---
title: "Cross-dataset CKD risk prediction with an explainable hybrid balancing ML framework"
id: "plos-one-4-an-explainable-machine-learning-framework-for-cross-dataset-chronic-kidney"
canonical_url: "https://medichelpline.com/clinical-feed/plos-one-4-an-explainable-machine-learning-framework-for-cross-dataset-chronic-kidney"
content_type: "clinical_feed_article"
specialty: "Nephrology"
source_name: "PLOS ONE (Medicine)"
source_url: "https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0362320"
published_at: "2026-11-05T14:00:00.000Z"
evidence_level: "Journal Feed"
license: "CC-BY-NC-4.0 / Informational Use"
---
# Cross-dataset CKD risk prediction with an explainable hybrid balancing ML framework
## Provenance & Clinical Metadata
- **Canonical URL:** https://medichelpline.com/clinical-feed/plos-one-4-an-explainable-machine-learning-framework-for-cross-dataset-chronic-kidney
- **Specialty:** [Nephrology](https://medichelpline.com/clinical-feed/nephrology.md)
- **Primary Source:** PLOS ONE (Medicine)
- **Source URL:** [Original Journal Publication](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0362320)
- **Published At:** 2026-11-05T14:00:00.000Z
- **Evidence Rating:** Journal Feed
## Executive GIST (TL;DR)
- The study developed an explainable hybrid machine-learning framework (HXAI-ML) for **chronic kidney disease** risk prediction designed for cross-dataset adaptation. It used two public datasets: the UCI Risk Factor Prediction dataset (source) and a Kaggle CKD dataset (target). - Preprocessing steps included label binarization, imputation, and standard scaling prior to model training and evaluation. The authors evaluated seven data-balancing methods and six classification algorithms; the source text did not enumerate every classifier or balancing method in full detail. - A target-domain adaptation protocol was applied: 10% of the target dataset was used for calibration and 90% reserved as a blind holdout for testing. This calibration-driven adaptation tested whether limited labeled target data could improve transferability. - Explainability methods used were **SHAP**, **LIME**, and Permutation Importance Analysis to interpret model predictions at both global and instance levels. The study emphasized interpretable outputs alongside predictive performance. - The combination labeled RO + TL + ETC (selected by accuracy) produced the top target-domain accuracy in the tested configurations. On the 1,494-sample blind holdout it achieved 91.97% accuracy, 92.24% precision, 99.64% recall, 95.80% F1-score and 67.52% ROC-AUC, per the source. - The same primary model showed extreme class-wise imbalance: specificity for the non-CKD class was 5.74%, and balanced accuracy was 52.69%, indicating poor discrimination for the minority class despite high recall for CKD. - Authors framed the results as methodological feasibility for cross-dataset adaptation with limited labeled target data, but they explicitly cautioned that the framework is exploratory and not a clinically deployable screening or decision-support system given limited ROC-AUC and low specificity. - Data availability: the two public datasets are listed in the source (Kaggle and UCI Machine Learning Repository). The authors reported no specific funding and declared no competing interests. - Primary methodological gaps noted by the study team include severe class imbalance in the target dataset (1,524 CKD vs. 135 non-CKD samples), cross-dataset distributional shifts, and the need for clearer minority-class representation for reliable adaptation. - The study did not report every hyperparameter choice, each classifier identity in full, or detailed per-class confusion-matrix counts beyond the summarized metrics reported in the article.
