Season 1, Episode 8 | Building AI Inside a Cancer Center and the Data Challenges No One Talks About

From Nurse Protocols to Knowledge Graphs: Building Clinically Rigorous AI

70% of patients think they know where to go for care. They're wrong. That's one of the surprising findings from Clearstep, which built its AI triage engine not on a general purpose LLM, but on the same nurse protocols that triage call centers have trusted for decades, the Schmitt Thompson guidelines, turned into knowledge graphs with rigorous unit testing baked in. In this episode, Clearstep CEO and Co-Founder Adeel Malik walks through how they encoded clinical reasoning into a system now trusted by Mount Sinai, HCA, Ochsner, and the U.S. military. He demos the patient-facing triage live, shows a new urgent care use case where AI predicts diagnostic tests before the provider even sees the patient, and breaks down what it actually takes to reroute patients to the right level of care at scale. The conversation ends with his vision for a future where AI has its own NPI, a National Provider Identifier that every doctor in the U.S. gets, and orders lab tests directly, eliminating an entire visit from the patient journey. Guest Bio Adeel is the CEO and co-founder of Clearstep, an AI care navigation company that has partnered with some of the leading healthcare institutions in the U.S. (such as Ochsner, HCA, Tufts Medicine, Novant Health, and others) and has helped millions of Americans access the best care for them. Prior to Clearstep, Adeel was a consultant at Accenture where he worked across pharma, health tech, health systems, and retail. Before Accenture, Adeel also spent 3 years as a neuro-immunology researcher at Johns Hopkins. Links https://www.clearstep.health/ https://www.linkedin.com/in/adeel-malik-3ba345ba/

What Physicians Actually Want From AI

Most healthcare AI tools get built and then don't get used. Why? Because they weren't built around how clinicians actually work. In this episode, Patricia Thaine sits down with Dr. Travis Bias, a family medicine physician and Deputy Chief Medical Officer at Solventum, to get the practitioner's perspective on what's working, what's falling flat, and what builders need to understand about clinical workflows before they ship. They cover ambient documentation and why it's seeing unusually high adoption, the "last mile" integration problem that kills otherwise solid products, how clinical data flow from a physician's note all the way to population health resource planning, the build vs. buy decision for health systems with local technical teams, interoperability and how far it's come and how far it still needs to go, what open banking-like system could look like if applied to health data, and the privacy and values questions that come with patient-facing AI tools like ChatGPT Health. Travis also shares a patient story that cuts to the heart of why context, not just data, determines whether AI is actually useful in clinical care. Guest Bio Travis Bias is a board-certified family medicine physician and Deputy Chief Medical Officer, Health Information Systems, at Solventum (formerly 3M Health). He has 15 years of experience across multiple clinical settings, has taught comparative health systems at the UCSF Institute of Global Health Sciences, and currently practices telemedicine. His work sits at the intersection of clinical medicine, health data infrastructure, and AI governance. Links - Travis on Solventum: https://www.solventum.com/en-us/home/health-information-technology/resources-education/experts/travis-bias-do-mph/ - Travis on LinkedIn: https://www.linkedin.com/in/travisbias/ - Ambient Documentation & Speech Recognition at Solventum: https://www.solventum.com/en-us/home/health-information-technology/ambient-documentation-speech-recognition/ About The Data FrontierHosted by Patricia Thaine, co-founder of Private AI. Each episode brings together builders, engineers, data scientists, product leaders, and domain experts working at the frontier of data innovation. From NLP and privacy to interoperability and AI-driven workflows, we get into the real problems, the real blockers, and what's actually working. 🌐 private-ai.com

Season 1, Episode 8 | Building AI Inside a Cancer Center and the Data Challenges No One Talks About

