Purpose or Case Report: Child abuse is the principal cause of traumatic injury and death in children 0-36 months old. Radiologic findings are a critically important piece of evidence necessary for assessment by a Child Protection Team to assess possible abuse. Current algorithmic strategies for the detection of abuse are sensitive, but compromise specificity. The purpose of this study was to determine the performance of deep learning to diagnose child abuse using Natural Language Processing of the unstructured free-text from Electronic Health Records (EHR), including essential information from Radiology Reports.
Methods & Materials: This retrospective study was approved by our IRB. The Child Protection Team of our Regional Referral Center processed 1123 patients between January 2015 and May 2019, roughly half were diagnosed as victims of abuse. We utilized these patients' EHR including Radiology Reports. Provider notes from the first encounter until the one prior to involvement of the Child Abuse Pediatrics Team (CAP-Team) were processed, enabling an assessment of abuse before CAP-Team involvement. We designed and optimized three natural language processing methods: Rules-based (88 rules), Bag-of-Words, and Word Embeddings, using Convolutional Neural Network approaches, and training with 10-fold cross validation.
Results: The best performing CNN approach was Bag-of-Words, achieving average accuracy of 89.9±2.6%, followed by Rules-based with accuracy 76.6±3.7%. Word Embeddings using two separate corpora had the lowest accuracy of 65.8±2.8% and 66.4± 3.8%. The average ROC-AUC was 93.1±2.2% for the bag of words, and 81.4±5.2% for rules-based. Word embeddings' results were not successful due to the lack of representative embedding. Saliency maps showed that radiologically determined words (negative, normal, fracture, musculoskeletal, head, swelling) were among those highly relevant to the classification.
Conclusions: We successfully applied deep learning networks trained on unstructured free-text; our methodology exhibited high accuracy and ROC-AUC for abuse classification. The Bag of Words algorithm is particularly successful, while word embeddings approaches fail due to a lack of representative embeddings. Radiological findings strongly influence the classification. By restricting the study to pre-CAP-Team involvement, these results are suggestive of a decision support aid that non-child-abuse-specialists may use to determine whether to involve a CAP-Team for the evaluation and management of a specific patient.