Investigating Gender Bias in the Trauma Bay: A Study of Team Responses and Patient Outcomes

Each year, over 2,000 critically injured patients receive specialized care at the U's Trauma Surgery Center. The prompt and effective collaboration of a team of specialists is essential in delivering quality care and minimizing the risk of human error in trauma emergencies. The Team Lead, a trauma surgeon, plays a critical role in the team’s success.
However, as the diversity of trauma surgeons increases, social factors like gender bias may impact team functioning. Research from other industries shows that members’ biases against women in leadership can negatively affect team performance. This study will be the first to investigate whether gender bias plays a role in the trauma bay, potentially compromising team effectiveness and patient care.
Using both behavioral coding and acoustic level analyses, we will analyze recorded footage of trauma bay resuscitations for potential gender biases that hinder team efficiency. Our first aim is to determine if team members’ responses to the Lead vary based on the Lead’s gender, independent of their professional rank and patient details. Our second aim is to assess whether these differential responses have a negative impact on patient care. The project will focus on identifying factors that can negatively impact patient care for future targeted interventions (e.g., trauma medicine training procedures).
This project supports the U’s missions to increase representation of women in trauma medicine and to provide exceptional patient care.

Current Status

2024-02-15
The team has made substantial progress on our methods and analyses. To date, we have analyzed 384 recorded trauma bay resuscitations from two trauma bays in a Level 1 Trauma Center. Three human coders annotated the Trauma Lead’s rank (resident, fellow, or attending physician) and Gender (man or woman).

We used two techniques to measure the quality of communication during the resuscitations. First, video sound files were analyzed to obtain a validated transformation of sound pressure levels recorded by a microphone measurement to closely match the perception of the human ear. This transformation is called “A-weight” and has been validated to quantify hospital noise previously (Busch-Vishniac et al., 2005). A-weight can be thought of as a measure of ambient noise in the human auditory range. To further validate A-weight, three human coders rated each video from Bay 1 on a five-point scale from very quiet to very loud. Indeed, the A-weight was significantly correlated with coders’ ratings of the videos in both bays, r(237) = .54, p < .001.

Next, we transcribed the video communications using OpenAI’s Whisper, an open-source neural net speech recognition system (Radford et al., 2023). Whisper is a powerful program that can produce near-perfect transcriptions when audio quality is high. When audio quality is not ideal, like in the resuscitation videos, Whisper will produce some instability between transcriptions. We transcribed each video eight times to obtain a stable estimate of the number of times “quiet” was uttered, taking the modal number of “quiet” utterances across the transcripts. It provided another validation of the A-weight measure, as noisy videos were more likely to contain “quiet” in the transcripts (e.g., “Everybody, quiet down!”). As expected, the number of times “quiet” was uttered significantly correlated with the A-weight (bay 1, r(238) = .20, p = .002; bay 2, r(142) = .28, p < .001).

Finally, we capitalized on the disagreement between Whisper transcriptions as a measure of communication clarity. We selected two transcripts at random and compared the level of agreement between the two. Agreement can be thought of as a measure of communication clarity.

In the next phase, we will collate patient records with these noise/clarity indices in order to test our research questions.