FedMM-CKD: Federated Multimodal Sensor-EHR Fusion with Cross-Attention Transformer and Contrastive Alignment for Early Chronic Kidney Disease

Abstract

Chronic kidney disease (CKD) is a global health emergency affecting over 850 million individuals worldwide, with more than 90% of early-stage cases remaining undetected until irreversible renal decline necessitates dialysis or transplantation. Early detection of reduced glomerular filtration rate (eGFR < 60 mL/min/1.73 m2) from routinely collected clinical and wearable sensor data represents a transformative opportunity for scalable AI-driven population screening. However, the decentralized, heterogeneous, and privacy-sensitive nature of clinical data imposes fundamental barriers on centralized model training. This paper presents FedMM-CKD, a novel federated multimodal learning framework integrating 12-variable structured EHR features with 7-day wrist actigraphy sensor sequences through a cross-attention Transformer fusion architecture, trained across five simulated non-IID clinical clients using the publicly available NHANES 2011-2012 dataset. The framework incorporates: (1) FedAvg aggregation augmented with FedProx proximal regularization for non-IID client drift mitigation; (2) InfoNCE contrastive cross-modal alignment enforcing representation consistency between EHR and sensor embeddings; (3) sinusoidal positional encoding for temporal sensor sequences; and (4) mean-imputation with z-score normalization for robust missingness handling. FedMM-CKD achieves a global test AUROC of 0.9910 and AUPRC of 0.9020 after 10 communication rounds, outperforming all 35 compared state-of-the-art methods from 2021 to 2026 by at least 9.0 AUROC percentage points. Ablation studies confirm that cross-modal attention and contrastive alignment each contribute critically, with the full model improving AUROC by 16.0 percentage points over a sensor-only baseline.