Development and validation of machine learning models for the prediction of blunt cerebrovascular injury in children

• Published in: Journal of pediatric surgery 2022-04, Vol.57 (4), p.732-738
• Main Authors: Farzaneh, Cyrus A., Schomberg, John, Sullivan, Brittany G., Guner, Yigit S., Nance, Michael L., Gibbs, David, Yu, Peter T.
• Format: Article
• Language: English
• Subjects: Adult; Cerebrovascular Trauma - diagnosis; Cerebrovascular Trauma - epidemiology; Child; Denver; Humans; Machine Learning; Memphis model; Pediatric blunt cerebrovascular injury; Retrospective Studies; Skull Fractures; Wounds, Nonpenetrating - complications; Wounds, Nonpenetrating - diagnosis; Wounds, Nonpenetrating - epidemiology
• ISSN: 0022-3468; 1531-5037
• DOI: 10.1016/j.jpedsurg.2021.11.008

Blunt cerebrovascular injury (BCVI) is a rare finding in trauma patients. The previously validated BCVI (Denver and Memphis) prediction model in adult patients was shown to be inadequate as a screening option in injured children. We sought to improve the detection of BCVI by developing a prediction model specific to the pediatric population. The National Trauma Databank (NTDB) was queried from 2007 to 2015. Test and training datasets of the total number of patients (885,100) with complete ICD data were used to build a random forest model predicting BCVI. All ICD features not used to define BCVI (2268) were included within the random forest model, a machine learning method. A random forest model of 1000 decision trees trying 7 variables at each node was applied to training data (50% of the dataset, 442,600 patients) and validated with test data in the remaining 50% of the dataset. In addition, Denver and Memphis model variables were re-validated and compared to our new model. A total of 885,100 pediatric patients were identified in the NTDB to have experienced blunt pediatric trauma, with 1,998 (0.2%) having a diagnosis of BCVI. Skull fractures (OR 1.004, 95% CI 1.003–1.004), extremity fractures (OR 1.001, 95% 1.0006–1.002), and vertebral injuries (OR 1.004, 95% CI 1.003–1.004) were associated with increased risk for BCVI. The BCVI prediction model identified 94.4% of BCVI patients and 76.1% of non-BCVI patients within the NTDB. This study identified ICD9/ICD10 codes with strong association to BCVI. The Denver and Memphis criteria were re-applied to NTDB data to compare validity and only correctly identified 13.4% of total BCVI patients and 99.1% of non BCVI patients. The prediction model developed in this study is able to better identify pediatric patients who should be screened with further imaging to identify BCVI. Retrospective diagnostic study–level III evidence.