Machine Learning to Predict Complications After Percutaneous Native Kidney Needle Biopsy

10.6221/AN.202409_38(3).0005

Yu-Chen Chiu；Te-Chun Wang；Wei-Jie Wang；Hung-Chieh Wu

complication ； Joint Commission of Taiwan ； native kidney biopsy ； machine learning ； random forest

Acta Nephrologica

38卷3期（2024 / 09 / 01）

181 - 193

英文

BACKGROUND: Risk prediction models for post-biopsy complication have not been developed. The study aimed to develop and validate a model to predict post-biopsy complication. METHODS: Participants who underwent percutaneous native kidney biopsy in Taoyuan General Hospital from 2014 to 2023 were enrolled. Demographic data, comorbidities, laboratory data and procedure-related characteristics were assessed. We defined post-biopsy minor complications according to macrohematuria, perirenal hematoma and hemoglobin drop. Major complications were defined as those who need intervention. The 16 factors machine learning models, such as random forest (RF), logistic regression (LR), K-nearest neighbor (KNN), multilayer perceptron (MLP) and sector vector machine (SVM) were used and the 6 most critical risk factors were extracted by feature selection. Model performance was evaluated by area under receiver operating characteristic (AUROC) curve. To avoid overfitting and distorted effect of imbalanced data, we also compared accuracy, F1 score, and positive predictive value. RESULTS: From a total of 694 subjects, 561 vs. 133 participants were classified as development vs. validation dataset. The complication rates were 5.7% vs. 9.7%, respectively. After feature selection, bleeding time, creatinine, hemoglobin, proteinuria, parenchymal thickness, and liver function were essential to predict complications. Despite low sensitivity, RF is the best prediction model according to AUROC (0.943 [0.901-0.985] vs. 0.910 [0.837-0.982]), accuracy (0.973 vs. 0.960) and F1 score (0.727 vs. 0.758). The prediction performance of 6-factors RF model was non-inferior to 16-factors model (AUROC: 0.943 vs. 0.900 and 0.910 vs. 0.869, P > 0.05). CONCLUSION: The study demonstrated the feasibility of 6-factors RF model to predict post-biopsy complications.