Gastric cancer has become one of the major diseases that affect people's lives. Early and timely screening can effectively improve the survival rate of patients. However, the reality is that relying on a single type of data for prediction has a certain gap from the ideal level, and many prediction models ignore the connection of rich contextual signals in medical information. To address this challenge, this study proposes a hybrid model integrating machine learning and Bidirectional Long Short-Term Memory Network (Bi-LSTM) based on multimodal time series data, which fuses clinical laboratory test results and textual information of gastroscopy results at different times. In the proposed model, the Bi-LSTM module with residual connections is designed to effectively extract long-term dependencies in medical time-series data to generate high-dimensional feature vectors. The integrated machine learning module of XGBoost, LightGBM and CatBoost is employed to process data in parallel. Through dynamic weighting by the dual-head attention mechanism of features and models, the fusion prediction results become more accurate. The provided results approve that the prediction accuracy can reach 95.74%,the area under the curve (AUC) is 96.89%, and the F1 score is 93.23%. The designed hybrid model facilitates rational allocation of medical resources, thereby enhancing trust in patient-doctor relationships and efficiency in gastric cancer care.