What does it actually take to build AI systems inside a major cancer center? In this episode, Dr. Anyi Li, Attending Physicist and Chief of Computer Service at Memorial Sloan Kettering Cancer Center, pulls back the curtain on the messy, complex reality of working with clinical data and deploying AI where it matters most. We dig into Woollie, an LLM trained on nearly 40,000 radiology notes to detect cancer progression, and what it unexpectedly revealed about disease pathways. The conversation covers why EHR data are far messier than most people assume, how duplicated clinical notes confuse language models, and why Dr. Li believes the biggest near-term opportunity for AI in healthcare is not diagnosis but reducing the crushing administrative burden on clinicians. Along the way, we unpack the tension between open-source and closed-source models when patient privacy is at stake, whether patient-facing chatbots should be regulated as medical devices (the answer may surprise you), and when fine-tuning your own model may no longer be worth the GPU burn when general-purpose models plus RAG and agents can get you most of the way there. Whether you work in healthcare AI, clinical operations, or data infrastructure, this conversation is packed with hard-won lessons from someone building at the intersection of all three. In this episode, we cover: * How Woollie was trained on radiology impressions to classify cancer progression at 91-92% accuracy * Why clinical notes are full of duplications that break LLM performance * The case for multi-agent architectures to clean and deduplicate patient records * Why ambient listening data and physician-to-physician communications could unlock new clinical insights * The argument that administrative AI is the fastest path to real clinical impact * How to think about FDA regulation for patient-facing vs. physician-facing AI tools * Why general-purpose models with RAG and agents may beat fine-tuned domain models on cost and speed, and vice versa Bio Dr. Anyi Li is an Attending Physicist and Chief of Computer Service in the Department of Medical Physics at Memorial Sloan Kettering Cancer Center. He leads enterprise digital transformation across radiation oncology and medical physics—building the platforms, operating model, and governance required to deploy AI and LLMs safely at scale. Dr. Li directs a multidisciplinary team spanning engineering, data science, physics, and operations, and oversees a portfolio of ~150 clinical systems used by clinicians and staff. His programs reduce administrative friction, improve coordination and throughput, and strengthen patient safety through AI-enabled decision support and workflow intelligence, including timeline analytics and workflow knowledge graphs. Prior to MSK, he spent many years in healthcare IT at IBM Watson Health and Explorys, following early training in theoretical nuclear physics. Relevant Links https://www.mskcc.org/profile/anyi-li [https://www.mskcc.org/profile/anyi-li] https://www.linkedin.com/in/anyili/ [https://www.linkedin.com/in/anyili/] https://pubmed.ncbi.nlm.nih.gov/40604229/

Season 1, Episode 7 | What Happens When Patients Tell Their Own Story

What if patients could complete their medical history before ever walking into the exam room, and actually enjoy doing it? Dr. Chris O'Connor is the CEO of FirstHx and a practicing internal medicine and critical care physician. In this episode, he walks us through how his company built an adaptive interview system that asks patients the right questions, in the right order, based on their answers. No clipboard. No 500-question forms. The results: 95% of patients love it. Completion rates top 90%. And the data they provide are remarkably accurate, matching what's reported in the medical literature. We dig into real outcomes. A provincial virtual care platform cut wait times from six hours to one. A weight loss clinic reduced its waitlist from seven months to one. Clinicians are seeing 10 to 50% efficiency gains depending on the practice. But the bigger story is what happens when you collect standardized, structured data at scale. Chris shows us real-time dashboards tracking heart failure medication adherence, viral outbreaks by geography, and clinical trial eligibility, all generated automatically from patient intake. This is the kind of population-level visibility that's been missing from healthcare. It's possible because the data is standardized and structured from the start, collected before the patient even sees their clinician. Topics covered: - Why adaptive questioning beats static forms - Patient accuracy and the skepticism around self-reported data - Real-time population health analytics from intake data - Screening for clinical trials without adding clinician burden - The cost of unstructured clinical notes and fragmented data

