AI-Powered Health Data Interoperability

Most healthcare AI doesn't stall because models underperform. It stalls because infrastructure is fragmented. We are no longer constrained by algorithmic creativity. We are constrained by data silos, privacy governance, interoperability gaps, compute access, and the operational friction of translating retrospective research into prospective clinical impact. This brief examines this structural bottleneck through the Mayo Clinic Platform. The authors focus on something foundational: building an AI-ready ecosystem designed to accelerate real-world clinical research at scale. The platform provides a secure, cloud-based research environment built on de-identified, standardized EHR data from more than 15 million patients. Key capabilities include: ⭐ OMOP-aligned data models for interoperability ⭐ Structured and unstructured data ⭐ Cohort-building and schema exploration tools ⭐ Integrated workspaces with scalable CPU/GPU infrastructure ⭐ Both no-code and advanced coding environments Unlike traditional institutional repositories, Mayo Clinic Platform enables access for external researchers, supports federated multi-institutional data contributions, and embeds analytics within a privacy-preserving architecture. The paper highlights four applied studies conducted within MCP: 1️⃣ RCT emulation for heart failure drug efficacy using observational data 2️⃣ Validation of antihypertensive medications and reduced dementia risk 3️⃣ Deep learning prediction of mild cognitive impairment progression to Alzheimer’s disease 4️⃣ Neural network prediction of major adverse cardiovascular events after liver transplantation Extracting a cohort of ~15,000 patients took approximately one week. Training and running a deep learning model required roughly 10 minutes on moderate compute resources. When infrastructure friction is minimized, research velocity changes materially. Competitive advantage in healthcare AI is increasingly defined by: 💫 Data harmonization at scale 💫 Federated, privacy-preserving architectures 💫 Reproducible research pipelines 💫 Integrated compute environments 💫 Lower barriers for clinician engagement The authors also point toward multimodal expansion (notes, imaging, genomics), large-scale cross-institutional validation, and “Clinical Trials Beyond Walls” models that broaden participation and diversify real-world evidence. For those shaping AI strategy in health systems, pharma, or digital health, this paper offers a concrete example of production-grade, AI-ready infrastructure. The future of healthcare AI will not be won by isolated models. It will be won by platforms that integrate data, governance, compute, and workflow into a coherent operating system for translational impact. John Halamka, M.D., M.S. and team, great work! #HealthcareAI #HealthSystems #RealWorldEvidence #ClinicalResearch #DigitalHealth #TranslationalMedicine #PrecisionMedicine #HealthData #AIInfrastructure #MedicalInnovation

🧬 We talk about “health data” as if it’s one thing, but it’s really hundreds of incompatible languages trying (and failing) to talk to each other. Every layer speaks a different dialect: • EHRs: HL7 v2, CDA, FHIR • Claims: X12 837, UB-04, CMS-1500 • Labs: LOINC, SNOMED CT • Devices: DICOM, IEEE 11073 • Genomics: VCF, FASTQ, BAM Each was built for a single purpose, not interoperability. The result? 🚑 A patient’s data is scattered across 40+ systems, each with its own schema, timestamps, and access controls. But things are shifting. Newer models are moving beyond formats to: • Graph-based data structures • Semantic layers • Federated architectures These approaches preserve context, not just content, across systems. FHIR paved the road. But the next frontier is semantic interoperability. That’s not just data exchange; it’s data understanding. 🧠 The future of healthcare intelligence isn’t in collecting more data, it’s in connecting meaning. #HealthTech #DataInteroperability #FHIR #HealthcareAI #KnowledgeGraphs #SemanticWeb

AI is only as smart as its data. Bad data breaks everything. Good data builds the future. AI in healthcare is not magic. It is math, logic, and trust—stacked on a backbone of clean, connected data. Here’s the truth: • AI can’t fix broken data. • Automation fails if the data is a mess. • Connected care needs a solid data foundation. Think of data as the bones of a body. If the bones are weak, nothing stands. If the bones are strong, you can build muscle, move fast, and stay healthy. To build smarter AI and real connected care, start with these pillars: 1/ Data Quality:   Garbage in, garbage out.   Every record, every field, every update must be right.   No duplicates. No missing info. No errors.   Clean data is the first rule. 2/ Interoperability:   Systems must talk to each other.   Break down silos.   Use standards like HL7, FHIR, and APIs.   If your data can’t move, your care can’t connect. 3/ Privacy and Security:   Trust is everything.   Encrypt data.   Control access.   Follow HIPAA and GDPR.   Patients own their data—protect it. 4/ Governance:   Set the rules.   Who can see what?   Who can change what?   Audit trails, clear roles, and strong policies keep data safe and useful. 5/ Infrastructure Flexibility:   Cloud, on-prem, or hybrid—pick what fits.   Scale up as you grow.   Don’t get locked in.   Your data backbone must bend, not break. 6/ Continuous Improvement:   Data is never “done.”   Check, clean, and update all the time.   Train your team.   Make data quality a habit, not a project. When you get these right, you unlock: • Smarter automation • Real-time insights • Scalable AI that learns and adapts • Seamless patient care across systems The best AI in the world can’t save bad data. But with the right data backbone, you build care that connects, scales, and lasts. Start with better data. Build the future of healthcare—one clean record at a time.

