A few years ago, I started experiencing symptoms I couldn’t explain, and no one could explain them either. Fatigue that wouldn’t lift, heart palpitations, an inability to lose weight regardless of what I did. I went to several doctors. I ran tests. Every time, I got the same answer: your labs look normal. The implication was clear. The problem was me.
What I eventually discovered, through my own research and insistence on looking deeper, was that my thyroid wasn’t converting hormones properly due to a genetic variant and increased stress over a short period of time, something that wasn’t visible through standard TSH testing. Combined with iron and vitamin D deficiencies and other nutrient gaps, the picture only made sense when I stopped looking at each marker in isolation and started looking at how they interacted. I was experiencing insulin resistance due to my PCOS genetics, but it was more than that. Individually, nothing looked alarming. Together, they told a different story entirely.
The doctors I saw weren’t bad doctors. They were working within a framework that wasn’t built to see what I needed them to see. And that framework is straining under the weight of what patients are now bringing to their appointments. People are arriving at primary care visits with wearable data, lab reports, genetic panels, and microbiome reports, armed with more information about their own biology than their doctor has time or infrastructure to interpret. The sequencing revolution gave us the richest map of human biology ever created. We are still running out of ways to reason across it in the room where it matters most.
Standard lab reference ranges carry enormous authority. They shouldn’t, at least not uncritically. Those ranges were largely calibrated against populations that didn’t adequately represent people across different life stages, genetic backgrounds, or hormonal profiles. Someone can receive a normal TSH result and still be symptomatic. For women alone, research suggests the majority with normal TSH levels have low T3, meaning the conversion from T4 isn’t happening adequately. That’s not an edge case. That’s a systematic blind spot built into the tools we depend on. Believe it or not, it’s happening with men, too.
The more precise question, the one personalized care actually demands, isn’t whether a result falls within a reference range. It’s whether that result is optimal for prevention for this specific person, given their genetics, history, and the relationship between their systems. This is where AI has a real role to play: not comparing a patient’s biomarkers to a flawed population average, but interpreting them alongside their individual genetic profile to identify what optimal actually looks like for their biology. Biomarkers read in isolation can’t do that. Standard ranges weren’t designed to. Those ranges also reflect the health of the population they were drawn from. Given the trajectory of chronic disease over the past two decades, using that population as a baseline target is not a neutral clinical decision. It’s accepting a floor for prevention that’s already too low.
The natural response to an infrastructure gap is better tools. In medicine, that conversation has become largely a conversation about AI, and specifically about large language models. The appeal is easy to understand. These systems are fast, fluent, and capable of synthesizing vast amounts of information quickly. But fluency is not the same as reasoning. Large language models pattern-match against the text they were trained on. They predict what a plausible clinical statement looks like. They are not working from causal biological logic.
The result is what clinicians are beginning to call “faithful hallucinations,” outputs that are confidently stated, clinically plausible, and factually wrong. In a recent study 91.8% of clinicians using AI tools reported encountering them, and 84.7% believed these hallucinations could cause patient harm. Consider a concrete example. An AI might state that DHEA is a precursor to both cortisol and sex hormones. To a non-specialist, this sounds authoritative. It has the texture of real endocrinology. But the biological pathway doesn’t support it. A clinician who catches it loses trust in the tool. One who doesn’t may act on it. The alternative is AI built on transparent, causal reasoning, systems that ground every output in validated biological mechanisms, show the pathway behind each recommendation, and cite their sources. It’s what makes the difference between a tool that erodes clinical trust and one that rebuilds it.
As doctors, you already know that a primary care physician has 15 minutes per patient. You’re trained to address population-level risk, not to run additional panels for patterns that standard protocols don’t flag. The system wasn’t designed for the kind of cross-system reasoning that personalized care requires, and in most cases, it doesn’t give patients visibility into the amount of work that goes on behind the scenes, outside of face-to-face contact.
The result is that the depth of care a patient receives depends heavily on where they live and what they can afford. A functional medicine specialist with the time and tools to reason simultaneously across endocrinology, immunology, and genomics is not equally accessible to everyone. This level of care, as it currently exists, is largely a concierge offering. A primary care physician in rural Arizona is no less capable than a specialist in Manhattan. They just have less infrastructure, and the gap is closeable!
Transparent, causal AI, the kind that connects biomarkers to genetic context, shows its biological reasoning, and cites its sources, doesn’t replace clinical judgment. It extends it. It gives a primary care provider the cross-system pattern recognition that currently requires years of subspecialty training, making that level of diagnostic capability available at the point of care regardless of zip code.
Trust in clinical AI will not come from better design or smoother interfaces. It’s built by giving clinicians transparent tools that cite their sources and show their biological reasoning, tools that let a physician interrogate the logic before acting on it, and catch it when it’s wrong. That kind of infrastructure changes what’s possible at the point of care. It doesn’t replace clinical judgment, but instead gives clinicians something to work with that’s actually equal to the complexity of what patients are now bringing through the door. The framework that treats the body as a collection of isolated parts, measured against a population average that was never universal, is not the ceiling of what medicine can be. It’s just where we’ve been. The space between that and what’s possible is real, it’s measurable, and it’s worth demanding we close it.
Photo: peterhowell, Getty Images
Elena Ikonomovska, PhD, is Co-Founder and CEO of Diadia Health, a causal reasoning health platform for clinicians managing complex chronic disease cases.A serial founder with nearly two decades of experience, Elena holds a PhD in machine learning. Her career includes work at Google, where she prototyped technology that became BigQuery’s backbone, and at Reddit, where she served as the company’s first data scientist, building content recommendation systems to detect harmful behavior.
Elena’s other ventures include co-founding Nuntio Labs, an AI writing tool that claimed customers such as Wells Fargo and Square, serving as Head of AI at Change.org, where she founded and led the ML/AI department, and co-founding Mnemonic, Inc., serving as CEO and Chief AI Officer, developing AI-enhanced blockchain intelligence. After a long personal health battle where she was perpetually dismissed, Elena co-founded Diadia Health to offer solutions to others facing similar challenges with standardized medicine.
This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.
