MOMA: multi-expert framework with missing pattern awareness for rectal cancer neoadjuvant therapy

Neoadjuvant therapy plays a critical role in the management of rectal cancer. To improve the prediction of treatment response, we propose MOMA, a multi-expert fusion framework with missing pattern awareness, which is developed through three key steps. First, large language models (LLMs), optimized via human-AI interaction, are employed to structure multimodal medical records. Second, the structured data are grouped according to their missing patterns. Third, multi-expert feature interaction modeling is designed to generate predictive outcomes. The framework is empirically evaluated on a real-world dataset from China-Japan Friendship Hospital. Results show that MOMA achieves an accuracy of 81.5% in predicting neoadjuvant therapy response, outperforming the best traditional machine learning model by approximately 3.3%. These findings suggest that MOMA can serve as an efficient and accurate decision-support tool for clinicians in personalizing neoadjuvant therapy strategies for rectal cancer patients.