Shaped in 2012, the Parkland Middle for Scientific Innovation (PCCI) is a know-how analysis and improvement group affiliated with Parkland Well being & Hospital System in Dallas. Healthcare Innovation just lately interviewed PCCI President and CEO Steve Miff, Ph.D., about among the highlights of PCCI’s 2026 annual report, which has a deal with predictive analytics and AI.
Healthcare Innovation: Steve, your annual report notes that PCCI has pioneered a novel framework to make sure reliable and sustainable AI improvement, and it now has 14 fashions in manufacturing, seven in testing, one in early exploration, and others underneath improvement. Might you begin by speaking concerning the improvement of that reliable AI framework?
Miff: It has developed over time, however notably as we began to deploy fashions, we observed that it’s one thing that is required not solely on the entrance finish as you construct fashions, but additionally after deployment as you proceed to take care of and help them.
We have recognized 4 key pillars that we consider are vital to offering the required transparency to create belief. One is prediction transparency. The second is efficiency transparency. Third is safety transparency, and forth is compliance.
With prediction transparency, what we have observed is that it is nice to have the ability to predict rising danger and the extent of danger for people, however except you’ll be able to give the small print behind what’s driving the danger, the data is helpful, however not as helpful because it may very well be, by way of giving customers the consolation that what they’re seeing is smart. We develop this know-how known as “Islet” that permits real-time visualization of the data behind a mannequin. With a click on of a button from the digital medical file, you may pop up a window that provides you not solely the present predictive rating, however the historic values. After which dynamically it brings forth the highest 5 elements which might be influencing the prediction essentially the most at that cut-off date. After which it provides you all of the precise information that is feeding into that.
HCI: That is fascinating as a result of we frequently hear from well being system leaders that providing that type of transparency is vital to getting clinician buy-in.
Miff: One other pillar is round compliance. It’s so vital to be sure that any fashions which might be being deployed meet the rigor of the newest compliance necessities. We have been a part of the Well being AI Partnership, one of many founding members with Duke and Mayo and Berkeley. They’ve printed some actually good standards and rubrics about parts that ought to then go into each the compliance on the entrance finish after which a lifecycle administration of AI. We have now recognized a rubric of 20 to 30 totally different parts that we put each single mannequin by means of earlier than is being deployed and evaluated as an internally generated service.
The third pillar is round safety. No matter occurs with the info must be within the safe setting, since you’re managing PHI and managing a number of information sources that want to come back collectively. You will need to spotlight that and consistently take note of it, and have all of the the rigor, the accreditations and all these elements in place.
The final one is round efficiency transparency. The extra fashions we deployed, the extra time we’re spending really monitoring them to be sure that they carry out based on nonetheless they have been designed, how they’re educated, and that they don’t seem to be beginning to deviate. That turns into overwhelmingly time-consuming, and we’re spending extra time on monitoring issues than really being able to develop new issues. So we constructed and are within the course of deploying an AI monitoring dashboard that automates a variety of these statistical capabilities of the fashions which might be being deployed. We’re additionally doing that now for LLMs and ambient listening fashions. You will need to have the ability to create these guardrails of what is anticipated, from a statistical perspective after which be alerted when the mannequin begins to deviate from the parameters that you’ve got recognized.
HCI: Can we stroll by means of among the AI improvements described within the annual report? However first I wished to ask whether or not a few of these improvements may very well be commercialized or exported past Parkland’s use?
Miff: Sure, we design them that manner. We’re not ourselves a business entity, however we’re at all times trying to have the ability to replicate these in different environments. For instance, our trauma mortality mannequin, which is exclusive, is somewhat bit extra area of interest as a result of it applies to Stage 1 trauma facilities and predicts real-time mortality — we’re within the technique of deploying that at Grady Well being in Atlanta.
