Integrating Patient Clinical Risk Factors, Serum Biomarkers, and Compact Depthwise-Separable CT Embeddings for Calibrated Multi-Class Renal Disease Stratification

• Abilly Elly

  Texas University

  Author

Chronic kidney disease affects nearly 800 million adults globally and remains among the few leading causes of death with rising mortality rates. Despite advances in renal imaging and biomarker discovery, existing computational models for renal disease stratification suffer from critical limitations: they rely predominantly on single-modality data, lack rigorous patient-disjoint validation protocols, and rarely address probability calibration—a prerequisite for clinical decision support. This study develops and validates a multimodal framework that integrates patient clinical risk factors, serum biomarkers, and compact depthwise-separable CT embeddings for calibrated multi-class renal disease stratification. Using a retrospective cohort design, we extracted 12,446 CT images across four diagnostic categories (Normal, Cyst, Tumor, Stone) and integrated these with clinical variables and serum biomarker panels. A depthwise-separable CNN backbone, adapted from the NephroNet architecture, generated image embeddings that were fused with tabular clinical and biomarker features. Post-hoc temperature scaling addressed calibration. The framework achieved 89.4% accuracy (95% CI: 87.8–91.0%) with an expected calibration error of 0.031, substantially outperforming imaging-only (82.1%) and clinical-only (76.8%) baselines. Feature importance analysis identified age, eGFR, FGF-23, and the CT embedding dimension corresponding to cortical morphology as top predictors. These findings demonstrate that calibrated multimodal integration significantly improves renal disease stratification accuracy and reliability, offering a replicable framework for clinical deployment.

• pdf

Copyright (c) 2026 Abilly Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.