CARE-Net: Causal-aware risk embedding for venous thromboembolism prediction in orthopedic inpatients

Venous thromboembolism is a major preventable complication among orthopedic inpatients, yet existing risk scores and correlation-driven models often miss complex physiological interactions and provide limited interpretability. We propose Causal-Aware Risk Embedding Network (CARE-Net), a causal representation learning framework for VTE prediction on structured clinical data. CARE-Net first infers a directed causal graph among laboratory, demographic, and therapeutic variables, then performs message passing strictly along causal directions to construct mechanism-aligned patient embeddings. A causal contrastive objective further aligns patients with similar causal signatures, enhancing robustness and suppressing spurious associations. Extensive comparisons with statistical, ensemble, deep tabular, transformer-based, and graph-based baselines show that CARE-Net delivers consistently superior discrimination and a more balanced sensitivity-specificity profile. Ablation and feature-importance analyses confirm that each causal component contributes meaningfully and that learned risk factors align with established clinical pathways. These findings suggest that embedding causal structure into representation learning offers a principled route to reliable VTE decision support in orthopedic care.