David Lareau is CEO of Medicomp Systems, which makes medical data relevant, usable and actionable.
getty
Sit in on any healthtech vendor pitch to a health system this year, and the pattern becomes impossible to miss: Nearly every solution on display has been wrapped in some form of AI. Whether ambient documentation, chart summarization, coding support or clinical search, the race to ship AI-enabled clinical products has accelerated. But it has also exposed a failure point most teams are working around rather than solving: There is no structured, validated clinical knowledge layer beneath the model.
Without that layer, outputs vary based on context, patient population and the quality of the underlying data. The same query can produce different answers on different days. A diagnosis inferred from an ambient transcript may be plausible in one chart and inappropriate in another. A summary can sound authoritative while simultaneously omitting the clinically relevant negatives that shaped the clinician’s reasoning. These are the outputs clinicians cannot act on with confidence, and it’s also why so many AI pilots stall short of enterprise deployment.
The industry is beginning to recognize what this missing piece represents, calling it deterministic AI, a clinical knowledge graph, a validated reasoning layer or a clinical data foundation. The label matters less than the function. The underlying idea is that probabilistic large language models (LLMs), on their own, cannot be trusted with decisions that affect patient safety, reimbursement and regulatory standing.
They require a deterministic foundation of curated clinical knowledge: a structured representation of clinical concepts, relationships and rules against which outputs can be validated before they reach a chart, a claim or a clinician's eyes.
LLMs are extraordinary at language. They were engineered to produce fluent, contextually reasonable responses, and they do so at a remarkable scale. Clinical decision-making, however, demands more. It requires consistency, traceability and alignment with evidence that has been curated and maintained over time. When an AI model produces three different summaries of the same encounter across three separate queries, that variability isn't just a reliability problem; it reflects a structural flaw in the technology.
After all, when probabilistic outputs flow unchecked into the medical record, the downstream consequences accumulate. Unsupported diagnoses migrate into problem lists, where they shape future care decisions. Coded data ends up too vague to support risk adjustment, quality reporting or interoperability. Documentation can appear coherent on the page while lacking the evidentiary support needed for audits, appeals and continuity of care. The answer isn't to abandon the model. It's to stop treating it as the entire product and build around it accordingly.
A deterministic clinical knowledge layer standardizes the terminology, relationships and clinical relevance that an AI output must conform to before it can be trusted. In practice, that means validating whether a suggested diagnosis is supported by documented findings, flagging coding too unspecified to be defensible, identifying missed conditions supported by the evidence in the chart and confirming that structured data generated from narrative reflects what the clinician intended to capture.
This layer is what makes AI outputs reproducible and explainable. A product team can then tell a health system leader, with specifics, why the technology surfaced a particular suggestion and why it suppressed another. That same layer allows sensitive clinical queries to be answered without exposing protected health information to external models, because structured queries and validated responses can stand in for narrative data.
The trajectory of clinical knowledge in healthcare AI mirrors how other categories of enterprise software have matured over the past two decades. Cloud infrastructure, identity management and observability all began as differentiating features before becoming non-negotiable layers. No serious enterprise product ships today without them. A well-tuned model may be sufficient for a narrow consumer-facing tool, but for a product that touches diagnosis, documentation, coding or care decisions, a validated clinical knowledge foundation is becoming the floor.
That shift changes the technology selection process for health system leaders. Evaluating AI roadmaps should now begin with a small set of direct questions. What clinical knowledge sits underneath the model, and who curates it? How are outputs validated before they're written to the record? Can the solution explain why it reached a particular conclusion? What happens when the underlying evidence in the chart changes?
If the honest answer to any of these is that the model is expected to figure it out, the product has a foundational problem that no amount of prompt engineering will resolve.
Those evaluation questions, and likely several others, will increasingly separate the products that thrive from the ones that stall. In the next wave of healthcare AI, differentiation will turn less on model sophistication and more on the quality and structure of the clinical knowledge beneath it.
Teams that recognize this early will build products that clinicians trust, customers adopt at scale and regulators can engage with constructively. Those that don't may spend the coming years patching outputs, managing liability and wondering why their pilots never graduated to full deployments.
The success of these solutions comes down to the less dazzling side of AI, but it's also where the real product work happens. Think of it this way: The model is the visible part of the product. What sits beneath it, the clinical knowledge layer, is what decides whether the product is worth using at all.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。