Article
Nov 7, 2026
What women's medical records leave out, and what it means for AI
A new STAT essay and Gallup data show many women's records miss dismissed symptoms and misdiagnoses. Here is why that matters for clinical AI.
AI tools in medicine learn from the electronic health record, and they summarize it, search it and draw conclusions from it. A new essay in STAT argues that for many women, the record leaves out the parts of the story that matter most. If that is true, AI built on those records will miss them too, unless clinics change how they capture what patients say.
What happened
On November 6, 2026, STAT published a First Opinion essay by epidemiologist Annemarie G. Hirsch and sociologist Jennifer Silva titled "AI depends on medical records that don't always tell the full truth about women." The two are the authors of the book Seen but Not Heard: What Medical Records Don't Tell Us About Women's Lives, published this year by the Russell Sage Foundation.
Their argument is simple. "Medical records are narratives, shaped by the questions clinicians ask and the assumptions they bring into the room," they write. AI scribes, record chatbots and diagnostic models treat that narrative as data, and "these tools are only as good as the data in the medical record."
What the records left out
The book draws on interviews with 87 non-college-educated, economically disadvantaged women living in rural America, matched with their health records from a large nonprofit health system. The researchers compared what the women described with what clinicians had documented. The gaps were large:
Broken trust: 53 of the women described an encounter that broke their trust in a health system. Only 7 had records that documented mistrust.
Misdiagnosis: 32 women reported a misdiagnosis. It was documented for only 8 of them.
Disrespect: 30 women said they felt clinicians treated them with disrespect. None of that appeared anywhere in their records.
This is a qualitative study of one group of women, not a national estimate. But it shows a specific mechanism: experiences that shape whether a patient comes back, follows a plan or discloses symptoms often never reach the chart.
A wider pattern in national data
The essay points to national survey data that suggests the problem is not limited to rural patients. The Gallup-Pivotal Women's Health Survey, fielded July 14 to 22, 2026, with 5,232 US adults, found:
34% of women said a provider had minimized or dismissed their symptoms or pain in the past five years, compared with 21% of men.
33% of women said a provider had misdiagnosed them or could not give a diagnosis, compared with 21% of men.
51% of women reported at least one problem when accessing care, compared with 39% of men.
47% of women said they waited three or more months for their most recent diagnosis.
Put the two sources together and the concern is clear. If dismissed symptoms and missed diagnoses are common, and the record rarely captures them, then the record systematically understates what women are experiencing.
Why this matters for AI
Most clinical AI does not talk to patients. It reads what is already written. That creates three risks.
Summaries inherit the gaps
A record chatbot or pre-charting tool that summarizes a patient's history can only surface what was documented. If years of dismissed pain were never written down, the summary will look tidy and complete when it is not.
Models learn from incomplete labels
Diagnostic and risk models are trained on coded diagnoses and notes. When misdiagnoses go undocumented, a model learns the wrong answer as the right one, and it can repeat that pattern at scale.
Scribes capture what is said in the room
Ambient scribes record the visit, which helps. But they still reflect the questions asked in a short appointment. A patient who feels rushed or unheard may not raise what matters, and the scribe has nothing to capture.
What this means for clinics
The essay's authors end with a plain goal: "medical records need to tell the complete story." For clinics in Canada and the US, that suggests a few practical steps.
Ask before the visit, in the patient's words. Giving patients time to describe symptoms, history and past experiences before the appointment surfaces details that a short visit often misses.
Document the patient's account, not only the clinical conclusion. Prior dismissals, delayed diagnoses and concerns about care are clinically relevant history.
Ask vendors what their AI reads. A tool that only summarizes the existing chart will carry its blind spots forward. Ask how it captures new information from patients.
Keep a clinician in the loop. AI output should be reviewed by someone who can notice what is missing, not just what is present.
This is the problem Athel is designed around. It interviews patients before the visit by voice, web chat or video and gives the physician a structured history built from what the patient said, and the clinician reviews every summary.
AI can make records faster to read. Whether it makes them more accurate depends on whether the patient's own story gets into the record in the first place.