Why Healthcare AI Needs Better Data Before It Needs Better Models by Isobel Glanville-Pearl and Mark Stanton-Bennett

Why do 90% of Western healthcare budgets go towards disease management rather than prevention?

The system was simply never designed for prevention. It was instead designed for intervention. Healthcare operates in reactive silos, with funding models, workforce structures, and data systems all geared towards rewarding activity after disease occurs. Governments set KPIs around managing conditions like COPD and diabetes rather than preventing them.

Prevention also demands longitudinal visibility, but healthcare remains largely episodic, meaning we see patients when they’re sick, not before. COVID, however, has accelerated a shift in public appetite. More people now want to take greater autonomy over their own health and to engage with preventative options.

Fragmented healthcare data is often cited as a major barrier. Why is this an even greater problem in 2026?

We have more health data than ever, from wearables to personalised tests to regulated functional diagnostics that sit outside conventional medicine. But rather than closing the gap, this explosion of data has widened the divide between our current system and the preventative model we need to move towards.

In practice, clinicians across primary, secondary, and acute care are all working on different platforms such as Epic, Cerner, and others with GP systems that don’t integrate with any of them. There is no longitudinal record of a patient’s health journey. Add to that the growing sophistication around genomics and complex datasets, and you have a situation where vital data exists but cannot be meaningfully connected.

Most clinical data is unstructured such as notes, scans, voice recordings. AI has largely focused on the structured 20%. What are we missing?

We are missing clinical reasoning, context, and change over time. Structured data tells you what is happening; unstructured data tells you why. Without integrating both, you never achieve a complete view of the patient.

The consequence is significant: we are currently training AI on a heavily redacted version of reality. Until unstructured data is properly incorporated, the outputs will always be incomplete.

Most medical AI is not trained on real patient data. What does that reveal about the true state of the field versus the hype?

We are still in a performative phase. Most systems are trained on synthetic, narrow datasets, which produces approximations, not intelligence. From a clinical standpoint, that carries genuine risk. The hard problem was never building the models; it is building the infrastructure to safely and ethically learn from real patient journeys. Until that is solved, medical AI will remain impressive in demonstration but limited in transformative impact.

The constraints go deeper than data availability. We must be honest: this is still machine learning, and narrow or synthetic datasets quickly expose the limitations of the underlying technology. Without sufficient breadth, depth, and context, you cannot reliably determine which data points matter, how to test them, or whether a model is surfacing the right features. Results will, at best, be adequate.

There is also a fundamental issue of expectation. Truly predictive AI does not yet exist and we are making informed guesses based on historical data. Context can sharpen those guesses, but the field will not reach its potential until it has access to deeper, richer, and more representative datasets to underpin real clinical decision-making.

If unifying longitudinal clinical data is so clearly valuable, why hasn’t it been done? What are the blockers?

The blockers are multiple and systemic. Data is a commercial asset, which immediately introduces competition. Layer on top of that cultural risk aversion, governance and compliance complexity, and widespread misunderstanding of AI, and the result is an ecosystem of siloed data that organisations don’t know how to safely integrate – except within tightly controlled research contexts, which are themselves heavily bureaucratic.

There is also a fundamental incentive misalignment. Without funded research and a defined outcome, there is no motivation to bring disparate datasets together. And critically, there is still no category-defining platform capable of unifying clinical data at scale, meaning the problem remains unsolved, not just for technical reasons but structural ones.

Clinical data is both a commercial asset and a public good. Is that tension holding progress back?

It is, and it largely goes unspoken. The moment data is associated with monetisation, the conversation shuts down, despite the fact that commercial use of patient data is already commonplace. Large hospital groups and EHR providers are already building redacted longitudinal datasets for research purposes, often without meaningful patient notification.

The deeper issue is one of agency and accountability. Most patients assume they own their data, unaware that it has already passed through multiple systems and providers. Ensuring patients retain genuine control is the right hope, but it requires accountability structures that do not yet exist at scale. People would be shocked to understand how widely their data has been used without their knowledge.

Is regulation really the villain in the data-sharing debate?

No, and that framing is itself part of the problem. Regulation is frequently used as a scapegoat by those who lack a clear understanding of governance, and that reflects a deeper issue: a lack of trust. With greater transparency, clearer processes, and auditable data flows, regulation can become an enabler that builds patient confidence rather than obstructing progress. What the sector needs is not less regulation, but standardised frameworks for the ethical use of data.

Investment is flowing into generative AI tools such as scribes, co-pilots and imaging models. What are their real limitations?

The industry needs an honest conversation about the limitations of the underlying technology. LLMs and the broader generative AI suite share the same fundamental model architecture and that architecture has real constraints.

