The investigators tested an adaptation pipeline that trained on the UCI CKD Risk Factor Prediction dataset and adapted to a Kaggle CKD target dataset. They applied a target adaptation split using 10% of the target data for calibration and 90% for blind testing. Across the evaluated target-domain configurations, the combination designated RO + TL + ETC achieved the highest overall accuracy under the study's accuracy-based selection rule. On the withheld 1,494-sample blind holdout the selected configuration produced 91.97% accuracy, 92.24% precision, 99.64% recall, 95.80% F1-score and 67.52% ROC-AUC, as reported by the authors.
Short sentence: accuracy was high. Longer context: despite those strong-sounding accuracy and recall figures, the model’s class-wise performance was markedly uneven, and that imbalance materially affects real-world interpretability and potential clinical use.
The study confronted marked class imbalance in the target dataset; the source text reports 1,524 CKD samples versus 135 non-CKD samples. To address this, investigators evaluated seven balancing methods alongside six classifiers. The manuscript describes hybrid data-balancing as a core feature of the HXAI-ML framework, but the source did not enumerate every balancing algorithm and classifier with full hyperparameter detail. After preprocessing steps (label binarization, imputation, and standard scaling), the team compared oversampling variants, SMOTE-style techniques, and hybrid approaches to determine which improved cross-dataset generalization.
Short observation: class imbalance remained central. Longer point: even with hybrid balancing and target calibration, the selected model showed extreme skew in predicting the minority non-CKD class, so balancing methods improved some metrics but did not resolve per-class discrimination entirely.
The authors integrated explainability into the pipeline using SHAP, LIME, and Permutation Importance Analysis. These methods were applied to both global and instance-level predictions to provide interpretable feature contributions alongside model outputs. Investigators reported use of these XAI techniques to inspect how features influenced risk predictions, and to support transparency in cross-dataset adaptation experiments.
Short sentence: explanations were generated. Longer sentence: the article emphasizes that explainability was treated as a necessary complement to performance evaluation because high predictive accuracy alone is insufficient when model decisions must be communicated to clinicians; however, the source did not provide exhaustive examples of individual SHAP or LIME plots within the text provided here.
The primary model’s metrics illustrate the trade-offs encountered. Recall (99.64%) and F1-score (95.80%) were elevated, signaling strong sensitivity for the CKD class. Precision (92.24%) and overall accuracy (91.97%) also appear favorable on the blind holdout. The investigators also reported a comparatively modest ROC-AUC of 67.52% and a specificity of 5.74% for the non-CKD class, yielding a balanced accuracy of 52.69%.
Short analysis: sensitivity favored the majority class. Longer clinical note: this pattern means the model will label most patients as CKD, capturing nearly all true CKD cases but failing to reliably exclude non-CKD individuals—an outcome that undermines screening specificity and risks high false-positive rates if deployed clinically.
The datasets used are public: the Kaggle Chronic Kidney Disease Dataset and the UCI Risk Factor Prediction of Chronic Kidney Disease dataset, as cited by the authors. Preprocessing included label binarization, imputation of missing values, and standard scaling. The experiment assessed seven balancing techniques and six classifiers; the source did not list every algorithm and hyperparameter in the provided excerpt. The target adaptation approach used 10% labeled target data for calibration and 90% as a blind test set, specifically to evaluate adaptation with limited labeled target-domain examples.
The authors explicitly framed the work as an exploratory cross-dataset evaluation rather than a clinically deployable system. They reported methodological feasibility for target-domain adaptation with limited labeled data but also acknowledged important performance shortcomings—particularly the low specificity (5.74%) and modest ROC-AUC (67.52%)—which limit discriminative ability. The source text also emphasized technical challenges that persisted: severe class imbalance, potential noisy or overlapping samples near class boundaries, and cross-dataset distributional differences that complicate transfer.
Short closing note: the study stops at methodological feasibility. Longer final question: unanswered practical items include whether additional balancing strategies, richer non-CKD representation, alternative calibration fractions, or different classifier ensembles could materially improve specificity and ROC-AUC; these remain open for subsequent work as noted by the authors.