AI trained on incomplete women’s medical records risks worsening healthcare

A recent analysis highlights a major shortcoming in the datasets used to train artificial intelligence systems for healthcare: electronic medical records frequently exclude critical aspects of women’s health experiences, creating blind spots that can deepen biases in diagnosis and treatment decisions.

Medical records omit critical details about women’s health

When the team compared electronic records with in-depth interviews from 87 women, however, they uncovered a significant mismatch.

Fifty-three participants described encounters that shattered trust in the healthcare system, yet only seven of those instances appeared in their records. Thirty-two women reported misdiagnoses, but only eight were documented. Thirty women said they felt disrespected by clinicians, yet no mention of this appeared in their files. The findings suggest medical records—often treated as neutral documentation—are instead shaped by clinician assumptions and patients’ reluctance to disclose uncomfortable truths.

A case study illustrates the consequences. A woman, identified as Kate, endured debilitating pelvic pain for over a decade, with doctors repeatedly attributing her symptoms to obesity. Her records included notes like “She can’t control what she is eating” and “She admits to poor eating habits.” After weight-loss surgery, the narrative shifted to mental health explanations: “Gastrointestinal malfunctions arising from mental factors.” Only when her husband attended an appointment and noted “Pain makes pt stop having intercourse” did doctors conduct the necessary examinations. Kate was finally diagnosed with endometriosis after ten years of suffering.

While her medical record did document discussions about weight and mental health, it omitted her frustration at not being believed, the decade spent undiagnosed, and the erosion of trust. “They’re not hearing what you’re saying,” she said, “and that compounded the issues that were going on.”

How AI inherits flawed data

Medical records are not objective; they reflect clinician decisions, preconceptions, and the language used to describe patients. The details recorded can reinforce stereotypes, shaping how future providers interpret symptoms. For example, if a woman’s pain is repeatedly framed as psychological, that framing may follow her across providers.

AI systems trained on these records risk amplifying existing biases. If a model learns that a woman’s symptoms are “likely stress-related” based on past notes, it may replicate that assumption in future recommendations. The issue extends beyond individual cases: when women fear judgment or dismissal, they may withhold critical information, leaving records incomplete.

Many clinics and hospitals now use AI scribes to transcribe doctor-patient interactions and generate notes. While these tools can improve efficiency, their output depends on the quality of the input. If a record omits a patient’s full story, such as the emotional toll of repeated misdiagnoses, AI may overlook patterns that could lead to better care.

AI could address disparities, if the data improves

This gap has serious implications because AI systems trained on incomplete data may perpetuate historical biases. Over time, such omissions could train models to deprioritize certain symptoms in women.