Using supervised machine learning and ICD10 to identify non-accidental trauma in pediatric trauma patients in the Maryland Health Services Cost Review Commission dataset.

Background: Identifying non-accidental trauma (NAT) in pediatric trauma patients is challenging. We developed a machine learning model that uses demographic characteristics and ICD10 codes to detect the first diagnosis of NAT.

Methods: We analyzed data from the Maryland Health Services Cost Review Commission (2015-2020) for patients aged 0-19 years. Relevant ICD10 codes associated with NAT and trauma were identified. Health records preceding the patients' first trauma diagnosis were analyzed. Random forest models were built using covariates selected through penalized regularization. Models were developed for confirmed and suspected NAT. Data was divided into 80/20 split for model training and testing. We conducted analysis in R.

Results: We analyzed 128,351 non-NAT trauma patients, 522 confirmed NAT patients, and 2128 suspected NAT patients totaling 364,217 encounters. Variable selection identified 55 covariates for confirmed NAT and 65 for suspected NAT for model development. These covariates were primarily musculoskeletal injuries of the head and extremities. Model testing results are summarized in Table 1.

Conclusion: Our study uses machine learning to identify NAT within the pediatric trauma cohort. Analyzing ICD10 categories before the first traumatic diagnosis may allow for earlier detection of NAT. Additional research in building learning models with ICD10 codes is needed to better understand how clinician and billing biases may impact predictive models. Supervised machine learning can potentially augment clinical decision-making and enhance pediatric trauma care.