficient data for analysis.

Conclusion:

Using machine learning techniques, we developed a spine specific surgical risk calculator with potential to predict the days from operation to death, hospital admission >30 days, reoperation, and hospital readmission.