From Multi-Omics Prediction to Clinical Workflow: An Interoperable, Bias-Audited, Human-in-the-Loop Decision-Support Framework for Precision Oncology

Abstract

- Multi-omics survival models for precision oncology are increasingly capable of producing patient-level risk estimates, subtype projections and molecular explanations. Yet a model that performs adequately in retrospective validation is not automatically ready for clinical use. The translational problem is broader than model architecture: oncology teams must determine whether genomic risk predictions can be exchanged through interoperable health information systems, audited for subgroup bias, interpreted by clinicians, monitored for drift and governed as decision support rather than autonomous diagnosis. This article develops an interoperable, bias-audited and human-in-the-loop decision-support framework for multi-omics precision oncology, using lung adenocarcinoma (LUAD) as the applied case. The empirical basis combines four evidence layers: a TCGA-LUAD and MSK-IMPACT transformer manuscript, a DNA/RNA multi-omics survival thesis, an individualized LUAD patient report, and a reproducible secondary-data design linked to public TCGA/Kaggle and cBioPortal data pathways. The attached research record shows that a ridge RNA+clinical baseline achieved the strongest discrimination (C-index = 0.724), while the full DNA/RNA multi-omics model achieved lower aggregate discrimination (C-index = 0.682) but improved biological interpretability, temporal stability and clinical-decision value. The transformer system achieved moderate internal discrimination and modest external transportability, while still producing meaningful risk ordering and Integrated Gradients explanations. A simulated silent-mode workflow analysis then demonstrates how a standards-based clinical implementation layer can reduce review burden, improve missing-data controls, strengthen override documentation and surface subgroup-specific calibration risk before any prospective deployment. The paper argues that precision-oncology AI should be evaluated not only by C-index, but by an integrated evidence package: discrimination, calibration, decision utility, subgroup fairness