This paper introduces RAGCTM (RoBERTa-wwm-ATT-BiGRU-CNN-Traffine-MLP), an innovative model specifically designed to address the complexities inherent in nested and discontinuous phenomena commonly found in biomedical entities. Traditional methods often encounter significant challenges when dealing with these intricate structures, leading to suboptimal identification of relevant entities. To overcome these limitations, RAGCTM enhances entity recognition by accurately predicting the relationships between individual words within their contextual framework. The architecture of RAGCTM comprises four essential modules: the encoding layer, which processes initial data; the hybrid convolutional layer, responsible for extracting relevant features using both convolutional and recurrent techniques; the prediction layer, which generates predictions for entity identification; and the decoding layer, which reconstructs the outputs for final interpretation. To bolster performance and resilience, adversarial training based on Projected Gradient Descent (PGD) is integrated into the training process, enhancing the model’s ability to generalize and resist overfitting. Furthermore, a strategic weighting mechanism balances the contributions of Dice Loss and cross-entropy loss functions, effectively addressing the prevalent issue of data imbalance in biomedical datasets. Experimental results reveal that RAGCTM achieves remarkable F1 scores of 80.98% on the nested dataset GENIA and 73.01% on the discontinuous dataset CADEC, underscoring the model’s effectiveness and its significant potential to advance biomedical entity recognition, ultimately contributing to improved outcomes in clinical data analysis and biomedical research.