In 2024, eight out of ten healthcare professionals reported experiencing workplace violence in the preceding year. The CDC's Quality of Worklife Survey has tracked the rise in the trend since the COVID-19 pandemic, linking distressing work environments to burnout and poor mental health among healthcare workers. The violence itself takes many forms including harassment, racism, as well as verbal, physical, and sexual abuse.
Last month, Vega Health announced a partnership with the Parkland Center for Clinical Innovation (PCCI) to distribute five of their AI models nationwide.We prioritized a specific model that addresses the rise in workplace safety events for frontline healthcare workers.
Hospitals have sought to address workplace safety in a variety of ways, such as adding personal alarms and video surveillance to patient rooms. While both may help a staff member call for help, Parkland Health, Dallas’ largest safety-net hospital, turned to PCCI to address the problem upstream.
For PCCI's data scientists, this model was a significant departure from the 18 others they've built, most of which draw on vital signs or lab results. The Workplace Safety Model instead pulls from a much more varied set of inputs, including the hospital's behavioral event reporting system.
The model works by identifying the 70% of patients unlikely to pose a safety risk, freeing the health system to direct additional resources toward the 30% who present a higher risk during their encounter.
“I think one of my take homes from this project is that you can identify patients that are low risk with reasonable accuracy, which makes this an incredibly useful tool in the tool belt,” Alex Treacher, principal data scientist at PCCI, said.
To help keep workers safe, the model flags high-risk patients within 15 minutes of admission, triggering a survey that prompts the nurse to assess the patient's risk level. From there, the care team determines what safety measures to put in place.
“The health system could determine that they should take extra staff in [to the room] or put signage on the patient's door so that dietary staff and others that don't have access to the medical record are aware of the risk of this patient,” Dr. Jacqueline Naeem, VP of Clinical and Social Health at PCCI, said. "It's a strong model in that it's very flexible to provider’s needs, which is one of its great advantages.”
Unlike predicting an adverse drug reaction or clinical deterioration, forecasting patient behavior is far less clear-cut. PCCI's data scientists had to comb through the health system’s data, combining several data streams to identify patterns associated with a lower risk of violence.
“It was really exciting to be able to dig into the data and work with Alex and team and figure out what those pieces or those patterns are,” said Naeem. “What can we identify in the patient charts and records and information that could give us a clue that perhaps there's something we can change in the course of their inpatient stay to stop an event from happening.”
While the data scientists started by looking at relatively static data points, they realized that a more predictive approach was possible by identifying not just who the patient is on paper, but the situation they’re in during that specific encounter.
Information about patients’ race, ethnicity, religion, and primary language were excluded from the model to prevent bias. These factors also did not reliably predict which patients were violent.
“We started off trying to build a predictive model to identify patients that were at low risk of having a violent incident, using patient-specific information only,” Treacher said. “This means that on some of the early iterations of the model, if the patient came back 10 times, the model would have made the same prediction. We quickly found out that while some of the patient information is useful, more useful for prediction is the situation the patient is in, meaning what type of encounter, what department, what's happening to them, and how did they arrive?"
Treacher noted that stress, confusion, and pain are key drivers of safety incidents. Arrival by ambulance or police car (rather than personal vehicle) and admission through the ED both signal elevated stress and urgency. Because Parkland Health has a tobacco-free campus, a patient with habitual tobacco use might pose a higher risk of agitation. A history of dementia is also correlated with higher risk.
The model was piloted in a few Parkland units, then integrated into most medical and surgical units in July 2025. Notably, the team did not implement the model in the psychiatry unit or medium security unit, which have different training and protocols for de-escalating patients.
When asked how the model has been functioning since its integration in the live clinical setting, Naeem was honest that there were some hurdles to adoption.
“Nurses are very, very busy. So, when you're asking them to do an additional task, we had to demonstrate that this would be helpful overall for them versus just another item on their checklist,” Naeem explained. “But we had really strong support from nursing leadership, and they conducted a number of training sessions, leading to improvements in buy-in over time.”
Naeem noted that other Vega Health partners that seek to use the model in their hospitals need a way of collecting workplace safety data. She strongly suggested having frontline clinician buy-in before implementing the model, which is baked into Vega Health’s approach to partnership.
She also noted that the model is not a panacea for healthcare workers’ safety.
“It's not going to solve everything on its own,” Naeem said. “But I think it could be an effective way to augment existing programs. If this is the first step in those interventions, that would work too.”
