ABSTRACT An innovative approach is needed for the early identification of GC (Gastric cancer) to improve the prediction of GC patients. This work presents a GC prediction system to identify GC depending on saliva data samples. The diagnosis of GC at an early stage will improve the survival rate. At first, Raman data of saliva samples are collected and pre-processed. Afterwards, efficient Raman spectral features are extracted from the pre-processed data. Then, the feature selection process is performed with a Levy search updated rainfall (LURF) optimisation approach. This optimisation scheme decreases the dimensionality of the features by integrating Levy’s flight and rainfall optimisation. Finally, the hybrid deep dual-stage bidirectional long short-term memory (Hybrid LURF) framework effectively classifies the data as normal or abnormal. This model efficiently addresses the issues of insufficient long-term dependency in GC prediction and also enhances the classification performance. The validation of the proposed approach is examined with various existing schemes and achieved better accuracy (98.5%), specificity (97%), sensitivity (96.5%), F1-score (93%), detection rate (98.4%) and ROC curve. Further, the accuracy is 0.06% better than multi-layer ANN and 10% better than SVM-polynomial and KNN models.