Machine-Learning Models Able to Predict Risk of Renal Function Decline
New research indicates that multiple machine-learning models can predict the risk of renal function decline on par with conventional methods using routine clinical data.
Source: Getty Images
By - Research has found that machine-learning (ML) models can use standard clinical data to predict renal function decline (RFD) with similar accuracy to that of traditional prediction methods.
In a new study published in Risk Management and Healthcare Policy, researchers trained and tested seven ML models to predict the risk of RFD. Then, they evaluated the models’ performances against that of a logistic regression model to determine whether ML methods could potentially be used for RFD prediction in clinical settings.
The researchers trained the ML models using retrospective data from 2,166 patients aged 35 to 74 collected between 2010 and 2020. From this data, 24 independent variables were extracted, with baseline estimate glomerular filtration rate (eGFR) being used to determine patient RFD status. The patient cohort was divided randomly into a training set and a test set, and the datasets were given to the models.
The seven ML models — random forest, gradient boosting, multilayer perceptron, support vector machine, K-nearest neighbors, adaptive boosting, and decision tree — achieved significant predictive accuracy overall. Among all algorithms, the gradient boosting model showed the best average predictive accuracy. Of the 24 factors measured, age, serum creatinine (Scr), serum uric acid (SUA), waist circumference, and systolic blood pressure (SBP) were the most important for prediction model performance.
All ML models, except for the K-nearest neighbors and decision tree algorithms, improved RFD prediction performance compared to the logistic regression model. However, the improvements were insignificant, indicating that more research is needed to determine whether significant improvements are possible.
The findings indicate that an ML algorithm for RFD prediction has the potential for clinical use and could improve disease surveillance and care management for those at risk, the researchers posited.
Other research has also shown how ML models can be used as predictive tools for kidney-related conditions.
In a recent study, researchers developed an algorithm to determine treatment response in lupus nephritis patients. Lupus nephritis is an inflammatory kidney disease that can cause scarring and fibrosis of the kidneys. Without proper treatment, the disease may lead to infection, stroke, heart attack, or death.
Standard treatment for lupus nephritis involves one of two immunosuppressive drugs being prescribed for six months, at which time a physician assesses how the patient has responded to the treatment. If the treatment is ineffective, second-line drugs are available for use.
Clinicians face challenges when treating these patients as second-line drugs are expensive, and they are tasked with predicting the likelihood that the second type of treatment will be effective.
The researchers developed an algorithm that uses the International Society of Nephrology/Renal Pathology Society biopsy scores for activity, chronicity, interstitial fibrosis, and interstitial inflammation as well as urine protein-to-creatinine ratio, white blood cell count, and hemoglobin level to help clinicians identify if a patient will respond to second-line treatment or require more frequent monitoring.
Initial tests of the model’s predictive ability were promising, but the researchers noted that further testing and validation would be needed before the tool could be fully implemented.
