MLP-CKD

Objective To develop and validate a clinically informed deep learning framework for admission laboratory-based screening and risk stratification of uremia-associated advanced renal dysfunction. Methods We propose MLP-CKD, a multi-branch feature interaction-enhanced multilayer perceptron that groups laboratory variables by clinical domain and incorporates missingness indicators to capture test-availability patterns. Adult ICU patients were extracted from MIMIC-IV, and laboratory measurements obtained within the first 24 h after ICU admission were summarized using mean, minimum, and maximum values. To prevent information leakage, only the first eligible ICU stay per patient was retained, and all model development was conducted using patient-level training, validation, and independent test splits. MLP-CKD was compared against conventional machine learning models, neural network ablation variants, and clinical rule-based baselines. Performance was assessed using discrimination, calibration, decision curve analysis, bootstrap-based 95% confidence intervals, and permutation importance. Results On the independent patient-level test set, MLP-CKD achieved an AUC of 0.913 (95% CI: 0.907–0.919) and an AUPRC of 0.848 (95% CI: 0.838–0.859), representing the best discrimination among the primary models. Its recall and F1-score were 0.796 and 0.764, respectively. Ablation analysis showed that the Branch-only MLP achieved near-equivalent discrimination (AUC 0.913; paired-bootstrap p = 0.952), indicating that clinically informed feature grouping was the main performance driver, whereas the interaction module provided only incremental benefit. MLP-CKD outperformed XGBoost (Delta AUC = 0.010, p < 0.001), random forest (Delta AUC = 0.016, p < 0.001), and the eGFR rule-based baseline (Delta AUC = 0.047, p < 0.001). Feature-restricted analyses suggested that the model integrated complementary laboratory information beyond creatinine and BUN alone. Conclusion MLP-CKD provides a clinically structured framework for admission laboratory-based screening and risk stratification of uremia-associated advanced renal dysfunction. The model achieved competitive discrimination and clinically plausible feature-importance patterns, with its main advantage arising from clinically informed feature grouping. Because this retrospective study used admission-window laboratory variables without a predefined future prediction horizon, MLP-CKD should be interpreted as a decision-support tool rather than as a standalone diagnostic system or a model for predicting disease onset

Confidence interval · Deep learning · F1 score · Missing data · Multilayer perceptron · Random forest · Receiver operating characteristic · Risk assessment · Risk Stratification · Acute Kidney Injury Research · Chronic Kidney Disease and Diabetes · Dialysis and Renal Disease Management

• Machine Learning in Medicine

  Alvin Rajkomar, Jeffrey Dean et al.•New England Journal of Medicine•2019

  Incluir item

• A guide to deep learning in healthcare

  Open Access•Andre Esteva, Alexandre Robicquet et al.•Nature Medicine•2019

  Incluir item

• High-performance medicine

  Open Access•Eric J Topol•Nature Medicine•2019

  Incluir item

• Deep learning

  Open Access•Yann LeCun, Yoshua Bengio et al.•Nature•2015

  Incluir item

• Forecasting cardiovascular disease mortality using artificial neural networks in Sindh, Pakistan

  Open Access•Moiz Qureshi, Khushboo Ishaq et al.•BMC Public Health•2025

  Incluir item