Date of Award

January 2022

Document Type


Degree Name

Medical Doctor (MD)



First Advisor

Gunjan Tiyyagura


Background: About 30% of children who suffered serious injuries from abuse were previously evaluated for injuries not recognized to have been due to abuse. Electronic health record (EHR)-based clinical decision support (CDS) has the potential to improve recognition of physical abuse. We developed and validated a natural language processing (NLP) algorithm to identify high-risk injuries suggestive of abuse in emergency department (ED) notes in the EHR that could be linked to a child abuse-specific CDS (CA-CDS).

Objective: To develop and test the usability of an NLP-based CA-CDS to help ED clinicians recognize possible physical abuse in infants who present with high-risk injuries. Methods: A team of stakeholders including experts in child abuse, ED, informatics, and user design developed a prototype that alerted providers to the presence of a high-risk injury via a digital “card” that also linked to a “protocol” with evidence-based questions, actions to consider, and space to document decision-making. The content was customized for medical and nursing providers and also differentiated between initial and subsequent exposure to the alert.

To assess usability and refine the CA-CDS, we conducted semi-structured interviews with 23 general and pediatric ED nurses and medical providers and asked them to think aloud about their impressions as they interacted with the prototype within a model EHR. Interviews were transcribed and coded by the research team using conventional content analysis. Participants also completed the System Usability Scale (SUS), a previously validated tool to assess usability of an electronic system.

Results: From the interviews, 5 main categories emerged:1. CA-CDS Benefits included providing an extra layer of protection, the ability to alert the entire clinical ED team to concerns about possible abuse, customization to provider type, inclusion of evidence-based recommendations, and adaptability despite case-specific variations. 2. User-centered, workflow-compatible design included soft-stop alert configuration, editable and automatic documentation, triggers to consider abuse resulting from multiple providers’ notes, clear presentation of the injury that caused the alert to trigger, attention-grabbing design elements, incorporation of a clear and memorable mnemonic to improve information gathering, and accessible recording of the previous provider’s actions. 3. Recommendations for improvement included clearer design elements, consolidating text, adding a hyperlink that connected to additional resources, adding further information about trigger source, emphasizing crucial history and physical exam features, modifying “protocol” title to emphasize interactivity, and modifications to better reflect provider workflow. 4. Facilitators for completion included reappearance at discharge, stakeholder buy-in, provider education, and sharing the test characteristics of the NLP algorithm that triggers the CA-CDS. 5. Barriers to implementation included concerns about alert fatigue, hesitancy to change, infringement on provider autonomy, and concerns about implications of documentation of child abuse.

The prototype was iteratively refined based on suggestions from the interviewees. Median SUS score was excellent at 80 (IQR: 75-92.5).

Conclusion: With its high usability and user-centered design, our CA-CDS can aid providers in real-time recognition and evaluation of child physical abuse and has the potential to reduce missed cases.


This thesis is restricted to Yale network users only. This thesis is permanently embargoed from public release.