Abstract
Artificial Intelligence (AI) is actively being integrated into Emergency Medical Services
(EMS). It assists in how providers train and how care is delivered in high pressure environments.
At the core of this shift is data, especially high quality video recordings from training simulations
and real emergency calls. These recordings are valuable because they capture more than just
clinical actions. They provide insight into how teams communicate, make decisions, and respond
under stress. For researchers building machine learning systems, this type of data is essential for
improving situational awareness and clinical decision support.
At the same time, these videos introduce a serious challenge. They capture patients and
providers in vulnerable moments, often with faces, voices, ID badges, and room layouts clearly
visible. This creates a fundamental tension between utility and privacy. The same data that
makes AI systems effective also creates risk if identities are not properly protected. When
protection fails, the consequences directly affect public trust in both the research and the
institutions behind it. Privacy is often treated as something addressed at the end of a pipeline
rather than as a core design decision. This project takes the position that this approach is
insufficient. If people do not trust how their data is handled, then even highly accurate systems
lose legitimacy.
To address this, my technical work focused on building a video anonymization tool
designed to serve as a final verification step in the UVA Link Lab’s de identification pipeline.
Existing systems in the lab used automated methods to detect and blur faces, but they were not
fully reliable and occasionally missed sensitive details. Instead of relying entirely on automation,
I implemented manual verification in the loop approach. The tool, built in Python using OpenCV
and PyAV, allows a reviewer to move through video footage frame by frame and manually mark
regions that need to be anonymized. This ensures that sensitive information is consistently
removed while preserving the data quality required for AI training.
One of the main technical challenges was handling both video and audio streams without
introducing synchronization issues or corrupting files during export. Integrating PyAV into the
pipeline allowed for consistent processing of both streams together. The tool was tested on 12
EMS recordings, totaling about 25 hours of footage. It reliably handled clips longer than 30
minutes, which had previously been difficult to process. These results demonstrate that
combining automated detection with human verification can improve both accuracy and
efficiency in a practical research setting.
Alongside the technical work, my STS research analyzes this problem through the Social
Construction of Technology (SCOT) framework. This approach focuses on how different
stakeholders influence what a technology becomes and how it is evaluated. Engineers tend to
prioritize data quality and model performance, while EMS providers focus on realism and
practical usability. Patients, although less directly involved, have a clear interest in how their
identity and dignity are protected. These priorities do not always align, and anonymization tools
sit at the center of that tension.
A key concept in this analysis is a responsibility vacuum in healthcare AI. Certain steps
in the pipeline, such as verifying that all identifying information has been removed, are critical
but often lack clear ownership. When responsibility is not well defined, gaps can emerge. This
project argues that anonymization tools are not just passive infrastructure but active mechanisms
for assigning responsibility. By requiring human verification and making that step explicit, the
system introduces accountability directly into the workflow.
This becomes even more important as traditional anonymization techniques such as basic
blurring or pixelation become less effective. Advances in re identification methods mean that
incomplete anonymization can still expose individuals. In this context, the design of the
workflow itself becomes a primary safeguard. It is not only about which algorithms are used, but
also about how verification is structured and documented.
Overall, this thesis portfolio shows that building effective AI systems for EMS is both a
technical and sociotechnical challenge, where performance and trust are tightly connected and
cannot be improved in isolation. The anonymization tool demonstrates that it is possible to
preserve data utility while strengthening privacy through structured human oversight, and the
STS analysis highlights how these design choices shape accountability and reflect broader values
in healthcare research. There are still clear opportunities for improvement, including reducing
export times, improving performance in low light EMS settings, and integrating automated
suggestions for regions of interest to better support reviewers. Together, these directions
reinforce the central idea that privacy must be treated as a core design principle, with
responsibility and verification built directly into the system to support a more reliable and
ethically grounded use of AI in emergency medical services.
