Diagnostic agreement between ChatGPT ‐based interpretation of laboratory‐generated flow cytometry graphical outputs and hematologist consensus in suspected acute leukemia

Accurate, rapid, and standardized interpretation of multiparametric flow cytometry findings is essential for clinical decision-making in patients evaluated for suspected acute leukemia. This study aimed to evaluate the diagnostic performance of ChatGPT, a general-purpose large language model, in interpreting de-identified, laboratory-generated flow cytometry graphical outputs from patients with suspected acute leukemia and to assess its agreement with the consensus interpretation of two hematologists in terms of diagnostic classification, blast percentage estimation, and hematogone detection. This retrospective observational validation study included 560 patients who underwent bone marrow aspiration and flow cytometric immunophenotyping during their initial diagnostic evaluation for suspected acute leukemia. All included patients were undergoing initial diagnostic evaluation and had neither a previous diagnosis of leukemia nor prior leukemia-directed treatment. The consensus diagnosis established by two hematologists was considered the reference standard. Between May 4 and 15, 2026, de-identified, laboratory-generated PDF images containing flow cytometry dot plots from a three-tube acute leukemia panel comprising 28 antibody-fluorochrome combinations were uploaded directly through the ChatGPT Plus web interface and interpreted using GPT-5.5 (OpenAI). Raw FCS files were not provided. Using the same standardized prompt for each case, the model was asked to estimate the blast percentage, select a single most likely diagnosis from five predefined categories (no pathology, MDS, AML, B-ALL, or T-ALL), and determine the presence or absence of hematogones. The initial response for each case was archived and directly encoded without adjudication. Diagnosis-specific performance was evaluated using one-versus-rest sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and F1-score. Overall agreement between ChatGPT-generated classifications and hematologist consensus was assessed using Cohen's kappa. Pearson's correlation coefficient was used to evaluate the association between blast percentage estimates, whereas Bland-Altman analysis was used to quantify agreement, systematic bias, and 95% limits of agreement. Hematogone detection was evaluated as a binary classification task using sensitivity, specificity, and accuracy. ChatGPT achieved an overall diagnostic accuracy of 78.8% (95% CI, 75.2%-81.9%), with substantial agreement with hematologist consensus (Cohen's κ = 0.712; 95% CI, 0.664-0.757). Diagnosis-specific sensitivity ranged from 69.2% to 87.3%, and PPV ranged from 60.0% to 93.5%. The highest sensitivity was observed for MDS, whereas the highest PPV was observed for no pathology. McNemar's test identified significant systematic differences between ChatGPT and hematologist classifications for no pathology (p = 0.041), AML (p = 0.008), and T-ALL (p = 0.023). Blast percentage estimates were strongly correlated (Pearson's r = 0.818; 95% CI, 0.789-0.844; p< 0.001); however, Bland-Altman analysis demonstrated a mean positive bias of 3.92 percentage points, indicating systematic overestimation by ChatGPT. For hematogone detection, ChatGPT achieved an accuracy of 89.5%, sensitivity of 70.3%, and specificity of 92.4%. ChatGPT-based flow cytometry interpretation showed promise for preliminary diagnostic classification and quality assurance applications. However, expert hematologist validation remains essential, particularly in scenarios associated with high diagnostic consequences, including AML and T-ALL classification, blast percentage estimation, and accurate discrimination of hematogones.