LLMs Enable Zero-Shot Chronic Kidney Disease Screening with Minimal Features

Researchers have investigated the potential of large language models (LLMs) for early detection of chronic kidney disease (CKD) without requiring extensive, labeled datasets or complex pathology tests. Their novel approach, called a feature-guided zero-shot framework, focuses on using a compact set of easily accessible, clinically relevant patient features. This method serializes tabular patient data into text prompts, allowing LLMs to perform inference without specific training on the CKD datasets. The study evaluated four prominent LLMs (LLaMA-3, Qwen-3, Mistral, and GPT-4o-mini) across three diverse international CKD datasets. The results consistently showed that using the selected, minimal feature set led to statistically significant improvements in screening performance, achieving accuracy levels suitable for practical application. This suggests LLMs can provide a valuable, training-free tool for CKD screening, particularly in resource-limited environments, by focusing on readily available community-based information.