<p>Early identification of chronic kidney disease (CKD) can prevent progression to renal failure, yet effective screening outside specialist settings remains limited. We present an Internet-of-Things (IoT) platform coupled with a leakage-free machine-learning (ML) pipeline for periodic remote CKD risk scoring. Using a retrospective dataset of electronic health records from 491 adults (21 variables) as a Proof-of-Concept (PoC) to simulate IoT-transmitted patient data, we compare seven classifiers combined with supervised feature selection (FS: Chi-square, ANOVA, Mutual Information) and dimensionality reduction (DR: PCA, UMAP) after imputation, scaling, and class-imbalance handling via model weighting. Performance is estimated with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(10\times \)</EquationSource> </InlineEquation> repeated, stratified 10-fold cross-validation and a nested threshold selection tailored to screening. To prioritize case detection, we evaluate models at a Screening Operating Point that maximizes recall while maintaining high specificity (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\approx 80\%\)</EquationSource> </InlineEquation>), accepting the trade-off of modest precision. Under this clinically aligned policy, Logistic Regression (LR) provides the best sensitivity–specificity balance across FS/DR settings, with recalls in the 0.773-−0.843 range and specificities 0.791-−0.803 (e.g., ANOVA FS: recall <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(0.843\pm 0.147\)</EquationSource> </InlineEquation>, specificity <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(0.796\pm 0.079\)</EquationSource> </InlineEquation>, ROC–AUC <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(0.880\pm 0.050\)</EquationSource> </InlineEquation>; PCA: recall <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(0.827\pm 0.137\)</EquationSource> </InlineEquation>, specificity <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(0.803\pm 0.074\)</EquationSource> </InlineEquation>, ROC–AUC <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(0.890\pm 0.045\)</EquationSource> </InlineEquation>). At this operating point, precision remains modest (PPV <InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(\approx 0.33\)</EquationSource> </InlineEquation>–0.35), reflecting a substantial number of false positives, whereas negative predictive value is consistently high (typically <InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(\approx 0.96\)</EquationSource> </InlineEquation>–0.97). For comparison, at a standard F1-optimal operating point, gradient-boosting ensembles increase precision at the cost of recall (e.g., CatBoost + Chi-square: recall <InlineEquation ID="IEq11"> <EquationSource Format="TEX">\(0.573\pm 0.202\)</EquationSource> </InlineEquation>), underscoring the importance of aligning the operating threshold with clinical priorities. Overall, methodological alignment outweighs model complexity for screening utility. Coupled with our scalable IoT architecture, a simple, highly discriminative LR + FS/PCA pipeline—operated under a specificity constraint—offers a low-cost, deployable solution for proactive CKD management.</p>

Iot enabled machine learning for early chronic kidney disease detection with a comparative study of feature selection and dimensionality reduction

期刊论文
出版模式：

开放获取
发表日期：

2026年4月7日

查看全文

El Mehdi Chouit,
Mohamed Rachdi,
Mostafa Bellafkih,
Brahim Raouyane

摘要

Early identification of chronic kidney disease (CKD) can prevent progression to renal failure, yet effective screening outside specialist settings remains limited. We present an Internet-of-Things (IoT) platform coupled with a leakage-free machine-learning (ML) pipeline for periodic remote CKD risk scoring. Using a retrospective dataset of electronic health records from 491 adults (21 variables) as a Proof-of-Concept (PoC) to simulate IoT-transmitted patient data, we compare seven classifiers combined with supervised feature selection (FS: Chi-square, ANOVA, Mutual Information) and dimensionality reduction (DR: PCA, UMAP) after imputation, scaling, and class-imbalance handling via model weighting. Performance is estimated with \(10\times \) repeated, stratified 10-fold cross-validation and a nested threshold selection tailored to screening. To prioritize case detection, we evaluate models at a Screening Operating Point that maximizes recall while maintaining high specificity ( \(\approx 80\%\) ), accepting the trade-off of modest precision. Under this clinically aligned policy, Logistic Regression (LR) provides the best sensitivity–specificity balance across FS/DR settings, with recalls in the 0.773-−0.843 range and specificities 0.791-−0.803 (e.g., ANOVA FS: recall \(0.843\pm 0.147\) , specificity \(0.796\pm 0.079\) , ROC–AUC \(0.880\pm 0.050\) ; PCA: recall \(0.827\pm 0.137\) , specificity \(0.803\pm 0.074\) , ROC–AUC \(0.890\pm 0.045\) ). At this operating point, precision remains modest (PPV \(\approx 0.33\) –0.35), reflecting a substantial number of false positives, whereas negative predictive value is consistently high (typically \(\approx 0.96\) –0.97). For comparison, at a standard F1-optimal operating point, gradient-boosting ensembles increase precision at the cost of recall (e.g., CatBoost + Chi-square: recall \(0.573\pm 0.202\) ), underscoring the importance of aligning the operating threshold with clinical priorities. Overall, methodological alignment outweighs model complexity for screening utility. Coupled with our scalable IoT architecture, a simple, highly discriminative LR + FS/PCA pipeline—operated under a specificity constraint—offers a low-cost, deployable solution for proactive CKD management.