A Multimodal AI Framework for Chronic Kidney Disease Prediction and Clinical Decision Support

Abstract

Chronic Kidney Disease (CKD) is often detected late, when the disease has already reached an advanced stage. Many existing prediction models rely on a single source of data, which may limit their ability to capture the full health profile of a patient. This research addresses this limitation by proposing a multimodal AI-based framework that combines complementary health data sources for CKD risk prediction and clinical decision support. The framework was built using harmonised NHANES data from 1988 to 2018. The selected predictors were organizd into three modalities: clinical variables, biochemical laboratory variables, and environmental/dietary variables. TabNet models were trained separately on each modality to evaluate their individual contribution. Early and late fu sion strategies were subsequently implemented to assess the benefit of combining modal ities. A temporal validation strategy was used to simulate real-world deployment, where models trained on past data must generalize to future patients. The multimodal approach outperformed single-modality baselines. The laboratory modal ity was the strongest individual predictor, while clinical variables provided meaningful complementary information. The best overall performance was achieved by the late fusion stacking model, with an AUROC of 0.9450, an AUPRC of 0.7004, and a sensitivity of 90.4% on the temporal test set. Explainability analysis using TabNet attention and SHAP showed that the predictions were based on clinically meaningful features. These results were integrated into NephroShield, a web-based clinical decision-support system that provides CKD risk prediction, feature-level explanations, automated PDF input ex traction, and LLM-generated clinical summaries.