评估基于AI的血细胞分析中用于卵巢癌风险检测的血液学模式变异

Ovarian cancer(OC) lacks reliable early detection methods, and blood cell-derived inflammatory signatures have emerged as potential non-invasive indicators. However, these hematologic patterns may vary across individuals and sample-processing conditions. This study applies an AI-driven blood cell analysis model to a new cohort and evaluates feature-level variability relevant to OC risk detection. Peripheral blood was collected to evaluate hematologic patterns associated with OC detection from asymptomatic controls, patients with benign ovarian or uterine tumors, and patients with newly diagnosed OC. Samples were transported under refrigerated conditions and processed within 28 hours for complete blood count analysis. Composite hematologic indices were calculated from measured parameters and input into our previously developed AI-driven blood cell analysis model to evaluate OC risk signals. Residual samples were additionally used to obtain platelet images for exploratory morphological assessment. Among 135 samples analyzed, including 16 OC cases, the AI-driven blood cell analysis model produced 6 true positives, 10 false negatives, 17 false positives, and 102 true negatives, yielding a sensitivity of 37.5%, specificity of 85.7%, PPV 26.1%, and NPV 91.9%. True-positive cases showed pronounced inflammatory activation with markedly elevated SII, NLR, and PLR and the lowest lymphocyte counts. False positives demonstrated similarly high inflammatory indices but disproportionately elevated platelet counts. False-negative OCs exhibited near-normal inflammatory markers yet had the highest RDW-CV, MPV, and PDW, indicating a morphology-dominant rather than inflammation-dominant hematologic phenotype. Although overall performance was lower than in our previous dataset using the same AI-driven blood cell analysis model, the consistency of feature importance suggests that the inflammatory signal remains stable. Reduced sensitivity is likely related to pre-analytical differences, as samples were refrigerated and processed after transport. False-negative cases showed a morphology-dominant profile not captured by inflammation-based indices, indicating that some OCs lack strong systemic inflammatory signatures. To address this limitation, we are acquiring platelet-level image data to develop a complementary morphology-focused diagnostic approach. The reproducibility of key hematologic patterns across datasets supports the biological relevance of AI-driven blood cell analysis for OC signal detection. However, reduced performance under altered processing conditions and the presence of morphology-dominant tumors highlight the limitations of inflammation-focused models. Incorporating platelet-image features may enhance detection of under-represented phenotypes and improve diagnostic coverage across OC subtypes.