OR WAIT null SECS
Investigators reported that AI programs such as LightGBM could potentially significantly benefit the diagnosis of Berger disease.
Machine learning could help provide improved immunoglobulin A (IgA) nephropathy (Berger disease) diagnostic improvements compared to conventional logistic regressions, according to new research.1
In data from a Japan-based trial, presented at the American Society of Nephrology (ASN) 2023 Kidney Week Annual Meeting in Philadelphia this month, a team of investigators reported that artificial intelligence (AI) programs such as LightGBM could potentially significantly benefit the diagnosis of Berger disease—enough so that the team recommended validation research support its clinical use going forward.
A number of research programs in the last decade has unearthed greater utility and potential with AI and machine learning in diagnostics, especially in chronic conditions. Data from earlier this year presented at the American Diabetes Association (ADA) 2023 Scientific Sessions showed a high-percentage point accuracy for diagnosing diabetic retinopathy progression with a novel machine learning model informed by patient imaging.2
Led by Ryunosuke Noda, MD, of the department of internal medicine at St. Marianna University School of Medicine in Kawasaki, investigators sought to interpret the diagnostic performance of machine learning algorithms for Berger disease.1
“IgA nephropathy often requires therapeutic modalities associated with potential complications, such as steroids, and so requires definitive diagnosis by invasive renal biopsy rather than non-invasive clinical diagnostic measures,” Noda and colleagues wrote. “Although the efficacy of machine learning for diagnostic purposes has been underscored in recent years, its application in the context of nephrology remains unclear.”
The team conducted a retrospective cohort study of 1419 cases in which patients underwent renal biopsies at the university hospital between January 2006 – September 2022. They excluded any cases wherein diagnoses were indeterminate and pathologies overlapped.
Investigators then randomly divided the remaining cases 8:2 into train and test datasets. Among the 44 explanatory variables used for the analysis were age at time of renal biopsy, gender, blood tests and urinalysis. Noda and colleagues evaluated 5 machine learning algorithms:
They identified the model with the highest mean Area Under the Curve (AUC) via stratified 5-fold cross-validation analyses in the train set. The model with the highest mean AUC was compared to the logistic regression model in the test set. Investigators used the Shapley Additive Explanations (SHAP) methodology to interpret the predictive outcomes of the algorithms.
Among the 5 machine learning models in the train set, LightGBM provided the highest mean AUC (0.92); it again provided a 0.92 AUC versus logistic regression (0.88) in the test set. What’s more, Noda and colleagues identified a number of explanatory variables linked to the prediction of IgA nephropathy, including urinary red blood cell count; serum albumin, IgA/C3 ratio; urinary protein / creatine ratio; and patient age.
“Our study indicated that machine learning, particularly the LightGBM model, could improve IgA nephropathy diagnostic performance beyond conventional logistic regression,” investigators concluded. “The influential variables identified were consistent with those reported in the existing literature. This highlights the potential utility of machine learning in IgA nephropathy diagnosis, necessitating further validation for clinical use.”