AKI-Detector: A Multi-Agent Framework by Integrating Machine Learning and Large Language Models for Early Prediction...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Critical Care Medicine

Background:

• Acute kidney injury (AKI) is a critical ICU condition requiring early detection.
• Traditional machine learning (ML) models for AKI prediction lack interpretability, hindering clinical adoption.
• Explainable AI is essential for integrating predictive models into clinical workflows.

Purpose of the Study:

• To develop and evaluate AKI-Detector, a novel multi-agent framework for enhanced AKI prediction.
• To improve the accuracy and interpretability of AKI prediction by integrating ML, LLMs, and RAG.
• To provide clinically actionable and explainable reports for AKI risk.

Main Methods:

• Developed a multi-agent framework (AKI-Detector) combining EHR-based ML models, LLMs, and retrieval-augmented generation (RAG).
• Integrated ML models with LLMs to mitigate hallucinations and enhance clinical reasoning.
• Evaluated AKI-Detector on MIMIC-IV ICU data, comparing its performance against traditional ML models.

Main Results:

• AKI-Detector achieved an accuracy of 0.827, precision of 0.672, recall of 0.542, and F1-score of 0.600 on the test cohort.
• The framework demonstrated superior predictive performance compared to ML models like CatBoost and GRU.
• The system successfully generated interpretable reports, bridging the gap between algorithmic output and clinical understanding.

Conclusions:

• The proposed AKI-Detector framework shows significant promise for trustworthy and explainable AI in clinical prediction.
• Multi-agent systems integrating big data and LLMs can enhance clinical decision support for critical conditions like AKI.
• This approach supports timely intervention and prevention of AKI in intensive care units.