ABSTRACT
BACKGROUND: The recognition of child physical abuse can be challenging and often requires a multidisciplinary assessment. Deep learning models, based on clinical characteristics, laboratory studies, and imaging findings, were developed to facilitate unbiased identification of children who may have been abused. METHODS: Level 1 pediatric trauma center registry data from 1/1/2010-1/31/2020 were queried for abused children and matched participants with non-abusive trauma. Observations were de-identified and divided into training and validation sets. Model 1 used patient demographics (age, gender, and insurance type) and clinical characteristics (vital signs, shock index pediatric age-adjusted, Glasgow Coma Score, lactate, base deficit, and international normalized ratio). Model 2 used the same features as Model 1, but with the text of the radiology reports of head computed tomography, brain MRIs, and skeletal surveys. Google's latest BERT Natural Language Processing (NLP) model, which was pre-trained on a large corpus, was used for fine-tuning Model 2. Accuracy, sensitivity, specificity, F1 scores, and positive predictive values were used to assess performance. RESULTS: Of 1,312 patients, 737 (56.2%) were abused. Model 1 had an accuracy of 86.3%, sensitivity of 87.2%, specificity of 85.1%, F1 score of 0.86, and positive predictive value (PPV) of 88.7% for the validation set with an area under the receiver Operating Curve (ROC AUC) of 0.86. NLP based Model 2 had an accuracy of 93.4%, sensitivity 92.5%, specificity of 94.6%, F1 score of 0.93, and PPV of 95.9% for the validation set, with a ROC AUC of 0.94. Most features had weak individual correlations with abuse (r < 0.3). CONCLUSIONS: Deep learning models accurately distinguished child physical abuse from non-abuse, and NLP further improved the accuracy of the models. Such models could be developed to run in real-time in the electronic medical record and alert clinicians when certain criteria are met, which would prompt them to pursue the diagnosis of abuse. LEVEL OF EVIDENCE: III STUDY TYPE: Diagnostic.
Subject(s)
Deep Learning , Natural Language Processing , Child , Electronic Health Records , Humans , Physical Abuse , RadiographyABSTRACT
BACKGROUND: The principal triggers for intervention in the setting of pediatric blunt solid organ injury (BSOI) are declining hemoglobin values and hemodynamic instability. The clinical management of BSOI is, however, complex. We therefore hypothesized that state-of-art machine learning (computer-based) algorithms could be leveraged to discover new combinations of clinical variables that might herald the need for an escalation in care. We developed algorithms to predict the need for massive transfusion (MT), failure of non-operative management (NOM), mortality, and successful non-operative management without intervention, all within 4 hours of emergency department (ED) presentation. METHODS: Children (≤18â¯years) who sustained a BSOI (liver, spleen, and/or kidney) between 2009 and 2018 were identified in the trauma registry at a pediatric level 1 trauma center. Deep learning models were developed using clinical values [vital signs, shock index-pediatric adjusted (SIPA), organ injured, and blood products received], laboratory results [hemoglobin, base deficit, INR, lactate, thromboelastography (TEG)], and imaging findings [focused assessment with sonography in trauma (FAST) and grade of injury on computed tomography scan] from pre-hospital to ED settings for prediction of MT, failure of NOM, mortality, and successful NOM without intervention. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were used to evaluate each model's performance. RESULTS: A total of 477 patients were included, of which 5.7% required MT (27/477), 7.2% failed NOM (34/477), 4.4% died (21/477), and 89.1% had successful NOM (425/477). The accuracy of the models in the validation set was as follows: MT (90.5%), failure of NOM (83.8%), mortality (91.9%), and successful NOM without intervention (90.3%). Serial vital signs, the grade of organ injury, hemoglobin, and positive FAST had low correlations with outcomes. CONCLUSION: Deep learning-based models using a combination of clinical, laboratory and radiographic features can predict the need for emergent intervention (MT, angioembolization, or operative management) and mortality with high accuracy and sensitivity using data available in the first 4 hours of admission. Further research is needed to externally validate and determine the feasibility of prospectively applying this framework to improve care and outcomes. LEVEL OF EVIDENCE: III STUDY TYPE: Retrospective comparative study (Prognosis/Care Management).