Abstract
Malignant intestinal obstruction (MIO) is a severe complication of advanced cancer. Traditional static assessment models struggle to capture its dynamic pathological mechanisms, limiting their clinical value. To address this, this study developed a multimodal machine learning framework. Core features (including the dynamic tumor enhancement ratio TER) were extracted via Lasso-Boruta dual-modality screening, and risk prediction was performed using an XGBoost-Random Survival Forest (RSF) cascade model. Results demonstrated an AUC of 0.84 ± 0.03 and Brier score of 0.19 in the internal validation cohort, with robust external validation performance. Clinical translation reduced mechanical ventilation duration by 41% and lowered antibiotic resistance rates from 37 to 14%. This approach ultimately provides dynamic, interpretable decision support for precise MIO diagnosis and treatment. This study enrolled 320 MIO patients, randomly divided into a training set (192 cases), internal validation set (64 cases), and external validation set (64 cases) at a 6:2:2 ratio.