Chronic kidney disease (CKD) constitutes a critical global public health challenge. Its early-stage symptoms are subtle and easily overlooked, frequently resulting in delayed diagnosis and escalated treatment expenditures. To enable timely early intervention, this study proposes a deep learning-based multimodal CKD stage prediction model, which integrates Western medical laboratory data and traditional Chinese medicine (TCM) symptom descriptions, thereby overcoming the inherent limitations of unimodal prediction approaches. For Western medical data, the synthetic minority oversampling technique (SMOTE) was employed to address class imbalance. Subsequently, a least absolute shrinkage and selection operator (LASSO) feature selection model was utilized to identify key biomarkers, including serum creatinine and serum chloride. A backpropagation neural network, enhanced with Adam optimization and regularization mechanisms, was constructed for predictive modeling. For TCM symptom texts (e.g., tongue manifestations and pulse conditions), the Google-pretrained Bidirectional Encoder Representations from Transformers (BERT) model was leveraged to learn latent semantic patterns in the textual data. Finally, a multimodal decision fusion strategy based on the attention mechanism was adopted to dynamically learn the relative importance of Western medical and TCM features in CKD staging prediction, culminating in the development of the proposed deep learning-driven multimodal CKD staging model. Validation results indicate that the proposed model outperforms classical machine learning models and unimodal models across multiple metrics, including accuracy, F1-score, area under the curve, and convergence efficiency. These findings confirm its clinical feasibility and effectiveness, providing an innovative multimodal data-driven prediction framework that synergizes the strengths of Western medical quantitative testing and TCM qualitative diagnosis.