Uncertainty-Aware Multimodal and Multiscale Treatment Response Prediction in Advanced Non-Small Cell Lung Cancer

Abstract

Managing patients with metastatic non-small cell lung cancer (mNSCLC) remains challenging due to highly variable responses to standard of care chemoimmunotherapy, with many patients experiencing toxicity and limited clinical benefit. These variable responses are associated with spatially complex heterogeneity that manifests across patients, between organs, and within tumors on medical imaging. Despite its widespread use to inform treatment decisions, programmed cell death ligand 1 (PD-L1) expression provides limited predictive value and is sensitive to sampling variability due to tumor heterogeneity. Furthermore, conventional response assessment by the Response Evaluation Criteria in Solid Tumors (RECIST) is based on changes in tumor size and does not capture functional or spatial variation within lesions, limiting its ability to guide treatment selection, adaptation, and radiotherapy planning. Current clinical decision-making is also primarily based on point predictions that do not quantify uncertainty, reducing their reliability in high stakes clinical settings. This dissertation addresses these limitations through a progression from biomarker discovery to an uncertainty aware, multimodal, and multiscale framework of informatics methods for treatment response prediction in advanced NSCLC. The work is structured around three distinct but complementary research objectives that collectively address these limitations. The dissertation first identifies fluorodeoxyglucose positron emission tomography (FDG-PET) imaging and circulating immune biomarkers, including T cell receptor and cytokine metrics, that are associated with treatment response and long-term outcomes in metastatic NSCLC, benchmarked against PD-L1. It then integrates these biomarkers into a multimodal framework that couples response prediction with conformal prediction to quantify patient-specific uncertainty, and demonstrates translational potential through a web-based clinical decision support prototype. To model spatial heterogeneity, it identifies a robust variogram that characterizes tumor spatial correlation across disease stages and treatment regimens. Building on this spatial foundation, it develops a multiscale framework for voxel-level response prediction across primary and metastatic lesions, and introduces a residual-variance conformal method that restores valid uncertainty under spatial dependence. Across these contributions, a central informatics theme is that predictions become clinically useful only when their reliability is explicitly quantified. By embedding uncertainty quantification at every scale, from patient- to voxel-level, this dissertation moves beyond prediction sets and intervals toward trustworthy, clinically actionable outputs. Although developed and evaluated in advanced NSCLC, the resulting informatics methods are generalizable and can be adapted to other solid tumors and broader biomedical informatics domains.