Abstract
<title>Abstract</title> <p> <bold>Purpose:</bold> Integrating heterogeneous multi-omics data for cancer subtyping remains a challenging problem due to differences in data dimens ionality, measurement scales, and noise characteristics across molecular platforms. Conventional graph-based integration methods rely on static similarity measures in general and have inadequate capabilities to distinguish biologically meaningful relationships, thereby limiting clustering robustness and clinical interpretability. This study proposes a confidence-guided graph reinforcement framework for constructing robust patient similarity networks from heterogeneous molecular data. <bold>Methods:</bold> A novel <bold>Multi-Criteria Graph Reinforcement Optimization (MCGRO)</bold> framework is presented to integrate multi-Omicdata through adaptive graph refinement. The proposed methodology introduces a hierarchical confidence estimation strategy based on a novel neighbourhood-consistency metric (Dst). Global confidence scores are used to assign adaptive weights to individual omics datasets, while local confidence scores quantify the reliability of individual graph edges for iterative graph reinforcement and confidence-guided pruning prior to spectral clustering . <bold>Results:</bold> The proposed confidence-guided reinforcement strategy generated more robust patient similarity graphs by strengthening biologically consistent neighbourhoods while suppressing unreliable graph connections. Experimental evaluation demonstrated improved subtype separation and clinically meaningful survival stratification across multiple cancer cohorts. Functional enrichment analyses further confirmed that the identified molecular subtypes were associated with biologically relevant pathways involved in tumour progression, immune regulation, and cellular signalling. <bold>Conclusion:</bold> MCGRO provides a robust computational framework for confidence-aware graph refinement and multi-omics integration. By combining neighbourhood-consistency estimation with adaptive graph reinforcement and confidence-guided edge optimization, the proposed framework improves the robustness, and interpretability of graph-based cancer subtype discovery. The modular design of the framework enables its application to heterogeneous biomedical datasets and provides a flexible computational foundation for future precision medicine studies. </p>