Leveraging Large Language Models for Predicting Postoperative Acute Kidney Injury in Elderly Patients

Objective: The objective of this work is to develop a framework based on large language models (LLMs) to predict postoperative acute kidney injury (AKI) outcomes in elderly patients. Impact Statement: Our study demonstrates that LLMs have the potential to address the issues of poor generalization and weak interpretability commonly encountered in disease prediction using traditional machine learning (ML) models. Introduction: AKI is a severe postoperative complication, especially in elderly patients with declining renal function. Current AKI prediction models rely on ML, but their lack of interpretability and generalizability limits clinical use. LLMs, with extensive pretraining and text generation capabilities, offer a new solution. Methods: We applied prompt engineering and knowledge distillation based on instruction fine-tuning to optimize LLMs for AKI prediction. The framework was tested on 2,649 samples from 2 private Chinese hospitals and one public South Korean dataset, which were divided into internal and external datasets. Results: The LLM framework showed robust external performance, with accuracy rates: commercial LLMs (internal: 63.73%, external: 68.73%), open-source LLMs (internal: 63.70%, external: 64.24%), and ML models (internal: 63.93%, external: 58.27%). LLMs also provided human-readable explanations for better clinical understanding. Conclusion: The proposed framework showcases the potential of LLMs to enhance generalization and interpretability in postoperative AKI prediction, paving the way for more robust and transparent predictive solutions in clinical settings.