Non-invasive prediction of occult peritoneal metastasis (OPM) in gastric cancer using logistic regression and random forest integrative models with CT radiomics and clinical parameters: machine learning prediction of gastric OPM

Background: Occult peritoneal metastasis (OPM) in gastric cancer is frequently missed on preoperative contrast-enhanced computed tomography (CT), leading to understaging in about 30–40% of patients and necessitating invasive laparoscopy for confirmation. A non-invasive tool to preoperatively identify OPM would help avoid unnecessary surgery and guide individualized treatment, but existing models lack interpretability and do not compare linear vs. nonlinear decision mechanisms. Therefore, we aimed to develop interpretable integrated models and compare their decision mechanisms and predictive performance for non‑invasive preoperative prediction of gastric OPM. Methods: We retrospectively enrolled 164 gastric adenocarcinoma patients (100 with OPM, 64 non-metastatic) from a single center. Inclusion criteria were biopsy-proven adenocarcinoma, no prior malignancy, complete clinical data, and contrast-enhanced CT at initial diagnosis showing no definite peritoneal metastasis, with OPM subsequently confirmed by laparoscopic exploration, peritoneal biopsy, or positive peritoneal lavage cytology. Exclusion criteria included prior treatment, poor-quality images, unconfirmed peritoneal lesions, other distant metastases, or remnant stomach. Nine clinical-laboratory variables [age, sex, body mass index, alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA19-9), albumin (ALB), tumor location, and tumor/node (T/N) stage] were collected, and three-dimensional (3D) radiomics features were extracted from primary-tumor CT volumes. The cohort was randomly split 7:3 into training (n=114, 70 OPM) and validation (n=50, 30 OPM) sets using stratified sampling. After intraclass correlation coefficient (ICC) (>0.75) filtering, least absolute shrinkage and selection operator (LASSO) with 10-fold cross-validation [1 − standard error (1-SE) rule] and Spearman correlation (threshold 0.75) selected six stable radiomics features from CT images. Integrated models combining age, T stage, and radiomics signature were built using five machine learning algorithms. Evaluation included area under the curve (AUC), calibration curves, Brier score, decision curve analysis (DCA), precision-recall (PR) curves, and average precision (AP). SHapley Additive exPlanations (SHAP) analysis was performed on both logistic regression (LR) and random forest (RF) models, and Friedman’s H-statistic quantified feature interactions. Results: Age and T stage were independent clinical predictors. The integrated LR model achieved training/validation AUCs of 0.907/0.868, with validation sensitivity of 0.833 and specificity of 0.800. The RF model yielded training/validation AUCs of 0.839/0.857, with validation sensitivity of 0.833 and specificity of 0.800. SHAP comparison revealed crossing decision paths in LR vs. cleaner boundaries in RF, with moderate feature interactions (H=0.100–0.136). A simplified nomogram (Rad-score + age + T stage) achieved a validation AUC of 0.846. Conclusions: This LR-based integrated model demonstrates promising performance and interpretability for preoperative OPM risk stratification in CT-negative gastric cancer patients in this single-center internal validation. Comparative SHAP analysis uncovers nonlinear dynamics and a novel paradoxical radiomics signature, providing biological insights beyond conventional prediction. However, prospective multicenter external validation is required before this approach can be considered for clinical implementation.