Season 1, Episode 6 | Rebuilding Every Pipeline: AI After an Epic Migration

When Unity Health Toronto migrated to Epic, Dr. Derek Beaton's team faced a familiar challenge: the same measures now had different names, technical feeds changed, and every tool they'd built needed to be reconnected to the new system. Derek Beaton, Director of Advanced Analytics at Unity Health Toronto (UHN), joins Patricia Thaine to talk about what that transition actually looked like. The UHN team has been deploying AI tools into clinical practice since 2017, with over 50 solutions now in production across the network. When the Epic migration happened, they spent months mapping old data to new, building monitoring dashboards to catch mismatches early, and prioritizing which tools to bring back online first based on how much historical data they needed. In this episode, Derek walks through the tools his team has built and rebuilt: ChartWatch, which predicts patient deterioration and has shown a 26% reduction in unanticipated mortality; the EDRN nurse assignment tool, which cut scheduling time from 2.5 hours to 3 minutes; and the ED wait time display that shows up on screens at St. Mike's. We also get into why Derek is skeptical of explainability in clinical AI, what interoperability gaps still slow teams down (scheduling systems, documentation formats), and how his team designs evaluation frameworks from the start of every project. Topics covered: * Migrating to a unified patient record system and reconnecting pipelines * Data monitoring strategies during system transitions * ChartWatch: predicting patient deterioration * EDRN: optimizing nurse assignments in the emergency department * Why explainability is for data scientists, not necessarily clinicians * The interoperability gaps no one talks about (scheduling, documentation) * Building evaluation frameworks that measure real impact About Derek Beaton:Derek leads the Advanced Analytics team in Data Science & Advanced Analytics (DSAA) at Unity Health Toronto. The Advanced Analytics team is comprised of data scientists with exceptional breadth and depth in statistics, machine learning, and artificial intelligence (e.g., evaluation, forecasting, predictive modelling, imaging, workforce analytics). Derek and his team are responsible for developing statistical, machine learning, and AI models that are integrated into clinical and operational workflows at Unity Health Toronto. The advanced analytics team also evaluates and monitors the performance and impact of the tools that DSAA creates. Previously he was a postdoctoral fellow at the Rotman Research Institute at Baycrest Health Sciences and one of the Ontario Neurodegenerative Disease Research Initiative's scholars; his work focused on multivariate methods, multi-modal data, biostatistics, neuroinformatics and translational work in neurodegenerative disorders. Derek has a MS and PhD in Cognition and Neuroscience from The University of Texas at Dallas, where his work was primarily on statistical analyses of big data (e.g., genomics, neuroimaging). Derek is the author of multiple widely used R packages (the ExPosition family of packages), has authored or coauthored over 70 journal publications, and is currently the joint world records holder for the fastest 5 person costumed half marathon and fastest 5 person costumed marathon. Resources & Links: * AI at Unity Health: https://unityhealth.to/about-unity-health/ai-at-unity-health/ [https://unityhealth.to/about-unity-health/ai-at-unity-health/] * DSAA Blog: https://lks-chart.github.io/blog/ [https://lks-chart.github.io/blog/] * Blog Post on Data Monitoring: https://lks-chart.github.io/blog/posts/2025-03-31-love-is-blind-but-data-shouldnt-be-spotting-the-red-flags/ [https://lks-chart.github.io/blog/posts/2025-03-31-love-is-blind-but-data-shouldnt-be-spotting-the-red-flags/] * Empowering Health Leaders for the AI Era: https://unityhealth.to/2025/12/empowering-health-leaders-for-the-ai-era/ [https://unityhealth.to/2025/12/empowering-health-leaders-for-the-ai-era/] * AI Tool Study: https://unityhealth.to/2024/09/ai-tool-study/ [https://unityhealth.to/2024/09/ai-tool-study/] * Unity Health Toronto LinkedIn: https://www.linkedin.com/company/unityhealthtoronto/ [https://www.linkedin.com/company/unityhealthtoronto/] * Derek's Google Scholar: https://scholar.google.ca/citations?user=rw1kVRcAAAAJ&hl=en&oi=ao [https://scholar.google.ca/citations?user=rw1kVRcAAAAJ&hl=en&oi=ao]