One other factor we have carried out with a number of entities and well being programs, and even with payers, is the work that we’re doing with our Neighborhood Vulnerability Compass, which is admittedly granular SDOH information, however it’s carried out on the block group degree. We reverse geocode, and attribute to a affected person file their block traits, so now we now have it on 100% of the sufferers with out the necessity to interview them. We simply printed a paper on this in JAMIA and it actually confirmed that it has unimaginable recall charges, not solely on the general index degree, however if you take a look at particular indicators, corresponding to whether or not anyone has meals insecurities or housing instability. It’s wonderful to have the ability to take a block group info attribute to a file after which for that to be so extremely indicative of what that particular person says. We have now 50-plus organizations that now use it.
HCI: I learn that’s being utilized by the United Means of their information capacity-building initiative within the Dallas space.
Miff: Sure, it’s been a six-year journey with them. What I simply completely love about that’s that it is foundational in a number of layers. United Means has been utilizing it for years to trace the influence that their investments within the communities are having, and monitor that 12 months over 12 months.
United Means additionally wished to deliver the group organizations in and improve their information capability. As an alternative of simply saying we’ll pay so that you can have licenses to entry this, they discovered 200 organizations and put them into cohorts that undergo a six-month curriculum to discover ways to apply it to their particular state of affairs. It’s been wonderful to see. That is thrilling, as a result of it’s educating folks use information.
HCI: Let me ask about a few different predictive instruments which might be used within the hospital setting. One is a office security prediction device. Does it display sufferers for the potential of violent interactions?
Miff: That is what it does. As you realize, violence towards frontline workers is actually a pandemic. It is gotten considerably worse after the COVID pandemic, and it continues to be an enormous problem. Many organizations are specializing in attempting to alleviate the issue. This pulls it from a number of sources. This additionally makes use of the Neighborhood Vulnerability Compass information. It even makes use of issues corresponding to smoking standing, earlier involvement with felony justice, or earlier violent occasions. It pulls all this complicated info collectively and mainly predicts the chance that that encounter will end in a violent occasion. It’s important to be very cautious that you just’re not profiling people. You are actually figuring out triggers. This is likely one of the most weak time in our lives, after we’re within the hospital for our personal well being or a cherished one, and also you add all these different issues that compile that anxiousness. For instance, smoking at all times reveals up within the high 10 predicting elements. All hospital campuses are smoke-free. If you’re a heavy smoker and never in a position to smoke, that provides to your stress, and begins to create the next danger.
HCI: You even have a pre-term start prevention program.
Miff: The pre-term start concerned constructing a predictive mannequin underlying elements to establish ladies who’re prone to have a pre-term supply. Initially that program initiated each training to ladies through texting and alerts to their suppliers. Then a broader coalition got here collectively to do extra work on this area, and we’re the analytical engine behind it. We’re utilizing CVC that we mannequin throughout these sufferers to grasp the non-medical boundaries and drivers of well being. We’re modeling with information from a neighborhood supply that is known as the DFW Hospital Basis, the place we now have near 100% of all pregnancies that happen throughout the 2 counties and the related collection of examine issues. So we’re in a position to geocode and mannequin these to grasp the place the very best density of those severe septic issues happen, and what’s the make-up of these neighborhoods. One of many intervention is iron distribution to have the ability to give pregnant ladies iron very early within the being pregnant. We’re utilizing this to establish areas the place the iron distributions happen.
We additionally construct a maternal well being forecasting mannequin. Beforehand we had constructed a diabetes surveillance system, and we’re modeling it after that. The diabetes mannequin predicts deterioration that may require ED visits and hospitalization 12 months out. It is the neighborhood degree, and it provides you each the medical points which might be driving that prediction and the non-medical drivers, and it ranks them, and it’s totally dynamic.
HCI: One other one featured within the annual report is a digital imaging surveillance system that leverages generative AI to establish missed prognosis for follow-up care. We have now written about well being programs which might be attempting to do a greater job of following up on incidental imaging findings. Is that this much like these efforts?
Miff: There are hidden issues within the notes from the radiology report corresponding to incidental findings and Parkland’s been doing this manually for quite a few years. We now use LLMs to scan by means of all these notes and establish these incidental findings. I feel it is wonderful how sturdy the accuracy is — the place it is really extra correct than people doing this.