MAGMED: A clinically grounded LLM-guided synthetic augmentation framework for admission-time ICU mortality prediction.

BACKGROUND: Mortality risk prediction for elderly intensive care unit (ICU) patients with severe infections remains challenging due to limited sample sizes and poor generalization across institutions. Existing models often fail to adequately address cross-center distribution shifts and long-tailed outcome distributions, while operating under strict data privacy and deployment constraints. METHODS: To address these challenges, we propose MAGMED, a semantically regularized synthetic data augmentation framework. The proposed approach integrates the distribution modeling capability of variational autoencoders (VAEs) with the semantic reasoning capacity of large language models (LLMs). By fine-tuning a locally deployable and lightweight medical LLM, MAGMED generates physiologically plausible admission-time laboratory profiles together with clinically aligned explanatory narratives in an offline setting. This design simultaneously satisfies data privacy requirements and limited computational resources. We conducted comprehensive evaluations on a large multicenter cohort, including three development centers and one independent external validation center, to assess predictive performance, robustness under distribution shift, and clinical interpretability. RESULTS: Across internal validation experiments, MAGMED consistently outperformed representative augmentation baselines in overall performance. Moreover, the proposed framework demonstrated robust generalization under cross-institutional distribution shifts, effectively mitigating the performance degradation commonly observed in multicenter settings. Clinician-centered evaluations further confirmed that MAGMED-generated samples exhibited high clinical fidelity, with physiologically coherent biomarker patterns and interpretable explanations. Importantly, the framework was successfully integrated into real-world hospital workflows, enabling seamless human-in-the-loop decision support at admission time under real-world clinical deployment constraints. CONCLUSION: By explicitly enforcing semantic consistency between synthetic biomarkers and clinical explanations, MAGMED overcomes key limitations of conventional data augmentation strategies. The proposed framework offers a privacy-preserving, interpretable, and deployment-ready solution for clinical prognosis modeling, and provides a scalable pathway for integrating trustworthy artificial intelligence into real-world critical care practice.