P0992 - Using Deep Learning and Natural Language Processing Models to Detect Child Physical Abuse

Background/Purpose:
Recognition of child physical abuse can be challenging and often requires a multidisciplinary assessment. Deep learning (DL) models, based on clinical characteristics, laboratory studies, and imaging findings, were developed to facilitate abuse recognition.

Methods:
Level one pediatric trauma center registry data from 1/1/2010-1/31/2020 was queried for abused children and matched participants with non-abusive trauma. Observations were de-identified and divided into training, validation, and holdout sets. Model 1 utilized patient demographics (age, gender, insurance type), clinical characteristics [vital signs, shock index pediatric age-adjusted, Glasgow Coma Score, and laboratory values (lactate, base deficit, INR)]. Patients with missing data were excluded from Model 1. Model 2 included features of Model 1 with natural language processing (NLP) of the radiology reports of head computed tomography, brain MRIs, and skeletal surveys. Google’s latest BERT NLP model, which was pre-trained on a large corpus, was used as the baseline for fine-tuning Model 2. Accuracy, sensitivity, and specificity were used to assess performance.

Results:
Of 1,312 patients, 55.8% were abused. Model 1 (n= 1,174) had an accuracy of 84.1%, sensitivity of 85.7%, and specificity of 82.4% for the validation set (Training set= 959; Validation set= 107; Holdout set= 108). NLP based Model 2 (n= 1,312) had an accuracy of 93.3%, sensitivity 93.8%, and specificity of 92.7% for the validation set (Training set= 1080; Validation set= 120; Holdout set= 112). Most input 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 accuracy. Novel DL and NLP models may facilitate abuse recognition. With additional data and training, these models could be used to improve triage and transfer decisions.