The pattern emerging is one of progressive narrowing. Faced with the genuine difficulty of building generalist AI, many businesses scope down their use cases to where data is available and results look strong. They go narrower and narrower, each time producing more impressive-looking outputs from an increasingly limited problem. Venture capital pressure accelerates this because investors want demonstrable returns, which incentivises the appearance of breakthroughs over genuine ones.

The result is a landscape of tools that perform well on very specific tasks but fall well short of the broad, generalist capability the industry originally promised. The technology isn’t being made more powerful and in many cases, it’s being made narrower to simulate that effect.

What needs to change to genuinely solve these problems?

The shift required is from data collection to data orchestration, and from isolated point solutions to a proper infrastructure layer. The focus must move away from episodic care towards longitudinal, patient-centred models built on interoperable systems, although achieving true interoperability at a global scale is, realistically, never going to happen. The practical question is how to build systems that enable data to flow freely enough to be meaningful. AI that learns continuously, rather than from static datasets, will also be essential.

Crucially, this remains a question of augmentation, not replacement. The data, models, and interoperable infrastructure needed for AI to operate independently do not yet exist. Domain experts must remain in the loop; providing context, enriching data, and shaping model development in ways that are genuinely suited to the problems at hand. We are still in the early stages, and human expertise is what bridges the gap between the technology’s current limitations and its long-term potential.

What are the essential components of a well-functioning healthcare data infrastructure?

Three things: a single unified patient view, real-time insights, and genuine confidence in data provenance. That third element is arguably the most important and it barely exists today. Data provenance encompasses how data is used, how consent is obtained and maintained, and how patients retain ownership and dynamic control over their own records. Greater patient engagement would improve data quality significantly. If people understood their data was being used to improve outcomes rather than for commercial gain, we could reduce the need for anonymisation and work with richer, more meaningful datasets.

There is also a critical and underappreciated risk in how AI systems are adopted. Many clinicians and healthcare organisations purchase AI tools and ingest their data without fully understanding what those models were trained on, or what the provenance implications are. Once data enters a model, it cannot be redacted. These conversations are not happening at the scale or depth they need to be.

What does a trusted, auditable, continuously learning data environment look like from a clinician’s perspective?

Clinical validation must remain central and with it, an honest acknowledgement that AI could increase risk as much as reduce it, if it breeds complacency. Human judgment and clinical intuition are not replicable. The ability to recognise that something is wrong with a patient, even when every objective data point looks normal, comes from experience and human interaction. AI cannot substitute for that. It is a decision-support tool, not a diagnostic system, and that distinction matters enormously.

Should clinical data be treated as a regulated asset class with ongoing transferable value rather than a by-product of care? What guardrails are needed?

Full auditability is the essential starting point. Knowing how data is used, where it goes, and being able to account for every instance of it. No healthcare system in the world can currently do that. As AI scales, that gap becomes increasingly dangerous.

The longer-term answer lies in genuine patient ownership, which means individuals holding their own records, with auditable, blockchain-enabled access and dynamic consent over how their data is used. That also opens the door to equitable participation: patients benefiting from the value their own data generates, rather than that value accruing solely to the organisations that hold it. We are not there yet, but it is where the framework needs to go.

If we fast-forward ten years, what does preventative healthcare look like for the average patient?

The vision is a continuous health record from birth to death, regularly updated and genuinely useful. Rather than visiting a doctor only when sick, patients attend meaningful annual check-ups, i.e. purposeful assessments, not box-ticking exercises. Interventions become personalised and proactive, replacing the protocol-driven, one-size-fits-all approach that has dominated healthcare for decades. Care becomes predictive rather than reactive. That is the goal.

Getting there will take at least ten years. Healthcare systems globally are in firefighting mode, and the mindset shift required, from treating illness to sustaining health is profound. But the direction is clear.

One point worth emphasising is why blockchain is not simply a technical detail in this vision, it is foundational to the entire solution.

Built correctly, blockchain embeds trust and compliance by design from the moment data enters a system. That matters enormously when the alternative is retrofitting governance onto infrastructure that was never designed for it.

It also addresses one of the most pressing problems in healthcare AI: auditability. As the field moves beyond generative AI and into agentic systems, models that take actions and make decisions autonomously, the ability to trace and explain those decisions becomes non-negotiable. The same prompt fed into the same model can produce different outputs. In most contexts that is an inconvenience; in healthcare it is a patient safety issue.

Blockchain provides the mechanism to ensure decisions follow consistent, traceable paths, and to capture and explain deviations when they occur. The combination of blockchain and AI offers something the sector urgently needs: not just capability, but credibility and the foundation for genuine confidence in these technologies among clinicians, patients, and regulators alike.