A Machine Learning Algorithm to Identify Patients with Tibial Shaft Fractures at Risk for Infection After Operative Treatment
- Additional Document Info
- View All
BACKGROUND: Risk stratification of individual patients who are prone to infection would allow surgeons to monitor high-risk patients more closely and intervene early when needed. This could reduce infection-related consequences such as increased health-care costs. The purpose of this study was to develop a machine learning (ML)-derived risk-stratification tool using the SPRINT (Study to Prospectively Evaluate Reamed Intramedullary Nails in Patients with Tibial Fractures) and FLOW (Fluid Lavage of Open Wounds) trial databases to estimate the probability of infection in patients with operatively treated tibial shaft fractures (TSFs). METHODS: Patients with unilateral TSFs from the SPRINT and FLOW trials were randomly split into derivation (80%) and validation (20%) cohorts. Random forest algorithms were used to select features that are relevant to predicting infection. These features were included for algorithm training. Five ML algorithms were trained in recognizing patterns associated with infection. The performance of each ML algorithm was evaluated and compared based on (1) the area under the ROC (receiver operating characteristic) curve (AUC), (2) the calibration slope and the intercept, and (3) the Brier score. RESULTS: There were 1,822 patients included in this study: 170 patients (9%) developed an infection that required treatment, 62 patients (3%) received nonoperative treatment with oral or intravenous antibiotics, and 108 patients (6%) underwent subsequent surgery in addition to antibiotic therapy. Random forest algorithms identified 7 variables that were relevant for predicting infection: (1) Gustilo-Anderson or Tscherne classification, (2) bone loss, (3) mechanism of injury, (4) multitrauma, (5) AO/OTA fracture classification, (6) age, and (7) fracture location. Training of the penalized logistic regression algorithm resulted in the best-performing prediction model, with AUC, calibration slope, calibration intercept, and Brier scores of 0.75, 0.94, 0.00, and 0.076, respectively, in the derivation cohort and 0.81, 1.07, 0.09, and 0.079, respectively, in the validation cohort. CONCLUSIONS: We developed an ML prediction model that can estimate the probability of infection for individual patients with TSFs based on patient and fracture characteristics that are readily available at hospital admission. LEVEL OF EVIDENCE: Prognostic Level III. See Instructions for Authors for a complete description of levels of evidence.
has subject area