PC2035: Factors Associated with Abusive Head Trauma in Young Children Presenting to Emergency Medical Services Using Machine Learning, and Natural Language Processing

Background:

Background: Abusive head trauma (AHT) is a leading cause of death and disability in young children. Traditionally, the identification of characteristics associated with certain clinical conditions presenting to Emergency Medical Services (EMS) are limited to structured data fields. Machine learning (ML) and natural language processing (NLP) can help enhance detection of high-risk conditions, such as AHT, through the identification of factors not previously found in structured data. The purpose of this study was to apply ML and NLP to free-text narrative documentation in EMS encounters of young children with AHT and describe potentially novel factors associated with this condition.

Methods:

Methods: This is a secondary analysis of the ESO Research Data Collaborative database that includes EMS data fields, free-text narratives, and hospital diagnoses. Eligible patient encounters were those aged < 36 months transported from the scene by EMS to a receiving facility over a 2-year period. We excluded scene deaths, refusals, against medical advice, and aeromedical transports. We developed a NLP pipeline to first extract the patient’s demographics, documented mechanism of injury, and key phrases to exclude ineligible encounters (Figure 1). A vector space model built text representation based on words. Chi-square feature selection was applied to select the top 10th percentile of ranked features to fit into the ML algorithms. We applied five different ML algorithms (Figure 2) with 5-fold cross-validation and compared results on predicting the abusive child head trauma (AHT) cases. We determined the best ML algorithm through the Area Under the Receiver Operator Curve (AUROC). SHapley Additive exPlanations (SHAP) allowed interpretation of the best performing model and the impact of individual words on predicting AHT.

Results:

Results: There were 697 encounters with a hospital diagnosis of Abusive Head Trauma. After chart review, 623 AHT cases met inclusion criteria with 74 encounters excluded for the following: motor vehicle collision (n=43), interfacility transfer (n=15), or age >36 months (n=16). Among the ML algorithms, Random Forest performed the best with AUROC 0.901, specificity 97.4%, and sensitivity 94.4%. (Figure 2) Using SHAP to interpret the trained Random Forest model, words positively correlated with AHT included "blood" (s=0.033), "objects" (s=0.026), "hands" (s=0.022), "fell" (s=0.020), "abrasions" (s=0.018), "avulsion" (s=0.010) and "cut" (s=0.009). Words negatively correlated with AHT included "mva" (s=-0.055), "mvc" (s=-0.045), "transported" (s=-0.029), "fracture" (s=-0.021), "pole" (s=-0.017), "vehicle" (s=-0.015), and "facility" (s=-0.007).

Conclusion:

Conclusion: Machine Learning algorithms have high sensitivity and specificity to detect AHT in EMS free-text narratives. Words associated with physical signs of trauma or a fall mechanism documented by EMS are strongly associated with AHT among preverbal children. Research is needed to evaluate the benefit of real time computerized decision support on identifying sentinel injuries documented by EMS as an aid to detecting AHT in young children.