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

Area of Science:

• Oncology
• Surgical Innovation
• Artificial Intelligence in Medicine

Background:

• Post-gastrectomy Clavien-Dindo (CD) grade ≥II complications affect ~20% of patients, reducing overall survival (OS) by hindering adjuvant therapy.
• Existing risk scores inadequately identify patients at high risk for these complications.
• A novel multi-modal deep learning framework was developed for simultaneous preoperative prediction of CD grade ≥II complications and long-term survival.

Purpose of the Study:

• To develop and validate a deep learning framework (DeepComp) for predicting postoperative complications and long-term survival in gastric cancer patients undergoing gastrectomy.
• To assess DeepComp's performance against existing clinical scores and its impact on surgical decision-making.
• To evaluate the potential of DeepComp in guiding individualized perioperative management strategies.

Main Methods:

• A multicenter study involving 5,237 gastric adenocarcinoma patients from 11 Chinese centers.
• External validation using prospectively collected data from six clinical trials across three neoadjuvant settings.
• DeepComp integrates clinical data with foundation-model image features (lesion, peritumoral, L3 body composition) using a tabular dual-task architecture.

Main Results:

• DeepComp achieved high AUCs (0.888 internal, 0.824-0.869 external validation), significantly outperforming clinical baselines and scores.
• Surgeon sensitivity in identifying high-risk patients improved from 47.1% to 87.9% with DeepComp assistance.
• DeepComp-guided interventions demonstrated significant reductions in CD ≥II complications (e.g., 20.6% with nutritional support).
• DeepComp independently predicted OS (C-index 0.766), with 5-year survival varying widely across predicted risk quintiles.

Conclusions:

• DeepComp demonstrates consistent performance in preoperative risk stratification for postoperative complications and long-term survival.
• The model shows potential for supporting individualized perioperative management in gastric cancer patients.
• This AI-driven approach offers a promising tool for enhancing surgical outcomes and patient survival.