New research published in Journal of American Medical Informatics Association We demonstrate a framework for predicting end-stage renal disease (ESRD) in patients with chronic renal disease (CKD). The authors expressed optimism that such information could improve clinical decision-making through integrated multi-decoration and advanced analyses, and future research could further expand integration into other chronic disease spaces.1
Image credit: SANSERT | stock.adobe.com
CKD is a complex, multiple condition characterized by a gradual decline in renal function, which can ultimately progress to ESRD. The global prevalence of CKD ranges from about 8% to 16%, with approximately 5% to 10% of individuals diagnosed with CKD eventually reaching ESRD, representing a significant public health challenge. This is especially important in people with comorbidities such as diabetes and hypertension.1
To improve prediction of CKD progression to ESRD, researchers can utilize machine learning and deep learning models, and apply them to clinical trials and claims data using various observation windows (AIs) to enhance interpretability and reduce bias. Investigators used data from 10,326 patients with CKD between 2009 and 2018.1
After pre-processing, cohort identification, and functional engineering, the investigators evaluated multiple statistical machine learning and deep learning models using five different observation windows. The importance of functionality and Shapley Addive explanatory analysis were employed to understand key predictors. The models were specially tested for robustness, clinical relevance, misclassification patterns, and bias.1
The authors observe that the integrated data model is superior to the single data source model, with long-term short-term memory achieving the highest region under the receiver operating characteristic curve (AUROC; 0.93) and F1 score (0.65). Additionally, the 24-month observation window optimally balances the accuracy of early detection and prediction. The estimated glomerular filtration rate equation for 2021 improved prediction accuracy and reduced racial bias, particularly in Black patients.1
“Our study presents a robust framework for predicting ESRD outcomes and improving clinical decision-making through integrated multiple data and advanced analytics,” Lemapadman, a professor of management science and healthcare informatics at Carnegie Mellonsheinz College, said in a news release. “Future research will expand data integration and expand this framework to other chronic diseases.”2
If CKD progresses to ESRD, dialysis or implantation is required for patient survival. The economic impact of CKD is also important, with the relatively small proportion of CKD patients with CKD relative to US Medicare, which contributes a disproportionately high share of Medicare costs, particularly when going to ESRD. In particular, more than one-third of ESRD patients will be readmitted to the hospital within 30 days of discharge, highlighting the important need for early detection and management of CKD to prevent progression to ESRD, improve patient health outcomes, and reduce medical costs.1,2
In the study limits, investigators write that reliance on data from one agency may limit the generalizability of the model to other care settings. Additionally, the use of data from electronic health records may introduce observational bias, incomplete records, and underestimation of specific patient groups that may impair both accuracy and fairness in the assessment.1,2
“Our work fills the critical gap by developing a framework that uses integrated clinical and billing data rather than isolated data sources,” explained Yubo Li, a doctoral student at Carnegie Mellonsheinz College. “By minimizing the observation window required for accurate prediction, our approach balances patient-centered practicality with clinical relevance. This integration improves both predictive accuracy and clinical utility, allowing for more informational decisions to improve patient outcomes.”2
