A Logistic Regression Model Integrating Flow Cytometry-Derived Immune Cell Profiles and Hematological Parameters for Preoperative Prediction of Peritoneal Metastasis in Gastric Cancer.

BACKGROUND: Peritoneal metastasis (PM) is a lethal and often occult event in advanced gastric cancer (GC), and its preoperative detection remains difficult. The role of peripheral NK and NKT-like cells in predicting PM has not been well defined. METHODS: We retrospectively analyzed 433 patients who underwent surgery for GC from 2016 to 2020. Patients were stratified by PM status and randomly assigned to training and held-out internal validation cohorts. Flow cytometry-derived lymphocyte subset percentages were treated as compositional data, and NK and NKT-like cell variables were entered into the model after log-ratio transformation. Candidate variables were selected using LASSO logistic regression in the training cohort and further assessed by multivariable logistic regression. Model performance was evaluated using AUROC, AUPRC, and confusion matrices in the validation cohort. Nomogram and SHAP analyses were used for model interpretation. RESULTS: Seven predictors were retained in the final model: log-ratio-transformed NKT-like cells, direct bilirubin, prealbumin, lymphocyte count, lactate dehydrogenase, CA125, and log-ratio-transformed NK cells. The model achieved an AUROC of 0.879 (95% CI: 0.805-0.936) and an AUPRC of 0.698 (95% CI: 0.524-0.848) in the validation cohort. At the conventional threshold of 0.50, the model achieved an accuracy of 0.915, sensitivity of 0.821, specificity of 0.941, PPV of 0.793, and NPV of 0.950. At the Youden-optimal threshold of 0.101, sensitivity, specificity, accuracy, PPV, and NPV were 1.000, 0.892, 0.915, 0.718, and 1.000, respectively. CONCLUSIONS: A preoperative model integrating log-ratio-transformed NK/NKT-like cell variables and routine hematological parameters showed good ability to identify PM in GC and may help select patients for diagnostic laparoscopy or closer preoperative evaluation.