Interoperability is the riverbed that determines the course of healthcare AI initiatives, ultimately dictating their success or failure. Imagine a reality where healthcare data flows as effortlessly as a river, each droplet representing trusted information that moves seamlessly, just like a conversation between colleagues. In this vision, every piece of medical data, from a healthcare consumer’s historical records to real-time diagnostic insights, can be instantly connected, understood, analyzed, and acted upon in real-time.
At the ViVE 2025 conference, Imran Qureshi, Chief AI and Technology Officer at b.well Connected Health, delivered a provocative message that cut through the AI hype. “Without Interoperability, There Is No AI,” he declared, challenging healthcare leaders to look beyond technological buzzwords and focus on the essential challenge of data integration. This idea was broken down into three categories of interoperability: data and AI, people and AI, and the interoperability of workflows and AI. This article examines the intricacies of data and AI interoperability.
Lesson 1: FHIR is the Best Format for LLMs
Healthcare is drowning in data. Electronic health records, claims databases, and clinical systems generate unprecedented volumes of information. But volume doesn’t translate to value. The real challenge lies in ensuring all of these data points are connected and transformed into meaningful, actionable insights that can genuinely improve individual care.
The Fast Healthcare Interoperability Resources (FHIR) standard is a critical bridge between fragmented healthcare systems. FHIR provides a common language that allows disparate systems to communicate effectively, reducing the long-standing friction in healthcare data management, making it the best format for LLMs.
In addition, LLMs are better at FHIR than proprietary formats because they have already been trained on FHIR specifications. They can document data better than relational, normalized data.
Lesson 2: Higher Quality Data Matters More Than Choice of LLM for Healthcare AI
In reality, it doesn’t matter what LLM or new AI tool you’re using. What matters most is the data that powers them. For example, at b.well, our health data management foundation is designed to aggregate, clean, and normalize health records across systems. However, we consistently produced similar results across top-tier AI platforms. The reason? High-quality data. Most healthcare organizations often focus on the wrong priority. Instead of endlessly comparing and switching between the latest AI technologies, leaders should concentrate their efforts on something far more impactful, improving the comprehensiveness and quality of their health data.
The message is to invest your time and resources in creating more comprehensive, cleaner, and higher-quality longitudinal health records. By doing so, you’ll achieve far more meaningful improvements in healthcare AI performance and insights than you ever would by simply chasing the newest language model.
Lesson 3: Use AI Agents instead of RAG
The challenge of processing healthcare data is fundamentally a matter of scale and precision. A typical longitudinal health record spans approximately 40 million tokens, yet most large language models can only process 200,000 tokens. This limitation creates a critical bottleneck in healthcare AI applications.
Traditional data retrieval methods prove inadequate when it comes to healthcare data. The solution lies in AI agents—intelligent systems that function like specialized medical consultants. These agents can selectively retrieve only the most relevant information for a specific query, addressing two critical challenges:
- Computational Efficiency: By retrieving just the necessary tokens, AI agents dramatically reduce processing costs. While processing a full 40-million-token record could cost around $32, a targeted query might only incur a fraction of a cent.
- Hallucination Management: The probability of AI hallucinations increases proportionally with the number of tokens in the context window. By carefully selecting and limiting the retrieved information, AI agents significantly mitigate this risk.
Hallucinations are not a “bug” but an inherent characteristic of complex information systems. The goal is not to eliminate uncertainty but to manage it intelligently. Our approach prioritizes curated, authoritative data sources, strict instruction parameters, and robust query moderation mechanisms. At b.well we use an eight-step AI safety framework to evaluate any AI features.
By treating AI as a nuanced tool, much like medical professionals approach clinical uncertainties, we can harness its full potential while maintaining rigorous standards of accuracy and reliability. The key is not to process everything, but to process precisely what matters.
Lesson 4: Use Feedback from Internal People and Then from External People
LLMs are remarkably sensitive to few-shot examples and fine-tuning, presenting a powerful opportunity for optimizing healthcare AI. The most valuable training data comes from those who understand healthcare’s nuances best—your internal team.
By capturing structured feedback from internal clinical staff and using it as few-shot examples, organizations can significantly improve the quality of AI responses. This method does more than enhance technical performance; it builds confidence among clinical professionals by directly incorporating their expertise into the AI’s learning process.
Successful healthcare AI is not about replacing human expertise, but augmenting it. AI agents should function as specialized consultants, intelligent tools that complement clinical knowledge rather than compete with it. This requires building interdisciplinary teams that understand both technological capabilities and healthcare challenges, creating feedback mechanisms that continuously improve AI performance, and prioritizing data quality and domain-specific training.
By developing systems that can learn, adapt, and seamlessly communicate, we move closer to a truly intelligent healthcare ecosystem. The key is not just technological sophistication, but meaningful collaboration between human expertise and artificial intelligence.
The Path to Meaningful Healthcare AI
The potential is significant, but the path is complex. Organizations must view interoperability not as an optional enhancement but as a fundamental requirement for meaningful AI implementation.
This journey requires commitment, strategic thinking, and a willingness to challenge existing paradigms. Healthcare’s future depends on our ability to transform vast amounts of data into actionable, trustworthy insights that genuinely improve individual care.
The transformation starts now. Are we ready to reimagine healthcare AI’s potential?
