Vega Health, a startup that helps health systems evaluate and implement AI, has partnered with Parkland Center for Clinical Innovation (PCCI) to license its AI model.
Five of PCCI’s AI models are now available on the Vega Health Marketplace and accessible to Vega customers. The model has been validated in a real hospital environment. Most models focus on clinical decision support, population health, or social determinants of health. Vega’s goal is to elevate innovations that might otherwise go unnoticed.
“One of the things we want to do at Vega Health is bring a lot of these use cases to market that are not really standalone companies, but that still have a huge opportunity to improve patient care and improve the health of the population,” Mark Sendak, MD, co-founder and CEO of Vega Health, told Fierce Healthcare.
In 2012, PCCI was spun out of Texas-based Parkland Health, one of the nation’s largest safety-net health systems. The ongoing collaboration between PCCI and Parkland is focused on identifying opportunities in AI and digital health, with a particular focus on the needs of vulnerable populations across North Texas.
The five PCCI models in Vega’s marketplace are:
- Inpatient Sepsis Prediction: Identify patients on the inpatient unit at risk of sepsis within the next 12 hours and identify key clinical factors for each prediction in the EHR.
- ED and Urgent Care Sepsis on Admission (POA): Identify patients who are already septic upon presentation to the ED or Urgent Care and trigger clinical alerts for intervention.
- Parkland Trauma Index of Mortality (PTIM): A predictive model updated hourly to assess the risk of in-hospital mortality in polytrauma patients.
- Patients at Risk for Adverse Drug Events (PARADE): Upon admission, this model stratifies patients by risk of experiencing an adverse drug event during hospitalization, allowing pharmacist intervention.
- Workplace Safety AI Model: This model screens inpatient admissions by identifying encounters that are most likely to proceed without a violent incident, based on EHR data, personnel records, and social needs.
The model tested at Parkland has achieved promising results so far. For example, inpatient sepsis prediction models alert clinicians well before a patient requires antibiotics for early intervention and better outcomes. The model alerted clinicians an average of 19 hours before routine antibiotic administration. This compares to 1.5 hours prior to dosing in current industry models, according to PCCI. Clinicians can snooze alerts if needed.
The trauma index correctly identified 89% of high-risk trauma patients and 92% of low-risk trauma patients. At Parkland, the adverse drug event model prevented more than 2,000 events and avoided more than $17 million in costs. The workplace safety model also accurately predicted 77% of violent incidents within 30 minutes of entry.
Vega was spun out of Duke University. Sendak previously served as Director of Population Health and Data Science at Duke University’s Institute for Health Innovation. The idea was to democratize access to effective clinical AI models built with front-line clinicians. In addition to curating models in the marketplace, Vega helps customers do the work necessary to deploy them in the wild, including evaluation and testing. Workflow integration. Fine-tune each model to unique patient populations. Post-implementation monitoring. This is especially important for hospitals that lack resources, Sendak explained.
“Few organizations have the internal capabilities to build and implement their own tools based on their own patient data,” he said.
“We have no intention of becoming a for-profit organization,” Dr. Steve Miff, president and CEO of PCCI, told Fierce Healthcare. PCCI has a small marketing team and no sales team. “We have been looking for the right partner to help us scale the impact of this work.”
Launched in late 2025, Vega currently operates two regional health systems, including critical access hospitals. The company has revenue sharing agreements with AI supplier partners, which so far include Duke and PCCI. This provides a path to commercialization for innovators.
Just because a model was developed within an academic medical center doesn’t guarantee it’s good, Sendak acknowledged. “You won’t know which model performs better until you test it,” he said. However, the benefit of having a partnership or in-house innovation department is the relationship between developers and clinicians who share responsibility.
In addition to Parkland, PCCI also works with the Dallas County Health Department, payers, and other health systems. PCCI currently has 19 fully deployed AI models. Since 2019, the model has identified approximately 3 million people at risk for intervention.
Health systems interested in leveraging PCCI’s models will work with Vega to evaluate the models based on their own local patient data before implementation. Data will be shared with Vega’s customers and relevant AI partners. If the model performs well, Vega supports clinical implementation and ongoing monitoring to track accuracy, implementation, and real-world outcomes.
If it doesn’t work as expected, Vega doesn’t recommend hospitals buy that particular model, Sendak said. Vega’s goal is not to tell you which model is better. It is also important that the model is trained on a diverse population, as it must be tailored to each institution.
“We’re not trying to be kingmakers,” Sendak said. “We want to help every health system find the model that works best for them.”
Both Sendak and Mif believe that AI has a future in the medical field. “Health care is so complex that no one physician or organization has the best expertise in every clinical area,” Sendak said.
“AI continues to play a huge role in healthcare, and we need it to scale what we do,” echoed Miff. However, he cautioned that administrative use cases are much more scalable and translatable between organizations. Complexities arise when AI is used for clinical decision support and population health management. In that case, models need to be co-developed with clinicians and tested in real-world settings.
“This is the most difficult part, but it also has the potential to have the greatest clinical impact,” Miff says.