Rethinking Epic's AI Future: A Call for Protocol-Level Innovation Bill Russell has long been a thoughtful voice in healthcare technology. A leader I respect and continue to learn from. His recent piece raises critical questions that every healthcare executive should be considering. At the heart of Bill’s argument is the Model Context Protocol (#MCP), a rapidly emerging open standard introduced by Anthropic in late 2024. MCP enables structured, secure sharing of organizational context between AI models and enterprise systems. It’s been called the USB-C of AI: a unified interface that simplifies integration across platforms. In practical terms, this means instead of building dozens of custom APIs to connect AI with data sources, organizations can connect once and scale widely. Bill rightly notes that Epic has been highly responsive to customer demand over the years, from Meaningful Use to #TEFCA to interoperability frameworks like #FHIR. But the next wave, context-aware AI, requires more than feature upgrades. It demands an architectural shift. I’m aligned with many of Bill’s observations: • MCP adoption is moving swiftly in other sectors. Financial services and software development are leveraging it to build agentic, context-rich applications. • AI-native platforms in healthcare are emerging, with built-in MCP support, that allow clinicians and analysts to create tools directly tied to real-world workflows. • Most importantly, CIOs are ready. Many are piloting AI solutions today but are held back by integration complexity, security concerns, or vendor constraints. That said, there are key tensions that must be navigated: • Security and governance cannot be compromised. Healthcare data is deeply sensitive, and any new protocol must align with the trust and safety expectations our patients and regulators demand. • Epic’s current closed-loop model was designed with auditability, traceability, and operational control in mind. Reimagining this through the lens of open context exchange requires rigorous oversight, not just technical feasibility. • And while the idea of mass customization is powerful, we must avoid fragmentation. Enabling bottom-up innovation should not come at the cost of standards or shared best practices. I believe Bill’s central thesis is directionally correct: if we want to accelerate safe, meaningful innovation in healthcare, context needs to be a first-class citizen in our AI strategies, and MCP, or something like it, could be the enabler. Epic has the data, the reach, and the infrastructure. What’s needed now is a willingness, from customers and vendor alike; to step into the next layer of interoperability: not just data exchange, but contextual intelligence at scale. As Bill points out, Epic moves when customers move. It’s time we collectively start that conversation. #HealthIT #AIinHealthcare #EpicSystems #MCP #DigitalStrategy #HealthcareLeadership #ClinicalInnovation #CxOInsights

Anthropic just validated every dollar you've spent on FHIR. Last week, one of the world's leading AI companies launched Claude for Healthcare. One detail buried in the announcement deserves attention: they included FHIR development tools as core infrastructure for their healthcare AI platform. The business case for FHIR has traditionally rested on interoperability. The slow, consensus-driven march toward better data sharing between organisations. The ROI is real but difficult to quantify, and it often loses out to more immediate operational priorities. Anthropic's move reveals a second, equally important use case: making data AI-ready. Four implications for Australian healthcare: 🔹 Your FHIR investment is paying a double dividend. Organisations already investing in AU Core and FHIR are building AI-ready foundations, even if that wasn't the original intent. You're further ahead than you think. 🔹 The work through Sparked and HL7 Australia is now doubly valuable. What we've been building for interoperability is precisely the infrastructure that global AI leaders require. Our standards work has become AI infrastructure work. 🔹 FHIR compliance is becoming table stakes. Organisations without structured, standardised data will be locked out of the most powerful clinical AI tools emerging from companies like Anthropic and OpenAI. 🔹 The "you go first" problem is dissolving. For years, interoperability stalled because the value for one hospital depended on every other clinic adopting the standard. AI flips this dynamic. Organisations now need FHIR for their own internal AI ambitions, and external data sharing becomes a bonus. The implications for Australia's national strategy go deeper. The global market is accelerating toward an API-first model where AI agents retrieve and harmonise data on demand. Tools like ChatGPT Health are entering the market. AI agents are becoming the primary interface between patients and their data. → Do we codify a formal "Right to API Access" into legislation, beyond the current right to download a PDF? → How do we accelerate AU Core compatibility with international SMART on FHIR standards while managing the complexity of federated governance? → Where should government investment prioritise: building better portals, or building the plumbing that enables a connected ecosystem? I don't claim to have all the answers. But I believe the conversation needs to happen now, before the velocity gap becomes unbridgeable. Where does your digital health strategy sit in a world where AI agents become the primary interface? And more importantly, what role do you believe government should play in that world? I'm genuinely curious to hear your perspective, particularly if you disagree. #DigitalHealth #FHIR #AIinHealthcare #Interoperability #HealthIT #AustralianHealthcare #DataStrategy #HealthcareLeadership #Sparked #HL7Australia

