Chronic kidney disease (CKD) is a progressive and largely irreversible condition that often remains undetected until advanced stages, and its clinical assessment depends jointly on disease stage, glomerular filtration rate (GFR), and serum creatinine. Most published machine learning models predict only one of these targets and therefore ignore the physiological dependencies linking them. This study introduces an attention-based multi-task learning (ABMTL) framework that predicts all three simultaneously from structured clinical data. A shared representation backbone learns common patterns of renal dysfunction, while three task-specific attention modules re-weight the shared latent features for each objective and feed independent classification and regression heads trained under a joint loss. A leakage-free preprocessing protocol combined correlation-driven adaptive imputation, smoothed target encoding, and training-set-only normalization. The framework was evaluated on 13,500 patient records comprising 13 clinical features, and compared against seven machine learning and deep learning baselines under identical partitions. ABMTL achieved 95.55% accuracy for six-class CKD staging, with coefficients of determination of 0.966 for GFR and 0.976 for serum creatinine, outperforming every baseline on all metrics. Ablation experiments confirmed that the attention modules and multi-task optimization contribute complementary gains, and paired testing indicated the improvements were statistically significant. Kernel Shapley additive explanations analysis identified albumin, blood urea, hemoglobin, hypertension, and diabetes as the dominant predictors, consistent with established nephrology. Serum creatinine and GFR were used only as prediction targets and never as inputs. At the clinically relevant cut-offs of stage ≥3 and stage 4–5, sensitivity exceeded 97% and specificity exceeded 99%. Because the design is single-center, retrospective, and cross-sectional, and calibration and decision-curve analyses have not yet been performed, the framework is presented as a candidate screening-support model that requires external, prospective, and outcome-based validation before clinical use.