In 2024, over 70 health AI companies received FDA clearance. Investment in clinical AI crossed $15 billion globally. Every major health system announced an AI strategy.
And the number of patients who deteriorate without warning in US hospitals has not meaningfully changed. Chronic kidney disease still takes years to catch. Diagnostic errors still affect an estimated 12 million American adults annually in outpatient settings. Your new doctor still asks you to remember your own medication list.
Something is wrong with the approach.
What most healthcare AI actually does
Most clinical AI products summarize notes, classify images, or fire alerts when a number crosses a threshold. All useful. All wrappers. They take clinical data in its existing form, run it through a model, and present a result. The underlying data hasn't changed. The way the system represents a patient hasn't changed.
A radiologist reading a chest X-ray with AI assistance is still reading one image from one moment. The AI may catch a nodule she would have missed. Genuinely valuable. But it can't tell her that this patient's pulmonary function has been declining across five tests over three years, that the decline accelerated after a medication change eight months ago, and that the trajectory projects toward COPD within eighteen months. That information exists. It's scattered across providers and years. No wrapper can reach it because a wrapper only sees the current encounter.
The unit of analysis in almost all clinical AI is one visit. One note. One image. In human disease, it is years. Until those match, AI in healthcare will keep producing faster answers to the wrong questions.
The unit of analysis in human disease is years.
The speed trap
Speed is the benefit everyone cites. Summarize a discharge note in seconds. Draft a referral while the physician finishes the conversation. Extract diagnoses from free text instantly.
Physician burnout is real. Documentation burden is crushing. Anything that returns time to clinical care has value. But speed on the existing infrastructure solves an efficiency problem. The physician who gets a perfect AI summary of today's visit still doesn't know that her patient's creatinine has been rising across three providers over the past two years. The summary is faster. The clinical picture is identical.
Healthcare systems adopt AI that makes workflows faster, report productivity gains, and call the AI strategy a success. Meanwhile, the gap between when disease starts developing and when it gets caught hasn't moved. The system is moving faster in the same direction. Nobody asked whether the direction is right.
Why the industry builds wrappers
Wrappers are rational. Take a pre-trained model, fine-tune it on clinical text, ship a product in months. Validation is simple: note in, summary out. Sales is familiar: "We plug into your EHR and make it smarter." No hard conversations about data architecture. No workflow disruption.
Building something that reads health over time is a different undertaking entirely. You have to represent the patient across years and providers, build clinical reasoning that operates on direction instead of position, validate across hundreds of sites, and convince hospitals that catching a disease two years earlier justifies deeper integration than a summarization plugin.
That takes longer, costs more, and requires conversations that most teams aren't equipped to have. It is also the only approach that addresses the actual clinical gap.
The black box makes it worse
Most clinical AI operates as a black box on top of being a wrapper. A model ingests data and produces an output. The physician can't see the reasoning. She can't question which inputs drove the conclusion. She can't apply her judgment to the intermediate steps because there are no visible intermediate steps.
In practice, clinicians treat these outputs the way they treat the hundreds of other EHR alerts: with suspicion and a trained instinct to dismiss. Opaque systems consistently achieve 10 to 15 percent engagement. The system fires. The clinician ignores it. The AI is running but clinically inert.
The systems that have achieved over 80 percent clinician engagement share one trait: the reasoning is visible. The clinician sees what drove the assessment and can disagree with specific parts. Without that, adoption doesn't happen. I watched it not happen for years before I started building.
What different actually looks like
A patient whose creatinine has been climbing 0.1 per quarter across two providers over 18 months. No single value triggered an alarm. Every visit, the number was "within range." A system that reads trajectories catches it. The physician gets an alert that shows the trend, the projected path, and the clinical reasoning behind the flag. She reviews it, agrees, and refers to nephrology a year before the patient would have crossed the threshold.
That is the kind of change we are building Chroneme to catch. We are validating it under a pre-registered protocol, and will publish what we find, including where it falls short.
The same intelligence that catches the slow drift in an ICU can catch the slow drift in a primary care office. The same system that flags a heart rate trending upward over weeks can flag an A1c trending upward over years. The architecture doesn't change. The time scale does.
Fifteen billion dollars. Seventy clearances. And the patient experience hasn't changed.
It will change when the infrastructure changes. When clinical AI reads trajectories instead of snapshots. When systems show their reasoning instead of hiding it. When health becomes something a person can see changing over time, not something that gets caught after the fact.
That's what we're building at Chroneme.
