Author: Yuanhan Chen, Ziqiang Chen, Qi Cheng, Feng Ding, Rui Fang, Shengwen Guo, Li Hao, Qiang He, Haiquan Huang, Yu Kuang, Xinling Liang, Yuanjiang Liao, Guohui Liu, Chen Lu, Qingquan Luo, Jing Sun, Yanhua Wu, Zhen Xie, Qin Zhang, Lang Zhou 👨🔬
Affiliation: South China University of Technology, Dongguan people's hospital, Sichuan Provincial People's Hospital, People’s Hospital of Xinjiang Uygur Autonomous Region, Second Hospital of Anhui Medical University, Guangdong Provincial People’s Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangdong Easy Life Information Technology Co., Ltd, Shanghai Ninth People's Hospital, Shanghai Jiaotong University School of Medicine, The First Affiliated Hospital of Zhejiang Chinese Medical University (Zhejiang Provincial Hospital of Traditional Chinese Medicine), Medical Physics Program, University of Nevada, Second Hospital of Jilin University, Chongqing Ninth People's Hospital 🌍
Purpose: Acute kidney injury (AKI) is a global healthcare issue with a rapid onset and severe consequences. Repeated measurement of serum creatinine (SCr) levels, a clinical standard of care, is crucial for the early diagnosis of AKI. However, the actual frequency of SCr testing is quite low, especially in low- and middle-income countries. Moreover, these regions also face issues with heterogeneous non-standardized electronic medical records (EMR) datasets across institutions. This study aims to develop a machine learning model to predict AKI, with or without the use of SCr data.
Methods: The Light Gradient Boosting Machine (LightGBM) was employed to develop predictive models in a multicenter study. Key features were extracted from non-standard EMRs collected from 2010 to 2016 across 15 hospitals in a middle-income country. The models’ performance was evaluated and compared using area under the curve (AUC), precision, sensitivity, specificity, and accuracy. Both internal validation, using 10-fold cross-validation, and validation on an independent test cohort were performed.
Results: A total of 561,137 hospitalized patients were eligible for the analyses, of whom 45,610 were diagnosed with AKI. The LightGBM models could effectively predict AKI occurrence within 24, 48, and 72 hours, with AUC values ranging from 0.860 to 0.986. Importantly, the model maintained strong predictive performance even without SCr data, with AUC values ranging from 0.861 to 0.882.
Conclusion: Non-standard EMRs are feasible for predicting AKI, even without SCr data. This approach is particularly useful in resource-limited settings, where traditional biomarkers are often unavailable, demonstrating the potential of other clinical features to compensate for missing SCr data in AKI prediction.