Related Experiment Video
Updated: Jan 8, 2026
Supervised Machine Learning for Semi-Quantification of Extracellular DNA in Glomerulonephritis
Published on: June 18, 2020
Enhancing Glomerular Hematuria Identification in Automated Urinalysis Using a Light Gradient Boosting Machine-Based
Rongrong Wang1, Jia Xu1, Jing Jin1
1Department of Clinical Laboratory, State Key Laboratory of Complex Severe and Rare Diseases, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
Background:
Glomerular hematuria (GH) is a key parameter assessed during urine analysis that is poorly identified using the automated UF-5000 system, which may misclassify GH and delay the detection of underlying glomerular diseases. We developed a light gradient boosting machine (LightGBM)-based model to improve GH detection using automated urinalysis data.
Methods:
We included 5,444 urine samples from patients with positive urinary occult blood results. All samples were manually classified into non-glomerular hematuria (NGH), GH, mixed hematuria, and non-hematuria groups, based on microscopic examination. We assessed 65 parameters using UF-5000 and UC-3500 analyzers and compared their performance with that of LightGBM, extreme gradient boosting, random forest, and logistic regression models. The final model was validated through 10-fold cross-validation with an independent test set and then compared with the performance of UF-5000. SHapley Additive exPlanations were applied to identify key predictive parameters.
Results:
Employing the LightGBM model substantially improved GH recognition accuracy to 44% during validation and 37% during testing (versus 16% and 11% achieved with the UF-5000 model, respectively). Sensitivity for GH increased from 0.3 in the UF-5000 model to 0.7 in the LightGBM model. A similar increasing trend was observed for the negative predictive value (0.6 to 0.9), accuracy (0.5 to 0.8), and Cohen's kappa agreement (0.4 to over 0.6). Key predictive parameters included red blood cell count, forward scatter peak in surface channel, and urinary protein level.
Conclusions:
This interpretable LightGBM-based model offers a substantial improvement in automated GH identification and is a promising tool for classifying hematuria sources.
