Supervised machine learning algorithms used to predict post-surgical outcomes following anterior surgical fixation of odontoid fractures

Abstract

Background: Odontoid fractures have a high mortality rate, and numerous classification systems have previously predicted surgical outcomes with mixed consensus. We generated a machine learning (ML) construct to predict post-operative adverse events following anterior (ORIF) of odontoid fractures. Methods: 266 patients from the American college of surgeons-national surgical quality improvement program (ACS-NSQIP) with anterior ORIF (CPT 22318) of odontoid fractures from 2008-2018 were analyzed using ML algorithms random forest classifier (RF), gradient boosting classifier (GB), support vector machine classifier (SVM), Gaussian Naive Bayes classifier (GNB), and multi-layer perceptron classifier (MLP), and were compared to logistic regression classifier (LR). Algorithms predicted increased length of stay (LOS), need for transfusion (Transf), non-home discharge (NHD), and any adverse event (AAE). Permutation feature importance (PFI) identified risk factors. Results: ML algorithms outperformed LR. The average AUC for predicting Transf was 0.635 (accuracy=77.4%), extended LOS=0.652 (accuracy 59.6%), NHD 0.788 (accuracy=71.9%) and AAE 0.649 (accuracy 68.1%). GB performed highest for Transf (AUC=0.861), identifying operative time (PFI 0.253, p=0.016). GB and RF performed equally for NHD (AUC=0.819), highlighting preoperative hematocrit (PFI=0.157, p<0.001). GB predicted AAE (AUC=0.720) also identifying preoperative hematocrit (PFI=0.112, p<0.001). RF predicted extended LOS (AUC=0.669) highlighting preoperative hematocrit (PFI=0.049, p<0.001). Conclusions: ML outperformed LR, successfully predicting Transf, extended LOS, NHD, and AAE for anterior ORIF of odontoid fractures. Our construct may complement conventional risk stratification to reduce adverse outcomes and excess cost.