Performance Evaluation of GPT-5, Grok 4, and DeepSeek R1 in Interpreting Complete Blood Count Reports for Hematologic...

Large language models (LLMs) demonstrate potential in the laboratory, yet rigorous clinical evaluation remains limited. The opacity of LLM decision-making constrains their safe application in interpreting complete blood count (CBC) reports for hematologic diseases.

This study aimed to conduct an exploratory evaluation of GPT-5, Grok 4, and DeepSeek R1 in interpreting real-world CBC reports, particularly their reasoning capabilities and clinical safety.

This single-center retrospective study analyzed 100 CBC reports from initial-visit patients with hematologic conditions. After responses were generated by the 3 LLMs using standardized Chinese prompts, four trained laboratory physicians blindly evaluated them across 6 quality and 5 task dimensions. Interrater reliability was assessed using intraclass correlation coefficients (ICCs), and performance differences were assessed based on 4-rater consensus scores and Friedman and Wilcoxon tests. For task 4 (ablation analysis), the McNemar test was used to compare top-1 diagnostic concordance with the gold-standard diagnosis within each model, with and without initial clinical suspicion in the prompt. Error types and distributions were documented during the task evaluation.

DeepSeek R1 demonstrated excellent interrater reliability across most quality dimensions (ICC ≥0.75). In the quality dimension, DeepSeek R1 significantly outperformed the other models in comprehensiveness, accuracy, clarity, relevance, and practicality. In the task 4 evaluation, GPT-5 demonstrated the highest concordance (93/100, 93%) with gold-standard diagnoses, followed by DeepSeek R1 (92/100, 92%) and Grok 4 (89/100, 89%). After removing the initial clinical suspicion, these rates decreased to 79% (79/100), 77% (77/100), and 72% (72/100), representing statistically significant within-model reductions for all models (P<.001). Post hoc error analysis revealed distinct patterns across task dimensions. GPT-5 exhibited 12 hallucinations in the analyzer alert processing task; DeepSeek R1 demonstrated 1 hallucination in the abnormal item identification task, whereas Grok 4 displayed none. All models exhibited reasoning errors and varying degrees of deficiencies in the correlation analysis and preliminary diagnosis tasks, characterized by unwarranted inferences of disease status from isolated results without clinical integration. Grok 4 generated 9 reasoning errors in the clinical management task by providing generic recommendations not tailored to case-specific CBC data, potentially compromising individualized treatment decisions.

While current LLMs demonstrate potential for interpreting CBC reports in hematologic diseases, they show performance heterogeneity across models. The ablation study findings underscore the necessity of integrating clinical context for accurate laboratory test interpretation. Low scores, hallucinations, and reasoning errors in model outputs indicate that current clinical deployment requires human oversight and quality control. As this single-center, Chinese-language exploratory assessment provides only preliminary, possibly context-dependent evidence, multicenter, cross-lingual prospective validation is needed to delineate the practical boundaries and safety standards for clinical deployment.