## Clinical Analysis & Structured Key Points
An explainable machine learning framework for cross-dataset chronic kidney disease risk prediction using hybrid data balancing | PLOS One Browse Subject Areas ? Click through the PLOS taxonomy to find articles in your field. For more information about PLOS Subject Areas, click here . Article Authors Metrics Comments Media Coverage Peer Review Reader Comments Figures Figures Abstract Chronic kidney disease (CKD) is a major global health problem, and early detection is important for reducing complications and improving treatment. Machine learning has shown strong potential for CKD prediction, but many existing studies rely on single datasets, face severe class imbalance, and provide limited interpretability. These challenges reduce model reliability and make cross-dataset use in real clinical settings more difficult. Therefore, this study aimed to develop an interpretable and adaptable machine learning framework for CKD risk prediction across different datasets. The proposed hybrid explainable machine learning (HXAI-ML) framework used two public CKD datasets: the UCI CKD Risk Factor Prediction dataset as the source dataset and the Kaggle CKD dataset as the target dataset. After preprocessing, label binarization, imputation, and standard scaling, seven balancing methods and six classifiers were evaluated. A target adaptation setting was applied by using 10% of the target dataset for calibration and 90% for blind testing. SHAP, LIME, and Permutation Importance Analysis were used to explain the model predictions. Among all tested combinations, RO + TL + ETC achieved the highest overall accuracy among the evaluated target-domain adaptation configurations and was therefore selected as the primary model under the accuracy-based selection criterion. On the 1,494-sample blind holdout it obtained 91.97% accuracy, 92.24% precision, 99.64% recall, 95.80% F1-score, and 67.52% ROC-AUC. However, its specificity of 5.74% and balanced accuracy of 52.69% indicate substantial class-wise imbalance in predictive performance. The findings demonstrate the methodological feasibility of cross-dataset target-domain adaptation and explainable machine learning under limited target-domain data. However, the relatively low ROC-AUC values and low specificity for the non-CKD class indicate limited discriminative ability. Therefore, the proposed framework should be regarded as an exploratory cross-dataset evaluation rather than a clinically deployable screening or decision-support system. Citation: Mumu JT, Hossain I, Tanvir Anzum KM (2026) An explainable machine learning framework for cross-dataset chronic kidney disease risk prediction using hybrid data balancing. PLoS One 21(11): e0362320. https://doi.org/10.1371/journal.pone.0362320 Editor: Jordan Llego, Healthcare and Technology Institute (HCT Academy), PHILIPPINES Received: June 18, 2026; Accepted: October 22, 2026; Published: November 5, 2026 Copyright: © 2026 Mumu et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability: The datasets used in this study are publicly available on 1. Kaggle as Chronic Kidney Disease Dataset https://www.kaggle.com/datasets/rabieelkharoua/chronic-kidney-disease-dataset-analysis 2. UCI Machine Learning Repository as Risk Factor Prediction of Chronic Kidney Disease: https://archive.ics.uci.edu/dataset/857/risk+factor+prediction+of+chronic+kidney+disease . Funding: The author(s) received no specific funding for this work. Competing interests: The authors have declared that no competing interests exist. 1. Introduction Chronic kidney disease (CKD) affects more than 850 million people worldwide, and approximately 1.5 million people die from CKD annually. This amounts to 1 death in every 20 seconds around the world. Cases of CKD are supposed to increase by 2050 [ 1 ]. CKD is another condition that poses a big economic challenge in the world. Approximately 93.1 million individuals worldwide in Europe have CKD. Kidney replacement therapy is implemented on approximately 750,000 patients. Within a year, the number of people dying as a result of CKD is approximately 210,000 [ 1 ]. According to World Kidney Day reports, there are severe impacts of kidney disease. There are at least 2.4 million deaths on kidney related disorders each year. Kidney disease is among the leading causes of death worldwide and is one of the causes of mortality showing substantial growth globally [ 2 ]. The development of kidney issues can have a number of causes. It may result in Acute Kidney Injury as a result of dehydration or low blood flow. The AKI can also be caused by infections, injuries, specific medicines, and kidney stones. Such conditions can also rapidly impair the functionality of kidneys [ 3 ]. Kidneys can also be damaged because of autoimmune diseases and exposure to harmful environmental factors. These factors can be identified early and lead to the correct prevention and treatment. Severe kidney complications can be diminished with the early treatment [ 4 ]. National kidney foundation classifies CKD in five stages. These are stages that are reliant on kidney abnormalities and their slowed rate of filtration. The Glomerular Filtration Rate is the measure of the availability of the kidneys to filter the blood [ 5 ]. CKD is a significant reason for heart disease and a serious cause of