Vega Health has licensed two sepsis prediction models from the Parkland Center for Clinical Innovation. Read about the partnership.
The Parkland Center for Clinical Innovation (PCCI) has a long history of innovation on behalf of Dallas County's underserved patients.
PCCI spun off from Parkland Health (Parkland), the Dallas-based safety-net hospital system, in 2012 with a mission to develop AI-driven systems that combine data science, clinical expertise, and social determinants data to support the most vulnerable community members.
Cutting-edge AI models are rarely built to serve safety-net populations, yet PCCI has been building AI models and digital health tools for the diverse communities of Dallas, which includes a high proportion of Spanish speakers, for nearly 15 years.
Now, through a licensing agreement with Vega Health, several of PCCI's most consequential innovations are poised to reach far beyond Dallas County.
Parkland Health and PCCI
Every year, Parkland serves hundreds of thousands of patients across Dallas County, many of whom have nowhere else to go. More than 40% of Parkland’s payor mix is made up of charity and self-pay patients, and in 2025 alone the health system provided $1.4 billion in uncompensated care. PCCI, founded as a department in Parkland that spun out to an independent non-profit in 2012, has developed industry-leading expertise in data science and non-medical drivers of health (NMDOH) and is a driving force in the application of innovative, responsible, and ethical artificial intelligence to improve lives. In just the past six years, PCCI has deployed 19 advanced AI models that have generated over 35 million predictions, touched more than 6.2 million patients, and identified 2.8 million high-risk individuals.
"We operate and act like a technology startup, but we're set up as a nonprofit, so we have the opportunity to try these innovations with a population that people don't normally think about," Steve Miff, PhD, President and CEO of PCCI, said in a podcast interview with Outcomes Rocket.
The PCCI’ AI models for use in clinical care include its clinical decision support (CDS) models that focus on one diagnosis or disease state and predict whether the patient will have high risk for the condition in the inpatient or outpatient setting. PCCI has developed CDS models for sepsis, trauma mortality, and HIV infection risk, implementing EMR alerting on the patient's risk of developing the conditions so physicians can intervene early and create better outcomes.
PCCI also develops innovative population health models, predicting the risk of chronic conditions that disproportionately affect certain demographic groups. Among these are pediatric asthma, diabetes, and hypertension risk models, and more recently, a maternal health index that identifies areas where mothers and infants are more likely to have a more severe diagnosis.
PCCI’s inpatient sepsis prediction model
Like many health systems across the country, identifying and treating sepsis has been a system-wide priority. Nationally, sepsis is the leading cause of death of hospitalized patients, and Parkland and PCCI teamed to find a solution.
The inpatient sepsis prediction model, now available through the Vega Health Platform, has been in operation for almost five years at Parkland. The model is predicting sepsis on average 18 to 24 hours before IV antibiotic administration.
PCCI’s AI models benefit from a cyclical process of feedback from the end users in the silent trial phase of model testing. The silent go-live was a key period for PCCI developers to learn how to better tune the model. During this phase, the developers took the accuracy from 33% to 50% based on clinician feedback, which in turn developed trust among the physician and nurse care teams.
"That actually improved the trust between the model and the providers, and that feedback from the stakeholders actually helped us better understand [their needs]," Yusuf Tamer, Principal Data Scientist at PCCI, said.
The physicians also wanted to better understand how the model was making its predictions. “Physicians are scientists too,” Tamer noted, and they didn’t just want to know who might develop sepsis, but why.
So, the developers created an alert in the electronic medical record that gave physicians the top five reasons the model was making the prediction. After even more requests for explainability, Tamer’s team created ISLET, a predictive model visualization tool that gave providers a comprehensive view of the patient's sepsis severity score and trends in their vitals.
"We opened the machine learning model up for the care providers so they can go through the list to see what has changed over time and why the score has happened," Tamer explained. "So that final dashboarding solution really put us in the best state with the care providers. Now we have good trust between us on our model."
Screening for sepsis in the emergency department
In September 2025, PCCI added a second front: a sepsis model designed to catch the infection at ED arrival. The challenge was steep. Parkland had the highest number of emergency department visits of any hospital in the country in 2025. Not only is the patient volume high, but sepsis is hard to diagnose. Its symptoms overlap with dozens of other conditions, and treatments vary widely. Missing the case often means the patient will deteriorate rapidly.
"Aspects that challenge early care include competing ED diagnoses and care, varying levels of evidence for sepsis recommendations, and treating patients with unnecessary therapy when they ultimately have diagnoses other than sepsis," a 2021 study noted. Outcomes improve with early intervention, which means the model has to move faster than the disease, and faster than the physician's diagnosis.
This model solves a fundamentally different problem than the inpatient model. In the hospital, PCCI's model monitors lab values over days, watching for the slow drift toward infection. In the ED, that runway doesn't exist. The model must work with what's available on arrival. It uses vital signs rather than lab results, which can take hours to return. It factors in medical history, including prior sepsis or chronic conditions that raise infection risk and data from previous interactions with the health system.
"The model itself is racing to understand any subtle changes that take the patient toward high risk for sepsis," Tamer said.
The PCCI team built a two-filter system to manage the complexity. The first identifies patients likely to need IV antibiotics for an infection. The second narrows that group further to those where sepsis, specifically, is the probable cause. The model runs for 24 hours — the window in which the physician must decide whether to admit.
PCCI’s challenge was Parkland's extraordinary volume, but the sepsis problem it faces in the ED is not unique. Any hospital with a busy emergency department — community hospitals, regional medical centers, other safety-net systems — encounters the same diagnostic pressure: a patient arrives, the clock starts, and the symptoms alone don't tell the full story. What makes the ED model valuable beyond Dallas County is precisely what made it hard to build. It was stress-tested in one of the most demanding clinical environments in the country, refined through real physician feedback, and designed to work with limited data under time pressure.
For health systems that lack the resources to build that kind of tool from scratch, access to a model with that pedigree could meaningfully change how quickly and accurately their clinicians identify sepsis from the moment a patient comes in.
What both models share is a foundation of hard-won clinical trust, built through transparency, iteration, and a genuine feedback loop between developers and care providers.
Vega Health wants the best AI innovations, built for a wide range of patients, available on our Marketplace. That rigor is exactly what Vega Health is licensing. For health systems that serve patients like Parkland's, these models carry the credibility of one of the highest volume hospitals in the country and over a decade of PCCI's work on behalf of underserved populations.
