Related Experiment Videos
MAGMED: A clinically grounded LLM-guided synthetic augmentation framework for admission-time ICU mortality prediction
Wenxiong Chen1, Chao Yu1, Zhihong Zuo2
1School of Information Science and Engineering, Hunan Normal University, Changsha, 410081, Hunan, China.
Computers in Biology and Medicine
|July 18, 2026
Summary
Predicting mortality risk in severe infections is hard. MAGMED, a synthetic data augmentation framework using AI, generates realistic patient data and explanations, improving predictions and clinical decision-making while ensuring data privacy.
Area of Science:
- Artificial Intelligence in Medicine
- Critical Care Medicine
- Data Science
Background:
- Mortality risk prediction for severe infections in ICU patients is challenging due to limited data and generalization issues.
- Existing models struggle with cross-institutional data shifts and privacy constraints.
- Accurate prognosis is crucial for timely intervention in critical care.
Purpose of the Study:
- To develop a novel synthetic data augmentation framework, MAGMED, for improved mortality risk prediction in elderly ICU patients.
- To address limitations of existing models, including data privacy, computational constraints, and cross-center generalizability.
- To create a privacy-preserving and interpretable AI solution for critical care prognosis.
Main Methods:
- MAGMED integrates variational autoencoders (VAEs) for distribution modeling and large language models (LLMs) for semantic reasoning.
- A lightweight, locally deployable medical LLM is fine-tuned to generate synthetic lab profiles and clinical narratives.
- The framework operates offline, ensuring data privacy and efficient resource utilization.
Main Results:
- MAGMED outperformed baseline augmentation methods in internal validation.
- The framework demonstrated robust generalization across multiple institutions, mitigating performance degradation.
- Generated synthetic data exhibited high clinical fidelity, with coherent biomarkers and interpretable explanations, validated by clinicians.
- Successful integration into real-world hospital workflows for admission-time decision support.
Conclusions:
- MAGMED overcomes limitations of traditional data augmentation by enforcing semantic consistency between synthetic data and explanations.
- The framework provides a privacy-preserving, interpretable, and deployment-ready solution for clinical prognosis.
- MAGMED facilitates the integration of trustworthy AI into critical care practice, enhancing decision support.