The fragility of post-earthquake public health environment has significantly increased the risk of gastrointestinal mucosal injury and infection spread. It is urgent to establish intelligent analysis methods that take into account recognition accuracy, early warning timeliness and complex data adaptability. In this paper, a multi-modal intelligent diagnosis and risk warning framework fusing pathological images, clinical texts and detection indicators is constructed for gastrointestinal mucosal lesions. Through standardized preprocessing, convolutional visual coding, medical text semantic representation, structured indicator embedding, shared space alignment, modal attention fusion and dual-branch collaborative decision making of lesion recognition-risk discrimination. The linkage output of lesion classification, risk score and hierarchical early warning was realized. The experimental results based on 1620 case samples show that the Accuracy, F1-score and AUC of the proposed method reach 93.2%, 92.5% and 0.964, respectively. The risk warning accuracy reaches 90.8%, the recall rate of Level IV cases reaches 94.1%, and the average warning advance time reaches 23.7 hours. The results of this study provide reliable technical support and practical significance for early screening, early diagnosis and stratified intervention of gastrointestinal mucosal lesions after earthquake.
