ChatGPT matched hematologists on 78.8% of 560 leukemia flow cytometry cases, study finds

Agreement was substantial (Cohen's kappa 0.712), and blast percentage estimates correlated strongly with the hematologists' (r = 0.818). But the chatbot overestimated blasts by an average of 3.92 percentage points and differed systematically from experts on AML, T-ALL and no-pathology classification.

The authors conclude the approach shows promise for preliminary classification and quality assurance, while stating that hematologist validation remains essential where diagnostic stakes are high, such as AML and T-ALL calls, blast counts and hematogone discrimination.

Between May 4 and 15, 2026, the team uploaded laboratory-generated PDF images directly through the ChatGPT Plus web interface. Each case used the same standardized prompt asking for a blast percentage, a single most likely diagnosis and whether hematogones were present. The first response for every case was archived and encoded without adjudication.

The reference standard was the consensus diagnosis of two hematologists. Diagnosis-specific performance used one-versus-rest sensitivity, specificity, predictive values, accuracy and F1-score, with Cohen's kappa for overall agreement, Pearson correlation and Bland-Altman analysis for blasts.

Results showed 78.8% overall accuracy (95% CI 75.2%-81.9%) and kappa of 0.712. Significant systematic differences from hematologists appeared for no pathology (p = 0.041), AML (p = 0.008) and T-ALL (p = 0.023).

Hematogone detection was more specific than sensitive: 92.4% specificity against 70.3% sensitivity, at 89.5% accuracy. The authors say expert validation remains essential for high-consequence scenarios.