Abstract
Introduction/Purpose:
Acute ankle fractures are a common injury and represent a significant financial burden to the healthcare system. Reimbursement following ankle fractures is complicated by bundled payments within the 90-day global period which do not account for unpredictable complications. Comorbidity indices, such as the American Society of Anesthesiologists (ASA) classification, are often correlated with complications; however, the positive predictive value of these metrics have not been investigated. As insurers continue to transition towards bundled reimbursements, greater attentiveness towards providing efficient care is paramount. The purpose of this investigation is to develop a machine learning (ML) model that is predictive of complications following operative management of acute ankle fractures. This study will also compare ML algorithms to legacy indices to assess the predictive value in predicting complications.
Methods:
The American College of Surgeons-National Surgical Quality Improvement Program database was queried via Current Procedural Terminology (CPT) from 2015-2018 for open reduction internal fixation (ORIF) cases for ankle fractures, including ORIF of the medial malleolus, posterior malleolus, lateral malleolus, bimalleolar, trimalleolar, syndesmosis. Training and validation sets were created by randomly assigning 80% and 20% of the data set. Included variables were age, body mass index (BMI), operative time, smoking status, comorbidities, diagnosis, and preoperative hematocrit and albumin. Complications included any adverse event, transfusion, surgical site infection, return to the operating room, deep vein thrombosis or pulmonary embolism, pneumonia, urinary tract infection, cerebrovascular accident, cardiac arrest, myocardial infarction. Each ML algorithm was compared with one another and to a baseline model using American Society of Anesthesiologists (ASA) classification. Model strength was evaluated by calculating the area under the curve (AUC) for receiver operating characteristic (ROC) of any adverse event.
Results:
We identified a total of 28,736 ORIF cases. Mean age and BMI were 32.2 ± 18.6 years, 30.8 ± 7.6 kg/m^2, respectively. The overall incidence of any medical complication was 3.4% (n=975) with surgical site infection (1.3%, n=386) and return to operating room (1.3%, n=367) being the most common. Percentage hematocrit, BMI, operative time, and age were of highest importance in outcome prediction. Logistic regression (AUC: 72%) and gradient boosting (AUC: 72%) ML algorithms outperformed ASA classification (AUC: 69%) for predicting any adverse event following ORIF of ankle fractures.
Conclusion:
Legacy comorbidity indices are simple metrics that can be easily constructed from patient demographic variables and past medical history. However, the predictive power is limited as they have shown a low positive predictive value for predicting complications. Machine learning algorithms can calculate patient-specific risk for postoperative complications which may adjust perioperative care and site of surgery. These models have the capability to optimize utilization of healthcare resources and minimize excessive expenditure by providing patient-specific care.
Receiver Operating Curve-Area under the Curve Analysis for Adverse Events Following ORIF of Ankle Fractures