244: Bergen New Bridge's CHIO Dr. Anthony Rosania: How Safety Nets Use AI Safely, What Informatics Can Learn From Pharmacists on ROI, and Why a Zero Denial Rate Means You're Losing Money

Subscribe on: RSS | SPOTIFY | APPLE PODCAST | GOOGLE | BREAKER | ANCHOR

On this episode of The Digital Patient, Dr. Joshua Liu, Co-founder & CEO of SeamlessMD, and colleague, Alan Sardana, chat with Anthony Rosania, MD, MHA, Chief Health Information Officer at Bergen New Bridge Medical Center, about "How Safety Nets Use AI Safely, What Informatics Can Learn From Pharmacists on ROI, Why a Zero Denial Rate Means You're Losing Money, and more..."

Audio:

Guest(s):

• Anthony Rosania, MD, MHA, Chief Health Information Officer at Bergen New Bridge Medical Center
• Joshua Liu, MD (@joshuapliu), Co-founder & CEO at SeamlessMD

Episode 244 - Show Notes:

[00:06:42] How an operations background shapes his approach as CHIO — Anthony built his informatics credibility in emergency medicine operations (ED observation, sepsis order sets, quality, safety, and diagnostic error) rather than through the traditional fellowship path. He sees it as a strength: it helps him communicate with operations, get into the weeds on issues like billing and metric definitions, and see the health system holistically. "When you think about informatics in a silo, then you can't do good informatics that way. Really, it is a team sport." To make up for not having a fellowship's academic rigor, he reads journals constantly.

[00:08:38] Why informatics is about the human experience with technology, not the technology itself — while welcoming the growing number of informaticists with computer science and engineering backgrounds, Anthony teaches residents and medical students that informatics is fundamentally about the human-computer interface. He is often the one in the room to point out that the EHR or a given technology "one, is not your issue, and two, not your solution," and to propose workflow solutions instead of anchoring on a technological fix.

[00:10:02] Why he goes by "Anthony" instead of "Dr. Rosania" — drawing on Amy Edmondson's Teaming and its focus on psychological safety and flat teams, Anthony explains that in fast-moving work like an EHR implementation, any barrier to communication can set a project back weeks if someone doesn't feel comfortable raising an issue. He wants team members to see him as part of the team, not an executive, so they bring problems to him right away rather than acting out of fear. Still, authority gradients never fully go away: "you're always gonna have something that doesn't get escalated to you that should have."

[00:12:07] The three ingredients for earning candid feedback — Anthony rounds daily in the project team's offices, uses his experience as a physician builder to share "life hacks" that make him another user of the same tools, and shows vulnerability by owning and apologizing for his mistakes. He describes vulnerability, relatability, and time as the three ingredients, and stresses that none works on its own: life hacks without daily presence don't build trust, and daily visits without relatability don't either.

[00:15:57] How AI strategy differs in a safety net, low-resource environment — Anthony, who has spent his whole career in academic and community safety nets, says low-resource environments need to stick with fundamentals and avoid casting too broad a net. That means an Epic-first approach, not because other tools lack value, but because organizations need to start with the tools they've already paid for. Epic's expanded investment in AI has made it easier for safety nets to engage in this work, and constrained environments are often better at asking the essential question: what problem does this solve?

[00:18:54] How constraints push safety nets to use frontier AI tools creatively — responding to Josh's example of using general-purpose LLMs to simplify and translate patient instructions, Anthony notes that such instructions aren't protected health information, so no specialized tool is needed. Being in a safety net forces teams to be scrappy, creative, and to better understand regulatory barriers, asking when out-of-the-box frontier models can be used safely for tasks like asking clinical questions and translating instructions.

[00:20:48] What useful AI governance actually looks like — Anthony warns against "weaponized governance," where governance becomes a way to get to no or avoid dealing with a problem. Drawing on Lorien Pratt's work on decision intelligence and her book Link, he argues that good governance starts by mapping a process, understanding its problems, and only then looking for high-leverage points for technology, including AI. "We build a building with purpose, and we need to deploy AI with purpose too."

[00:23:34] How to monitor AI tools with limited resources — once you've defined the problem a tool is meant to solve, Anthony says you can monitor it with the same operational metrics used for anything else: hospital readmissions for a readmission tool, or time to antibiotics and repeat lactates for sepsis. Monitoring where data goes and how tools are used falls to his CIO dyad partner and the security team; Anthony usually advocates for more AI access but respects their role in protecting the institution.

