Machine Learning Could Offer Noninvasive, Reliable Diagnostic Tool in IgA Nephropathy

News
Article

Findings suggest machine learning models may have promising utility in IgAN, highlighting the superior diagnostic performance of certain models versus standard logistic regression.

Ryunosuke Noda, MD | Credit: ResearchGate

Ryunosuke Noda, MD

Credit: ResearchGate

Machine learning models leveraging common clinical data could offer a promising new tool for predicting IgA nephropathy (IgAN) diagnosis, according to findings from a recent study.1

Results showed models employing XGBoost, LightGBM, and Random Forest had better diagnostic performance compared to a conventional statistical model and were able to handle complex relationships of prediction between key variables beyond simple linearity.1

“IgAN diagnostic prediction studies have predominantly employed logistic regression, a conventional statistical model assuming linear relationships and thus limiting predictive performance. Advanced machine learning algorithms, capable of modeling non-linear relationships and complex interactions, could improve predictive performance,” Ryunosuke Noda, MD, of the division of nephrology and hypertension in the department of internal medicine at St. Marianna University in Japan, and colleagues wrote.1 “However, the efficacy of machine learning in predicting IgAN diagnosis remains unexplored.”

An autoimmune condition involving the buildup of immunoglobulin A in the kidneys, IgAN causes inflammation and kidney damage potentially leading to chronic kidney disease, kidney failure, and death. Its presentation varies from patient to patient, taking years to develop in some while progressing more rapidly in others.2 Importantly, the exact causes of IgAN are not well understood and the only way it is definitely diagnosed is through kidney biopsy, underscoring the need for a noninvasive tool to facilitate early detection and treatment to improve outcomes for these patients.3

To develop non-invasive prediction models for IgAN using machine learning, investigators retrospectively collected data from electronic health records on demographic characteristics, blood tests, and urine tests of adult patients who underwent native kidney biopsy at St. Marianna University Hospital from January 1, 2006, to September 30, 2022. The total dataset comprised 1268 patients and was divided into 2 cohorts, with patients who underwent kidney biopsy between January 1, 2006, and December 31, 2019, included in the derivation cohort (n = 1027) and those who underwent biopsy between January 1, 2020, and September 30, 2022, included in the validation cohort (n = 241).1

Investigators employed 5 machine learning models, including XGBoost, LightGBM, Random Forest, Artificial Neural Networks, and 1 Dimentional-Convolutional Neural Network (1D-CNN), along with logistic regression. A total of 14 variables were selected as predictors and included in the machine learning models: age; hemoglobin; total protein; albumin, lactate dehydrogenase (LDH); creatine kinase (CK), C-reactive protein; immunoglobulin G (IgG); IgA; complement C3; complement C4; IgA/C3; urine red blood cells; and urine protein to creatinine ratio (UPCR). Investigators evaluated the performance of each model using the area under the receiver operating characteristic curve (AUROC), additionally exploring variable importance through SHapley Additive exPlanations (SHAP) method.1

Among the total study population, 353 (28%) patients were diagnosed with IgAN, including 294 (28.6%) in the derivation cohort and 59 (24.5%) in the validation cohort.1

In the derivation cohort, LightGBM achieved the greatest AUROC (0.913; 95% CI, 0.906–0.919), significantly greater than logistic regression, Artificial Neural Network, and 1D-CNN, although not significantly different from XGBoost and Random Forest. In the validation cohort, however, investigators pointed out XGBoost had the highest AUROC (0.894; 95% CI, 0.850–0.935), with no significant differences observed with any models. Of note, in the derivation cohort, the Area Under the Precision-Recall Curve (AUPRC) for XGBoost was 0.779 (95% CI, 0.771–0.794), significantly higher than logistic regression, Artificial Neural Network, and 1D-CNN but with no significant difference from LightGBM and Random Forest.1

Results of the group normalized AUROC, mean sensitivity, and mean specificity for each machine learning model in the derivation and validation cohorts by the deep ROC analysis showed XGBoost and LightGBM had favorable normalized group AUROC in all 3 groups divided by the false positive rate. The calibration plot demonstrated good calibration for all models, with the Brier Score ranging from 0.107 to 0.131.1

Investigators calculated SHAP values for the high-performing XGBoost, LightGBM, and Random Forest models and identified age, albumin, IgA/C3, and urine red blood cells as consistently being among the top 5 predictor variables across all 3 models. Further analysis revealed a negative correlation between age and the prediction of IgAN, while positive correlations were observed with albumin, IgA/C3, and urine red blood cells.1

Sensitivity analysis with 10-fold cross-validation across the entire dataset showed LightGBM achieved the highest AUROC (0.913; 95% CI, 0.906–0.921), significantly increased compared to logistic regression, Artificial Neural Network, and 1D-CNN, although not significantly different from XGBoost and Random Forest. The AUPRC for XGBoost was 0.785 (95% CI, 0.775–0.809), significantly higher than logistic regression, Artificial Neural Network, and 1D-CNN, with no significant difference from LightGBM and Random Forest.1

Investigators outlined several potential limitations to these findings, including the study’s reliance on data from a single center and lack of external validation; the limited sample size, particularly in the validation phase; and the inclusion of all patients undergoing kidney biopsy without a specific focus on those with clinical manifestations of IgAN like chronic glomerulonephritis.1

“This study demonstrated the utility of machine learning models using common clinical data in the diagnostic prediction of IgA nephropathy,” investigators concluded.1 “These models can be helpful for non-invasive and reliable methods to predict IgAN.”

References:

  1. Noda R, Ichikawa D, Shibagaki Y. Machine learning-based diagnostic prediction of IgA nephropathy: model development and validation study. Sci Rep. 2024;14(1):12426. doi:10.1038/s41598-024-63339-7
  2. American Kidney Fund. IgA Nephropathy. April 22, 2024. Accessed June 17, 2024. https://www.kidneyfund.org/all-about-kidneys/other-kidney-diseases/iga-nephropathy
  3. National Kidney Foundation. IgA Nephropathy. Accessed June 17, 2024. https://www.kidney.org/atoz/content/iganeph
Related Videos
4 KOLs are featured in this series
4 KOLs are featured in this series
M. Safwan Badr, MD: Novel Treatments for Central Sleep Apnea in Last 10 Years
Video 4 - Featuring 3 KOLs in, "Implementing Treat to Target in the Long-term in Inflammatory Bowel Disease "
Video 3 - Featuring 3 KOLs in, "How important is transmural healing as a treatment target in UC and CD?   Where does intestinal ultrasound fit in CD management?  "
How Elite Athletes Can Optimize Sleep for Peak Performance, with Jesse D. Cook, PhD
Boadie Dunlop, MD, Weighs in on FDA Advisory Vote on Lykos’ MDMA
HCPLive Five at ADA 2024 | Image Credit: HCPLive
© 2024 MJH Life Sciences

All rights reserved.