Objectives
Chronic kidney disease (CKD) presents a growing public health challenge in China, exacerbated by low patient awareness and limited nephrology resources. This study evaluated the potential of large language models (LLMs) to support CKD diagnosis in primary care using time-series electronic health record (EHR) data.
Methods
Longitudinal data were extracted from the EHR system of Weinan, China. Among 29 963 adults meeting inclusion criteria, 2300 participants were randomly sampled to generate clinical vignettes. CKD status was determined using diagnosis data, and individuals with early kidney injury indicated by laboratory abnormalities who lack a formal diagnosis were also identified. Using the DeepSeek LLM with prompt engineering, we generated binary CKD diagnoses and probability scores based on EHR data. Diagnostic performance was evaluated using accuracy, sensitivity, specificity, F1 score, area under the curve (AUC) and the detection rate for early kidney injury and compared with traditional machine learning (ML) models.
Results
Using 1-month EHR data, the LLM achieved an accuracy of 87.0%, AUC of 0.921, F1 score of 0.579, sensitivity of 89.1%, specificity of 86.8%, and detection rate for early kidney injury of 42.2%. With 1-year data, sensitivity increased to 92.2% and early detection to 52.6%.
Discussion
The LLM outperformed ML models across most metrics, achieving diagnostic performance comparable to clinicians and further enhancing detection for early kidney injury.
Conclusions
LLMs applied to time-series EHR data may serve as a clinical decision support tool to improve CKD diagnosis in primary care, with particular value in resource-limited settings.