premature mortality [ 6 ]. Individuals are supposed to understand that CKD is treatable and manageable. Vital diseases are observable and treatment can be delayed. With appropriate management, the significant issues with kidney might be avoided [ 7 ]. CKD will lead to kidney failure without treatment. Thus, timely diagnosis of CKD is very critical. Machine Learning is useful in the identification of CKD at its initial stages. It uses patient medical data by identifying trends and risks. Such models examine the tests of blood as well as the history of the patients. They assist in monitoring kidney damage without showing serious symptoms [ 8 ]. Federated Learning is a privacy-conscious machine learning method. Without sending raw data, models flow through many locations. Between systems only model updates are exchanged. Sensitive medical data are secured during the training process [ 9 ]. Without patient-record sharing, hospitals will be able to establish stronger models. They also exchange model parameters as opposed to raw information. This would enhance teamwork and preserve the privacy of patients [ 10 ]. ML has already been put in healthcare in order to enhance the diagnosis of CKD through the analysis of complex medical data and identifying patterns that can be detected by doctors without using particular equipment [ 8 , 11 ]. Clustering algorithms like k means identify latent patterns of patient data. Such patterns allow finding subgroups of patients to treat them specifically [ 8 , 11 ]. Deep learning models are also used to enhance the level of diagnosis of CKD. Convolutional Neural Network is an analysis of medical imaginations like ultrasound or MRI. Sequential data on patients are recycled via Neural Networks. These models describe the complex medical relationships that is not a linear association when using medical data. They aide in identifying the irregularities of the images and temporal health data [ 8 , 11 ]. The study of CKD prediction has continued to develop over the last several years. A great deal of research features sophisticated machine learning codes. They are a Random Forest and XGBoost. Neural networks are also popular. Researchers combine ensemble learning models with feature selection. Such a method enhances the accuracy of prediction and reliability of the model [ 8 ]. Through these steps, by following them systematically, researchers were able to come up with strong and correct prediction models in case of chronic kidney disease. Combination of feature selection and proper machine learning algorithms enhanced better model result and high prediction accuracy. Hyper parameter optimization also enhances the model performance. Parameters such as grid search and random search are used to optimize. These methods determine the optimal model set ups. Raised models are embedded in clinical decision systems as part of some studies. These systems are used to ensure that doctors make enhanced and faster decisions whenever making medical choices [ 8 , 11 , 12 ]. 1.1 Technical challenges While doing the experiment, several technical difficulties appeared. They are following below: Severe Class Imbalance: The CKD datasets have a much larger number of CKD cases than non -CKD cases. This imbalance will place models in favor of the majority class. Oversampling and Intersecting Samples: Superimposition of clinical data results in duplicated data around the limits of classes. Such noisy samples cause confusion among models and absence of accurate prediction. Elements of Minimization of Target Data Adaptation: A trained model that has been trained using one dataset distribution could not work on a different dataset distribution. Customization of models using little new data is still technically difficult. Nonexistence of Model Interpretability: A lot of machine learning models are black boxes in the medical predictions. Automated predictions cannot be trusted before the clinical is clearly explained to clinicians. [ 13 ] 1.2 Problem statement Despite the increasing use of machine-learning methods for CKD prediction, several technical challenges limit their reliable application across datasets. In the target CKD dataset considered in this study, the class distribution is highly imbalanced, with 1,524 CKD samples compared with only 135 non-CKD samples. Such severe imbalance can bias conventional classifiers toward the majority CKD class and reduce their ability to correctly identify non-CKD cases. In addition, differences between the source and target datasets create a cross-dataset adaptation challenge, as a model trained on one dataset may not directly generalize to another dataset with different sample distributions and characteristics. The limited representation of the minority class further restricts reliable target-domain adaptation. Moreover, high predictive performance alone is insufficient for clinical applications when the underlying model decisions cannot be clearly interpreted. Therefore, there is a need for a framework that simultaneously addresses class imbalance, cross-dataset adaptation, and model interpretability. These challenges motivate the development of the proposed HXAI-ML framework, which combines data-balancing strategies, source-domain model optimization, target-domain adaptation, and explainable AI techniques to provide more transparent