A Multimodal Deep Learning Model for Preoperative Prediction of Postoperative Complications in Gastric Cancer.

BACKGROUND: Clavien-Dindo (CD) grade ≥II complications occur in roughly one in five patients after curative gastrectomy for gastric cancer and independently shorten overall survival (OS) by compromising adjuvant therapy delivery. Current nutritional, inflammatory and global surgical risk scores identify only a minority of affected patients. We developed and validated a multi-modal deep learning framework for simultaneous preoperative prediction of CD grade ≥II complications and long-term survival. PATIENTS AND METHODS: This multicenter study analyzed 5,237 patients with gastric adenocarcinoma from 11 Chinese centers and additionally validated the model using prospectively collected data from six registered clinical trials covering three neoadjuvant treatment setting (chemotherapy, chemoradiation, and immunochemotherapy). DeepComp integrates clinical variables with foundation-model image features from the target lesion region, 5-mm peritumoral region, and L3 body-composition compartments through a tabular dual-task architecture to jointly predict CD grade ≥II complications and OS. RESULTS: DeepComp achieved an AUC of 0.888 (95% CI 0.854 - 0.921) in the merged internal validation set and 0.824 to 0.869 across nine external cohorts, outperforming the best clinical baseline by 15.3 percentage points (P<0.001) and all nine established clinical scores (all P<0.001). With DeepComp assistance, the mean sensitivity of 10 surgeons rose from 47.1% to 87.9% (P<0.001). Through target trial emulation, DeepComp-guided prophylactic ICU monitoring, preoperative nutritional support with delayed surgery, and minimally invasive triage yielded absolute CD ≥II risk reductions of 5.9%, 20.6%, and 11.7%, respectively (all P<0.01; E-values 1.90 - 5.73). DeepComp independently predicted OS (adjusted hazard ratio 3.08 per standard-deviation, 95% CI 2.91 - 3.25; pooled C-index 0.766), with 5-year survival ranging from 97.5% in quintile 1 to 2.4% in quintile 5. CONCLUSIONS: DeepComp showed consistent performance across validation cohorts for preoperative risk stratification of postoperative complications and long-term survival, and may support individualized perioperative management.