Using deep learning and natural language processing models to detect child physical abuse.

Area of Science:

• Medical informatics
• Artificial intelligence in healthcare
• Pediatric medicine

Background:

• Child physical abuse identification is challenging, often requiring multidisciplinary assessments.
• Deep learning models offer potential for unbiased identification of abused children.
• Models utilize clinical characteristics, laboratory results, and imaging findings.

Purpose of the Study:

• To develop and evaluate deep learning models for distinguishing child physical abuse from non-abusive trauma.
• To assess the impact of Natural Language Processing (NLP) on model performance.

Main Methods:

• A retrospective analysis of pediatric trauma registry data (2010-2020).
• Two models were developed: Model 1 (demographics, clinical data) and Model 2 (Model 1 features + NLP analysis of radiology reports).
• Google's BERT NLP model was fine-tuned for Model 2.

Main Results:

• Deep learning Model 1 achieved 86.3% accuracy, with an ROC AUC of 0.86.
• NLP-enhanced Model 2 demonstrated superior performance with 93.4% accuracy and an ROC AUC of 0.94.
• Individual clinical features showed weak correlations with abuse.

Conclusions:

• Deep learning models effectively differentiate child physical abuse from non-abuse.
• NLP significantly improves the accuracy of these diagnostic models.
• Real-time EMR integration could alert clinicians, prompting abuse diagnosis.