[00:25:23] Why automation bias, not AI itself, is what keeps him up at night — Anthony worries about physicians using AI in the same way he worries about them using a CAT scanner. Medical education gives biostatistics "lip service," and many physicians treat a lab or imaging result as an answer rather than a probabilistic finding with false positives and false negatives. AI isn't breaking something so much as revealing what's already inadequate: "We are not training a physician workforce that has the foundational knowledge necessary to utilize artificial intelligence in the care of patients." The fix is more education in biostatistics and the probabilistic nature of diagnostics, not just how large language models work.

[00:28:51] How overtesting and false positives compound the AI problem — radiologists can misread images, scanners can miss findings because of resolution, and mistimed contrast can create artifacts read as findings. Testing people who don't have a disease produces false positives, and Anthony says this happens often because patients expect it, clinicians fear being sued, and medical students aren't taught how to avoid that trap. Without that foundation, understanding AI becomes even harder.

[00:30:50] Why statistics education should start long before medical school — Anthony suggests adding statistics and calculus to the MCAT and going back even further to high school, where "stats is just the forgotten math." Beyond science and health, he argues a population with a better understanding of statistics would handle things like political polls better: "we can make the world better with statistics."

[00:32:33] Why observation medicine should not be a denial management tool — when health systems overreact to denials, Anthony says, they overcorrect and observation becomes a self-denial, accepting lower payment for patients who justified a full admission. Like a screening test that finds no disease, a zero denial rate signals that something is wrong. His math: if you admit four patients and get paid for only one, it's the same as putting all four on observation, so a 75% denial rate is a wash and a 50% denial rate means you're making money. The answer is to do the right thing for patients, then "fight like hell to get paid for doing the right thing for patients."

[00:35:57] How the word "observation" creates costly ambiguity — "observation" refers to a clinical specialty at the interface of the emergency department and the hospital, a payment classification, and simply the act of watching a patient. Anthony says the EHR should make it easier to separate location from class, since being on an observation unit doesn't mean a patient is observation class. He has advocated for calling the unit a "clinical decision unit" to remove the term observation from the language of the work and the place.

[00:38:41] Why clinicians need benchmarked data, not letters about single cases — physicians receive letters when an admission should have been observation, but never one saying an admission was justified and paid. Anthony argues the EHR should give providers information that reflects their behavior and benchmarks them internally and against each other, because the right denial rate "is not zero, and it's probably not 50." A hospital with a denial rate under 5% is losing money, since the denial rate is only an interval metric; the end metric is revenue. He notes some for-profit hospitals skip observation entirely, relying on clear admission criteria, strong documentation, and fighting for payment.

[00:41:31] What happens when revenue cycle removes the human in the loop — Anthony notes that revenue cycle is one of the areas where people are becoming comfortable with no human in the loop. "At some point it really will just be two AIs arguing with each other over whether or not the care should get paid for," which, he adds, means "it's really just one AI."

[00:42:37] How informatics leaders can make the case for their value — Anthony describes informatics as a loss leader; because he doesn't generate obvious revenue, he's "the end of the disposable income chain" in the C-suite. Beyond the work AMIA has done, he looks to pharmacists, who have assigned dollar values to medication loss prevention and reductions in medication administration errors. For informatics, that means quantifying soft but measurable outcomes like medication errors, patient satisfaction, wait times, time spent in the EHR, reduced alert fatigue, and length of stay, since there's no RVU for building an order set.

[00:45:45] The two things Anthony hopes listeners take away — first, understand the problem an AI tool is solving and implement it at the highest-leverage point in the workflow. Second, educate the workforce beyond basics like "don't put PHI into Claude tomorrow," helping clinicians understand the probabilistic nature of these tools. "I really do fear that it's just gonna become another answer machine, and that's where a lot of the risk to our patients lies."

Fast 5 Lightning Round:

1. What is your favorite book, or the book you've gifted the most?

Ego Is the Enemy by Ryan Holiday

2. If you could instantly master any skill, what would it be?

"Coding and data science"

3. Would you rather have super strength, super speed, or the ability to read people's minds?

"Mind reading"

4. What is something in healthcare you believe that others might find insane?

"Half my med school class never went to class, and a positive test doesn't mean you have the disease"

5. What is the last movie or TV show you saw, and what did you think of it?

"The Pitt. Definitely recommend."

The Digital Patient has been recognized as Feedspot's #1 Patient Engagement Podcast of 2025. Thank you to our listeners for making this happen!