CKD predictions. Despite the increasing application of machine-learning techniques for CKD prediction, several challenges remain, including severe class imbalance, limited availability of representative target-domain samples, variation between source and target datasets, and insufficient interpretability of model predictions. Conventional learning approaches may become biased toward the majority class and may not adequately distinguish CKD from non-CKD cases. Furthermore, models that provide high predictive performance without transparent explanations can limit their usefulness in clinical decision-making. Therefore, there is a need for a robust framework that integrates effective class-imbalance handling, cross-domain adaptation, and explainable AI to improve CKD prediction while providing transparent and clinically interpretable model decisions. 1.3 Research question RQ1: Which data-balancing technique, including oversampling, SMOTE, and hybrid balancing approaches, provides the most effective handling of class imbalance for CKD prediction? RQ2: Can XAI techniques explain how ML model predicts CKD risk? RQ3: Can target-domain adaptive generalization enable CKD models to transfer reliably across different medical datasets with only limited labeled target data? To address these questions, this study has some objectives. They are described in the next sub-section. 1.4 Objectives The study intended to address the challenges faced by the objectives following below: To identify the most effective data balancing technique for improving CKD prediction on imbalanced datasets. To develop an interpretable machine learning framework using XAI techniques for CKD risk prediction. To evaluate the ability of target adapting cross-dataset learning to generalize CKD prediction across different medical datasets. The remainder of the article is organized as follows: Section 2 reviews the research literature regarding the prediction of chronic kidney disease. Section 3 provides an overview of the suggested methodology and explains it. Section 4 provides an evaluation study of the suggested framework. Section 5 evaluates our method, explains the study topics, and discusses the key findings, limitations, and recommendations. Section 6 concludes the paper with a summary of the paper’s contributions. 2. Literature review Chronic Kidney Disease (CKD) is a condition which gradually deteriorates the performance of kidneys and is colossal health challenge in the world platform. The markers in the blood and urine estimated glomerular filtration rate (eGFR) and urine albumin-to-creatinine ratio (uACR)) tend to enable the physicians to diagnose CKD and also give insights on the progression. These are useful in the classification of CKD though they are not that specific and they can sometimes fail to capture the complex biology of the force behind the disease [ 14 ]. 2.1 Chronic kidney disease clinical risk determinants In the recent research, the inflammation has been found to be significant in CKD. In case the filters are broken, a type of protein Albumin is commonly released into the urine by the blood. This will not be allowed to happen in a healthy kidney. CKD is the consequence of high blood pressure, which may also occur due to CKD [ 15 ]. Imflammatory biomarkers are indicative of deterioration of kidney function and end-stage kidney disease. They include Tumor Necrosis Factor Receptor 1, Tumor Necrosis Factor Receptor 2 and Interleukin-6 [ 16 , 17 ]. Such inflammatory mechanisms also lead to the scar tissue and the destruction of the blood vessels in the kidney [ 14 ]. eGFR figure illustrates that your blood is being cleansed by your kidney. When the eGFR levels exceed 90, then the kidney is fine. When the eGFR is below 60, it implies that you have CKD [ 18 ]. Another test ordered by the doctor is urine test to determine the functionality of the kidneys since the kidneys produce urine. When the urine is filled with blood and protein, that is an indication that your kidney is not functioning properly [ 19 ]. Doctor takes blood pressure since Blood pressure range indicates the manner in which your heart is pushing blood. When the amount of eGFR at less than one hundred and fifty, then the patient has end-stage kidney disease. Now all that is available is treatment: (i) dialysis, (ii) kidney transplant. In case of dialysis failure, the physician is left with the only option, i.e., transplantation of kidney. But it is so very costly [ 20 ]. CKD is also caused by oxidative stress. When the body is over-production of reactive oxygen species (ROS) then the cells are destroyed, inflammation develops, and scar tissue is deposited in the kidneys. Measures of such oxidative stress like urinary 8 oxiogo, F2 isoprostanes have also been studied to establish the possibility of PC measurement of early kidney problems and predict long-term albumin in the urine and kidney disease [ 14 ]. 