Aiming at the problems that only depending on pathological conclusions to carry out efficacy explanation after intervention for gastric precancerous lesions, endoscopic descriptions do not keep consistent, and multi-source information is hard to read in unified way, this paper builds a multimodal intelligent analysis system which is made specially for the intervention situation of Curcuma wenyujin. This system is designed for the identification of efficacy responses at the patient level and the assessment of stage migration. This research has in it 228 patients whose gastric precancerous lesions got pathologically proved, there are 114 cases in the Curcuma wenyujin intervention group and 114 cases in the control group. Longitudinal following data on Baseline, Month 3, and Month 6 were arranged to make a structured sample which includes 5,472 endoscopic images, 1,824 pathological site labels, and 26 clinical serum variables. Methodologically, the system consists of an endoscopic image branch, a pathological branch, and a clinical-serum branch. It achieves patient-level feature alignment through an attention-weighted cross-modal fusion mechanism and simultaneously outputs efficacy response results and OLGA/OLGIM stage migration results. The results show that the response rate, pathological remission rate, and OLGIM/OLGA downstage ratio in the Curcuma wenyujin group are higher than those in the control group at 6 months; specifically, the response rate is 58.8% compared to 37.7% in the control group, and the pathological remission rates are 53.5% and 31.6%, respectively. Additionally, the Curcuma wenyujin group exhibits more consistent improvement directions in indicators such as pathological load, mucosal load, PGI, PGR, G-17, IL-6, and TNF-α. Model comparison results indicate that the multimodal system achieves in the patient-level effect identification, we have obtained an AUC value 0.923, an F1-score value 0.854, and a Brier score value 0.108, hence it has better performance than the control models which only use clinical data, only use endoscopy data, only use pathology data, and dual-modal control models. The ablation experiment outcomes further make known that the pathological branch possesses the strongest pulling force for patient-level explanation, the endoscopic branch offers important shape-related increases, and the clinical-serum branch greatly enhances probability adjustment and subgroup stability. The analysis of errors indicates that the errors of the model are mainly concentrated in the situations of boundary stage transfer, multiple focal lesions, low-quality endoscopic images, and situations where pathological changes and endoscopic changes are not synchronous. The outcomes of this research prove that the system has the ability to integrate pathological alterations, mucosal appearance characteristics, and clinical serum targets into one united assessment frame, hence allowing combined explanation on the patient level after the intervention is done. This approach can be utilized to support review arrangements and risk reassessment, and provides a scalable methodological foundation for the intelligent efficacy evaluation of traditional Chinese medicine interventions in precancerous lesions of the stomach.
