Robert Zura1, J Tracy Watson2, Thomas Einhorn3, Samir Mehta4, Gregory J Della Rocca5, Ze Xiong6, Zhe Wang7, John Jones8, R Grant Steen9. 1. Dept. of Orthopaedic Surgery, Louisiana State University Medical Center, New Orleans, LA, United States. Electronic address: rzura@lsuhsc.edu. 2. Dept. of Orthopaedic Surgery, Saint Louis University School of Medicine, St. Louis, MO, United States. Electronic address: watsonjt@slu.edu. 3. Dept. of Orthopaedic Surgery, NYU Langone Medical Center, New York, NY, United States. Electronic address: Thomas.Einhorn@nyumc.org. 4. Dept. of Orthopaedic Surgery, Hospital of the University of Pennsylvania, Philadelphia, PA, United States. Electronic address: Samir.Mehta@uphs.upenn.edu. 5. Dept. of Orthopaedic Surgery, Duke University, Durham, NC, United States. Electronic address: gregory.della.rocca@duke.edu. 6. Dept. of Statistics, North Carolina State University, Raleigh, NC, United States. Electronic address: zxiong@ncsu.edu. 7. Dept. of Statistics, North Carolina State University, Raleigh, NC, United States. Electronic address: zwang28@ncsu.edu. 8. Clinical Affairs, Bioventus LLC, Durham, NC, United States. Electronic address: John.Jones@bioventusglobal.com. 9. Dept. of Orthopaedic Surgery, Louisiana State University Medical Center, New Orleans, LA, United States. Electronic address: Grant.Steen@bioventusglobal.com.
Abstract
INTRODUCTION: The epidemiology of fracture nonunion has been characterized so it is potentially possible to predict nonunion using patient-related risk factors. However, prediction models are currently too cumbersome to be useful. We test a hypothesis that nonunion can be predicted with ≤10 variables, retaining the predictive accuracy of a full model with 42 variables. METHODS: We sought to predict nonunion with prospectively-acquired inception cohort data for 18 different bones, using the smallest possible number of variables that did not substantially decrease prediction accuracy. An American nationwide claims database of ∼90.1 million participants was used, which included medical and drug expenses for 2011-2012. Continuous enrollment was required for 12 months after fracture, to allow sufficient time to capture a nonunion diagnosis. Health claims were evaluated for 309,330 fractures. A training dataset used a random subset of 2/3 of these fractures, while the remaining fractures formed a validation dataset. Multivariate logistic regression and stepwise logistic regression were used to identify variables predictive of nonunion. P value and the Akaike Information Criterion (AIC) were used to select variables for reduced models. Area-under-the-curve (AUC) was calculated to characterize the success of prediction. RESULTS: Nonunion rate in 18 fracture locations averaged 4.93%. Algorithms to predict nonunion in 18 locations in the full-model validation set had average AUC=0.680 (±0.034). In the reduced models, average validation set AUC=0.680 (±0.033) and the average number of risk factors required for prediction was 7.6. There was agreement across training set, validation set, and reduced set; in tibia, reduced model validation AUC=0.703, while the full-model validation AUC=0.709. Certain risk factors were important for predicting nonunion in ≥10 bones, including open fracture, multiple fracture, osteoarthritis, surgical treatment, and use of certain medications, including anticoagulants, anticonvulsants, or analgesics. DISCUSSION: Nonunion can be predicted in 18 fracture locations using parsimonious models with <10 patient demography-related risk factors. The model reduction approach used results in simplified models that have nearly the same AUC as the full model. Reduced algorithms can predict nonunion because risk factors important in the full models remain important in the reduced models. This prognostic inception cohort study provides Level I evidence.
INTRODUCTION: The epidemiology of fracture nonunion has been characterized so it is potentially possible to predict nonunion using patient-related risk factors. However, prediction models are currently too cumbersome to be useful. We test a hypothesis that nonunion can be predicted with ≤10 variables, retaining the predictive accuracy of a full model with 42 variables. METHODS: We sought to predict nonunion with prospectively-acquired inception cohort data for 18 different bones, using the smallest possible number of variables that did not substantially decrease prediction accuracy. An American nationwide claims database of ∼90.1 million participants was used, which included medical and drug expenses for 2011-2012. Continuous enrollment was required for 12 months after fracture, to allow sufficient time to capture a nonunion diagnosis. Health claims were evaluated for 309,330 fractures. A training dataset used a random subset of 2/3 of these fractures, while the remaining fractures formed a validation dataset. Multivariate logistic regression and stepwise logistic regression were used to identify variables predictive of nonunion. P value and the Akaike Information Criterion (AIC) were used to select variables for reduced models. Area-under-the-curve (AUC) was calculated to characterize the success of prediction. RESULTS: Nonunion rate in 18 fracture locations averaged 4.93%. Algorithms to predict nonunion in 18 locations in the full-model validation set had average AUC=0.680 (±0.034). In the reduced models, average validation set AUC=0.680 (±0.033) and the average number of risk factors required for prediction was 7.6. There was agreement across training set, validation set, and reduced set; in tibia, reduced model validation AUC=0.703, while the full-model validation AUC=0.709. Certain risk factors were important for predicting nonunion in ≥10 bones, including open fracture, multiple fracture, osteoarthritis, surgical treatment, and use of certain medications, including anticoagulants, anticonvulsants, or analgesics. DISCUSSION: Nonunion can be predicted in 18 fracture locations using parsimonious models with <10 patient demography-related risk factors. The model reduction approach used results in simplified models that have nearly the same AUC as the full model. Reduced algorithms can predict nonunion because risk factors important in the full models remain important in the reduced models. This prognostic inception cohort study provides Level I evidence.
Authors: Bridget Sinnott; Cara Ray; Frances Weaver; Beverly Gonzalez; Elizabeth Chu; Sarah Premji; Mattie Raiford; Rachel Elam; Scott Miskevics; Stephen Parada; Laura Carbone Journal: JBMR Plus Date: 2022-01-05
Authors: Robert Zura; Sue C Kaste; Michael J Heffernan; William K Accousti; Dominic Gargiulo; Zhe Wang; R Grant Steen Journal: Medicine (Baltimore) Date: 2018-08 Impact factor: 1.817