A study published in September 2026 introduced ObGynLongBench, a benchmark designed to evaluate AI models on obstetric and gynaecological clinical decisions using longitudinal electronic health records.

The benchmark contains 1,500 clinical decision points from 976 pregnancy histories. The authors found that models perform better when relevant evidence is supplied directly and lose accuracy when they must locate it within long, complex records.

The challenge is finding the right evidence

The work highlights an important limitation for clinical assistants: answer quality depends on identifying and using information that was available at the time of the decision, without leaking later events into the reasoning process.

Source: ObGynLongBench, arXiv, 7 September 2026.