This study addresses the challenges of privacy protection and data silos in the intelligent analysis and health risk prediction of electronic health records by proposing a federated learning-based framework. In this framework, data from different medical institutions do not require centralized storage but instead achieve cross-institutional collaborative optimization through local model training and secure parameter aggregation, thereby improving model performance under conditions of compliance and privacy protection. A multimodal feature fusion mechanism is introduced to jointly model structured diagnostic information, clinical text, and time-series data, enabling the capture of complex semantic associations and temporal relationships across modalities. At the output stage, the model adopts a probabilistic prediction strategy optimized with cross-entropy loss, which effectively enhances the accuracy and stability of risk identification. Experiments conducted on a public electronic health record dataset show that the proposed method outperforms several baseline models in accuracy, precision, recall, and F1-Score, achieving a good balance between privacy protection and predictive robustness. Overall, this study establishes an integrated framework that combines federated learning with multimodal modeling, providing a feasible path for the efficient use of electronic health records and health risk prediction, while demonstrating significant advantages in improving the value of medical data and supporting better health management decisions.

This study addresses the problem of insufficient context capture in text classification and proposes a large language model method enhanced with contextual mechanisms. At the input layer, raw text is transformed into vector sequences that incorporate both semantic and sequential features through the combination of embedding representation and positional encoding. A context encoder based on self-attention is then introduced to capture global dependencies within the sequence. At the same time, a context gating unit is designed to achieve dynamic fusion of local and global information, which preserves fine-grained features while strengthening overall contextual consistency. Furthermore, a global context aggregation module integrates semantic information across sentences or paragraphs, enhancing the model's ability to represent long texts and implicit semantics. In the output stage, sentence-level pooling is used to generate a unified representation, followed by a classification head to complete label prediction. To validate the effectiveness of the method, comparative experiments were conducted on a public news text classification dataset. The results show that the proposed method outperforms traditional deep learning models and mainstream large-model baselines in terms of accuracy, precision, recall, and F1-score. It maintains a more stable classification performance when dealing with semantic ambiguity and topic shifts. In addition, sensitivity experiments on hidden dimension settings demonstrate that moderate model capacity significantly improves performance, while excessive complexity may introduce redundant representations and slight overfitting. This study demonstrates the practical value of context enhancement mechanisms in large language models and provides a more robust and effective solution for text classification tasks.