Dual-view cross-semantic graph neural network for predicting intraoperative complications in patients with acute myocardial infarction undergoing percutaneous coronary intervention.

For acute myocardial infarction (AMI) patients undergoing percutaneous coronary intervention (PCI), accurate prediction of intraoperative complications (IOCs) in AMI patients undergoing PCI is critical for timely risk stratification and clinical intervention. However, existing approaches lack a patient-indicator dual-view synergy modeling framework, leading to semantic isolation between patient and indicator representations. Yet current risk prediction approaches often treat patient characteristics and clinical indicators as separate, disconnected pieces of information. What's missing is a unified framework that jointly models both the 'patient' and the 'indicators' in a synergistic way, capturing how they interact and influence each other. Without this dual-view perspective, valuable clinical insights remain hidden, limiting the accuracy and real-world utility of risk predictions. Guided by a gain-aware feature distillation strategy, an indicator heterogeneous collaborative graph (IHCG) is constructed to disentangle inter-indicator semantics via meta-path-constrained propagation, thereby encoding structured clinical logic beyond raw statistical associations. Concurrently, a patient neighborhood topology graph (PNTG) is instantiated based on high-order phenotypic affinity modeling to capture contextualized risk signatures, with manifold-preserving neighborhood propagation enabling the encoding of population-level pathophysiological coherence beyond isolated clinical profiles. According to IHCG and PNTG, semantic integration across views is achieved through a patient-query asymmetric propagation paradigm. In the paradigm, patient nodes actively drive risk-informed semantic aggregation from clinical indicators via cross-attentional indicator fusion, while indicator nodes provide auxiliary contextual feedback through a lightweight residual semantic feedback pathway, thereby establishing a structured yet asymmetric information flow that aligns individualized risk assessment with population-level clinical semantics. To further ensure coherent bidirectional representation learning, aligning structurally projected embeddings with semantically cross-attended outputs. On the basis of the above modeling, this study develop a novel method for predicting IOCs in patients with AMI undergoing PCI called dual-view cross-semantic graph neural network (DCGNN), and the method resolves cross semantic isolation and grounds intraoperative risk prediction in mechanism-aware clinical reasoning. On the dataset collected from Shanghai Sixth People's Hospital South Campus, DCGNN achieved an AUC of 89.8% and a sensitivity of 90.2%, significantly outperforming baseline methods. Ablation studies, hyperparameter sensitivity analysis, and semantic-aware risk factor attribution (SARFA) further validated the effectiveness, robustness, and interpretability of the proposed method. The study method achieves superior predictive accuracy while enabling traceable risk attribution through dual-view asymmetric cross-semantic alignment, providing an intelligent decision support tool for perioperative risk stratification in AMI patients undergoing PCI with significant clinical translation potential.