60/100

Moderate evidence

Cite
More
EmailLinkedInX

ChatGPT misclassified more ovarian masses than frozen section in a 259-patient study

Study title: Diagnostic performance of ChatGPT in adnexal mass classification: A comparative study with frozen section and final histopathology

Evidence type
Observational study
Sample
n = 259
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Diagnostic performance of ChatGPT in adnexal mass classification: A comparative study with frozen section and final histopathology

A retrospective study of 259 patients who had surgery for ovarian masses between January 2021 and July 2025 compared ChatGPT with intraoperative frozen-section analysis. Final histopathology was the reference. ChatGPT reached a sensitivity of 90.0% and specificity of 92.5%, against 94.4% and 100% for frozen section, when borderline cases were excluded.

Read the study abstract

Objective To evaluate the diagnostic performance of ChatGPT in classifying adnexal masses as benign, borderline, or malignant, and to compare its performance with intraoperative frozen-section analysis, using final histopathology as the reference standard. Materials and methods This retrospective, observational, comparative study included 259 patients who underwent surgery for ovarian masses between January 2021 and July 2025. Preoperative clinical data, laboratory parameters, and ultrasonographic findings were…

The problem

In a retrospective study of 259 ovarian surgery patients, ChatGPT's positive predictive value for malignancy was 56.3% versus 100% for frozen section, with specificity of 92.5% versus 100%, so it produced many more false positives.

In brief

Histopathology classed 83.4% of cases as benign, 7.7% as borderline and 8.9% as malignant. ChatGPT's positive predictive value was 56.3% versus 100% for frozen section, though negative predictive values were close at 98.9% and 99.5%. The authors conclude ChatGPT could help rule out malignancy but cannot replace histopathology.

Performance dropped when borderline tumours were counted as malignant: sensitivity fell to 74.4%, specificity to 79.6% and PPV to 42.1%. Areas under the curve were 0.967 for frozen section and 0.893 for ChatGPT.

Main points

  1. The study covered 259 patients who had surgery for ovarian masses between January 2021 and July 2025.
  2. Final histopathology found 83.4% benign, 7.7% borderline and 8.9% malignant cases.
  3. With borderline cases excluded, ChatGPT had 90.0% sensitivity and 92.5% specificity; frozen section had 94.4% and 100%.
  4. ChatGPT's PPV was 56.3% versus 100% for frozen section; NPVs were 98.9% versus 99.5%.
  5. Counting borderline tumours as malignant cut ChatGPT's sensitivity to 74.4%, specificity to 79.6% and PPV to 42.1%.
  6. Areas under the curve were 0.967 for frozen section and 0.893 for ChatGPT.
  7. ChatGPT produced more false positives, particularly in the borderline and malignant categories.

The full story

Researchers gathered preoperative clinical data, lab parameters and ultrasound findings for each patient. ChatGPT assessed them using structured clinical inputs without prompt optimization, and its benign, borderline or malignant classifications were checked against final histopathology alongside frozen-section diagnoses.

Frozen section was highly accurate, and ChatGPT's high negative predictive value (98.9%) was close to that of frozen section (99.5%). That suggests ChatGPT may help in ruling out malignancy, in the authors' view.

Its lower specificity and PPV, however, led the authors to say it is limited as a standalone diagnostic tool. They describe it as a possible complementary decision-support tool that cannot replace histopathological evaluation.

Sources

1 on file
  1. Peer-reviewed2026-11-12

Talk it over

Three questions to start on, two more if you have time
ChatGPT's negative predictive value was 98.9% against 99.5% for frozen section, but its positive predictive value was 56.3% against 100%. How would you weigh its use for ruling out malignancy against its false positives?

Start with a question

ChatGPT misclassified more ovarian masses than frozen section in a 259-patient studyhttps://truvace.com/item/diagnostic-performance-of-chatgpt-in-adnexal-mass-classification-a-comparative-s

  1. ChatGPT's negative predictive value was 98.9% against 99.5% for frozen section, but its positive predictive value was 56.3% against 100%. How would you weigh its use for ruling out malignancy against its false positives?

    Answer in the thread
  2. The 259 patients were all in a retrospective study, and ChatGPT worked from structured inputs without prompt optimization. What does that design let us say about how ChatGPT would behave with a surgeon deciding in real time?

    Answer in the thread
  3. Suppose you were a patient whose ChatGPT classification said malignant while frozen section and final histopathology later said benign. What would you want the surgical team to have told you beforehand?

    Answer in the thread
If you have more time
  • When borderline tumours were counted as malignant, ChatGPT's specificity fell to 79.6% and its PPV to 42.1%. How might a hospital decide where borderline cases belong before using a tool like this?

    Answer in the thread
  • What would a prospective study of ChatGPT alongside frozen section need to measure for you to accept or reject it as a complementary decision-support tool?

    Answer in the thread

What readers say