Uncertainty-Aware Multimodal Biomarker-Guided Treatment Response Prediction: Integrating FDG PET, T-cell Repertoire, and Inflammatory Cytokines with Conformal Uncertainty Quantification
Abstract
Purpose
Uncertainty-aware multimodal biomarker-guided treatment response prediction in metastatic NSCLC remains a critical unmet need to support robust therapy selection and adaptation over time. We developed a multimodal framework integrating FDG-PET, T-cell receptor (TCR), and cytokine biomarkers with conformal prediction for uncertainty-quantified treatment response prediction.
Methods
Thirty-five metastatic NSCLC patients (19 female [54%], age: 33-84 years [median: 71 years]) receiving first-line platinum-doublet–pembrolizumab on the PET-BRIGHT trial (NCT04151940) from 2019-2023 underwent FDG-PET/CT and blood collection at baseline and week 3; 27 were evaluable (18 post-treatment responders, 9 non-responders). Multimodal biomarkers included FDG-PET metrics (standardized uptake value/total lesion glycolysis), T-cell receptor (TCR) diversity metrics, and inflammatory cytokines. Nested leave-one-out cross-validation with direction-aligned consensus feature selection identified a single biomarker per modality. Class-balanced logistic regression was implemented to ensure well-calibrated probabilities for conformal prediction. Response prediction performance was evaluated via AUCROC and Brier score. Full conformal prediction (80% coverage) generated sets of {responders, non-responders}, with certainty defined by average set size (singleton=definitive). Unimodal and multimodal fusion (early vs late) models of baseline-only and baseline+mid-treatment biomarkers for adaptation/consolidation were compared.
Results
At baseline, unimodal TCR was dominant (AUCROC 0.80, Brier: 0.19), outperforming PET (AUCROC: 0.76, Brier: 0.21) and cytokines (AUCROC: 0.55, Brier: 0.27). Late fusion PET+TCR improved discrimination (AUCROC: 0.85, Brier: 0.17) and reduced uncertainty (set size 1.11 vs 1.22), with increased singleton predictions (78% to 89%). When using baseline+mid-treatment biomarkers, unimodal cytokines improved relative to the baseline-only model (AUCROC 0.78, Brier 0.20), while PET was similar (AUCROC: 0.76) and TCR was attenuated (AUCROC: 0.67). Late fusion PET+cytokines achieved highest discrimination (AUCROC: 0.86, Brier: 0.17) with identical set size and singleton predictions at both time-points (1.15, 85%).
Conclusion
Multimodal late fusion of longitudinal biomarkers with conformal prediction enhanced response prediction while providing patient-level uncertainty quantification for clinical decision support in managing metastatic NSCLC.