2.2 Machine learning in CKD prediction Machine-learning methods have increasingly been applied to CKD prediction, with studies evaluating different classification algorithms and feature-selection strategies. Polat et al. [ 21 ] developed a Support Vector Machine (SVM)-based CKD prediction model using the UCI CKD database and reported an accuracy of 98.5%. Their findings demonstrated the importance of selecting informative features for improving predictive performance [ 21 ]. Medicine is being considered the field where new computer programs that learn data are being applied. They are able to read a plethora of patient records, identify patterns and assist in predicting and diagnosing diseases [ 22 ]. Almasoud et al. [ 23 ] have experimented with various algorithms in computers, such as the Logistic Regression, Random Forest, SVM and Gradient Boosting, to identify CKD. They discovered that Gradient Boosting was the most accurate and correctly identified 99.1% [ 23 ]. Random Forest would easily be applied in CKD prediction since it can deal with complex medical data. Several algorithms were used to come up with a prediction model of end stage kidney disease. Han et al. expressed that, Han et al. [ 24 ] evaluated several algorithms for predicting end-stage kidney disease and reported that Random Forest achieved the highest performance, with an accuracy of approximately 93.97%. [ 24 ]. Qin et al. [ 25 ] developed a composite model combining Logistic Regression and Random Forest for CKD progression prediction and reported an accuracy of 99.83%, suggesting that combining complementary models may improve predictive performance. [ 25 ]. Farjana et al. [ 26 ] compared multiple machine-learning algorithms for CKD prediction and identified LightGBM as the best-performing algorithm, achieving approximately 99% accuracy and a high AUC [ 26 ]. Similarly, Raihan et al. experimented an XGBoost predictive model with SHAP explanation and used to predict CKD. It was used to identify the accurate explanation to a 99.16% and demonstrates the significance of explainable models in healthcare [ 27 ]. Overall, previous studies demonstrate the strong predictive potential of machine-learning algorithms for CKD prediction, particularly ensemble and boosting methods. However, many existing studies rely on relatively small datasets or evaluate models within a single dataset, which may limit their generalizability to different populations and data distributions. These limitations highlight the need to investigate model robustness across datasets and to address class-imbalance and cross-dataset adaptation challenges [ 22 , 28 ]. 2.3 Explainable AI in CKD risk prediction In addition to predictive performance, the interpretability of machine-learning models has become an important consideration in healthcare applications. Many predictive models operate as black boxes, making it difficult for clinicians to understand the basis of individual predictions [ 29 ]. The Explainable Artificial Intelligence or XAI is a new research area which acquires a growing method in AI [ 30 ] of making these predictions more understandable. XAI methods including SHapley Additive explanations and Local Interpretable Model-agnostic Explanations are employed by researchers in explaining the model of CKD prediction. Ghosh and Khandoker [ 31 ] evaluated Logistic Regression, Random Forest, and XGBoost models for CKD prediction and examined the clinical features contributing to their predictions. Important features included creatinine, glycated hemoglobin (HbA1C), and patient age [ 31 ]. These predictions were explained using SHAP and LIME. SHAP was also used by Arjaria et al. when explaining CKD detection models. Their model was based on Logistic Regression to predict CKD. The analysis demonstrated the relevance of strides that define the presence of CKD. They consisted of serum creatinine, hemoglobin, albumin, and random blood glucose analyzes [ 32 ]. Gupta et al. [ 33 ] proposed a hybrid CKD prediction framework that combined ensemble learning with the Grey Wolf Optimizer and reported an accuracy of approximately 98.75%. This study demonstrates the
## Related Clinical Research

- [KDOQI guideline adherence and reporting plausibility in Bangladeshi hemodialysis patients](https://medichelpline.com/clinical-feed/plos-one-3-dietary-intake-patterns-and-kdoqi-guideline-compliance-among-hemodialysis.md)
- [Risk factors for 90-day RRT failure after cardiac surgery AKI treated with CRRT](https://medichelpline.com/clinical-feed/pubmed-42736125.md) (DOI: 10.3760/cma.j.cn112137-20260204-00393)
- [Transportable AKI Prediction with an Explainable XGBoost Model Using MIMIC‑IV](https://medichelpline.com/clinical-feed/medrxiv-2-toward-transportable-acute-kidney-injury-prediction-an-explainable-xgboost.md)
- [Long-term Renal Outcomes After Neonatal and Pediatric Renal Vein Thrombosis](https://medichelpline.com/clinical-feed/medrxiv-15-renal-outcomes-in-survivors-of-neonatal-and-pediatric-renal-vein-thrombosis.md)
- [PDK4 and Acute Kidney Injury: Metabolic Control of AKI and the AKI‑to‑CKD Transition](https://medichelpline.com/clinical-feed/pubmed-43059392.md) (DOI: 10.3892/mmr.2026.14075)

## Navigation
- [← Back to Nephrology Feed](https://medichelpline.com/clinical-feed/nephrology.md)
- [← All Clinical Specialties](https://medichelpline.com/clinical-feed.md)
## Medical & Regulatory Disclaimer

> [!CAUTION]
> MedicHelpline content is structured for research, educational, and professional discovery purposes. It does not constitute individual medical advice, clinical diagnosis, or treatment recommendations.
> Always verify dosing, contraindications, and regulatory alerts against official product labeling and primary regulatory sources before clinical decision-making.