**Making FHIR more ‘interoperable’ with AI training** FHIR is supposed to be healthcare’s common data model—yet every AI project I’ve touched still starts with days of flattening nested JSON and unraveling pesky arrays 🤯 This new spec from John Grimes and the SQL-on-FHIR working group caught my eye: https://lnkd.in/eQ7iWM_6 They introduce FHIR view definitions—blueprints that pull just the fields you need into a tidy table, no matter where your data lives. What this unlocks: - LLMs that query FHIR natively. With tables on tap, an agent can ask, “Show me ICU patients with rising creatinine,” and run the SQL behind the scenes using more robust tooling - Tabular FHIR → better FHIR foundation models. Eg, pairing these views with some of the latest tabular foundation architectures (like this recent Nature paper: https://lnkd.in/e2pvixb2) EHR AI can only progress as fast as the underlying data and standards. Formalising the “flatten-then-query” step is more than a convenience—it’s the prerequisite for scalable, trustworthy AI.

A modern AI integrated Electronic Medical Record (EMR) should evolve from a passive data repository into an intelligent, workflow embedded clinical partner that enhances decision making without adding cognitive burden. At its core, such an EMR must unify longitudinal patient data, clinical notes, labs, imaging, genomics, wearable streams, and social determinants into a dynamic, continuously updated patient timeline, supported by interoperable standards like HL7 and FHIR. AI capabilities should be seamlessly integrated at the point of care: ambient voice documentation that converts clinician patient conversations into structured notes, predictive analytics that flag deterioration risks or suggest differential diagnoses, and context aware clinical decision support systems (CDSS) that provide evidence based recommendations tailored to the patient’s profile rather than generic alerts. The interface should be intuitive and adaptive, prioritizing relevant information based on clinical context, specialty, and user behavior, thereby reducing alert fatigue and documentation overload. Importantly, explainable AI must be embedded to ensure transparency and trust, allowing clinicians to understand the rationale behind recommendations. A modern EMR should also support bidirectional patient engagement through portals and mobile apps, enabling patients to contribute real world data and participate actively in care. From an operational standpoint, it should incorporate AI driven automation for coding, billing, and workflow optimization, while maintaining strict data governance, privacy, and security frameworks. Ultimately, the defining feature of such a system is its ability to transform raw data into actionable, personalized insights in real time shifting healthcare from reactive documentation to proactive, intelligence driven care delivery.

AI turned health data into an active clinical tool. But we’re still regulating, and engineering, it like it’s trapped in a manila folder from 1996 waiting to be stolen. We built HIPAA-era policy on a simple belief: “more data sharing = more danger.” Back then, it made sense. Data barely moved, rarely integrated, and almost never helped in real time. The risks were concrete. The upside was theoretical. In an AI-driven system, that mindset breaks down. The biggest failures in care, cost, and equity now come from data we can’t use. Clinicians practice in a world where data acts, predicting risk, surfacing conditions, exposing inequities we’ve never been able to see. And here’s the real problem: Our privacy frameworks and our interoperability infrastructure come from the same outdated worldview, one that treats data as a liability to contain, not a clinical input to activate. TEFCA reduces fragmentation, but it’s still document exchange, not the responsible, governed data layer modern care depends on. If we want AI to improve outcomes, leaders must push EMRs toward real interoperability: open APIs, real-time exchange, actual data liquidity. The risk today isn’t sharing too much. It’s sharing too little.

Our health system still spends too much time moving and cleaning data across systems that weren’t designed to work together. That fragmentation slows providers, delays care, and limits our ability to deliver truly coordinated treatment. At Elevance Health we built Health OS to change that. It’s a bi-directional clinical data interoperability platform that securely connects systems and standardizes data—making it accessible, actionable, and AI-ready with privacy and security at the core. With AI and digital technologies, guided by human oversight, we’re replacing repetitive, disconnected work with intelligent systems that anticipate needs, automate routine tasks, and help care teams act faster. In the article below, Jeff Plante and I share how Health OS enables seamless information flow across providers, health plans, and member experiences—supporting earlier intervention, better coordination, and more proactive care at the right time. https://lnkd.in/gqx3UFfd