A friend of my mother’s went in for routine surgery prep and had her blood drawn. The test showed critically low hemoglobin, and no one followed up. Three months later she saw the surgeon, had the test redrawn, and the result was the same. The low hemoglobin had been an early sign of colon cancer, already advanced by the time she was diagnosed. The number that might have saved her had been sitting in her chart the whole time.
Her physicians weren’t careless. The first result was recorded in one part of the system and never reached anyone in a position to act on it. Months passed before a second test caught what the first had shown. This is how American medicine usually works, and the patients it hurts most are the ones whose conditions surface slowly, over years and across specialists, in records no one reads as a whole story.
Built for billing, not for people
When electronic health records arrived, they were built to document a transaction. That design made sense for getting hospitals paid. But it was never meant to follow one patient over time, or to gather what every other doctor and clinic recorded along the way.
Real physician voices, twice a week
Free, and one click to unsubscribe.
A landmark 2009 analysis in the Annals of Internal Medicine found that a typical primary care physician’s Medicare patients required coordinating with 229 other physicians across 117 practices. That data is almost two decades old, and the problem has only grown since. Between 2009 and 2019, the average number of primary care doctors a Medicare beneficiary sees rose by nearly half.
A system gets the results it’s designed for. The information a diagnosis needs usually exists somewhere in the record. Until recently, we couldn’t synthesize and surface it in the moment. Now, that’s starting to change.
Why women wait longest
This design fails women most, because so many conditions that affect them reveal themselves only as patterns over time. Endometriosis, for example, takes nearly seven years, on average, to diagnose. It affects roughly one in ten women of reproductive age, close to 190 million worldwide, with as many as 60 percent of cases never diagnosed.
In one survey, 58 percent of patients had seen a general practitioner more than ten times before a diagnosis. Those visits happened. Each generated symptoms that went into the record, but no one could read them as a single story. And because no clinician sees the full picture, the same tests get ordered again and again.
The delay costs more than time. A diagnosis delayed three to five years drives up spending by $12,971 to $34,460 per patient. The system charges for years of searching for a diagnosis, even though the signs were there from the first visit.
The story in the record
This is the work AI is genuinely suited for. A model built for health care can read across the fragmented record, pull in the free-text notes that a standard database query would miss, and structure them into something a clinician can track. Give it four years of scattered visits and it can surface the thread running through them, connecting the recurring pelvic pain, the missed cycles, the escalating requests for painkillers, and the emergency room visits.
The insight that once took seven years can surface on the second visit. The result that was missed in my mother’s friend’s chart gets flagged the moment it appears. And for a woman with endometriosis, that’s the difference between a decade of being told her pain is normal and a documented pattern that earns her a diagnosis.
Building AI to earn trust
AI earns trust in two ways. The first is the data it learns from. For most of modern medicine, women were left out of the research that set the standard of care. The FDA’s 1977 guidance excluded women of childbearing potential from early drug trials, and that restriction stood until 1993, when Congress and the FDA finally required their inclusion. As recently as 2019, women were still enrolled in trials below their share of the patients a treatment is meant to serve. Train AI on that history and it inherits blind spots, applied faster and at greater scale.
The second requirement for trust is transparency. I’ve watched physicians ignore an AI alert in the ICU until it could show every factor behind its reasoning, the way we expect a resident to walk us through a diagnosis. Traceability turns an alert from noise into something a doctor will use. AI can assemble the whole picture and point to the detail buried on page 180 of a chart. The choice of what to do with it belongs to the patient and the physician together, because the factors that decide a treatment plan are often the ones no dataset captures, like a fear of needles or a pill someone simply can’t swallow.
This is within reach. The interoperability rules now taking effect require health plans to share a patient’s records across systems. The next five years can put clinicians back where they do their best work, face to face with a patient, confident the record is finally capturing and connecting what matters.
You can’t diagnose a story your system was never built to remember. We finally have the means to build one that does, for the woman on her tenth visit still waiting for an answer, and for the friend whose result was read too late.
Niki Panich is a physician executive.
