A pilot machine learning study for preoperative prediction of pathological lymph node status in PIPAC-treated gastric cancer in Kazakhstan: A single-center study

Pathological lymph-node status after neoadjuvant treatment and surgery (ypN) is prognostic in gastric cancer but unavailable before resection. We developed a pilot prediction model to estimate ypN risk from seven variables recorded before radical surgery in patients treated within a single-center PIPAC program; the timing of blood counts relative to neoadjuvant treatment could not be established. Of 62 patients, 47 with recorded ypN were analyzed (18 ypN-positive). Predictors were clinical T and N stage, tumor grade, age, body-mass index, white-cell count, and platelet count. Ridge logistic regression (primary), random forest, and Firth logistic regression were evaluated by cross-validation. Out-of-fold AUCs were 0.701 (95% confidence interval 0.539-0.851), 0.667 (0.488-0.831), and 0.703 (0.532-0.851), respectively; Brier scores were 0.227, 0.227, and 0.222 (prevalence-only reference 0.236). Performance was modest and imprecise; because the predictor panel was selected using the same cross-validation, estimates may be optimistic. These findings are an exploratory signal requiring independent validation.