Machine Learning for Prognostic Assessment in Elderly Patients with Gastroduodenal Perforation

Abstract Background Elderly gastroduodenal perforation (E-GDP) is prone to severe adverse effects (SAEs) and high mortality rates. This study aims to develop an intelligent prognostic assessment system for the comprehensive evaluation of E-GDP. Methods A total of 161 patients with E-GDP who underwent emergency surgery at two medical centers from June 2017 to June 2024 were included. Patients' clinical features, laboratory findings, and postoperative complications were retrospectively collected. Patients were categorized into superior prognosis group(SG) and inferior prognosis group(IG) using two-step clustering (TSC). Machine learning models predicted E-GDP prognosis, interpreted via SHapley Additive exPlanations (SHAP) analysis, and were validated at an external center. Results After TSC, patients were divided into two groups: the SG (76 patients) and the IG (32 patients). TSC better distinguished hospital days than SAE grade (p < 0.001 vs. p = 0.01). Compared with the SG, the IG had greater postoperative white blood cell counts (10.61 ± 4.82 vs. 8.01 ± 2.41, p = 0.006), intra-abdominal infections (94% vs. 0%, p < 0.001), anastomotic leakage (44% vs. 0%, p < 0.001), fluid time (8.01 ± 3.41 vs. 6.05 ± 2.12, p = 0.004), length of intensive care unit stay (114.20 ± 188.35 vs. 18.07 ± 48.55, p = 0.008), and costs (92180.79 ± 70920.55 vs. 38416.18 ± 20479.48, p < 0.001). The support vector machine (SVM) had the best performance, with an AUC of 0.90 (0.61, 0.84) and 0.77 (0.61, 0.92) in the external validation set(30 patients). Conclusion We developed a robust SVM-based machine learning model using the TSC assessment system to preoperatively predict the prognosis of E-GDP patients.