Toward uncertainty-aware clinical decision support for treatment response prediction in metastatic NSCLC: integrating FDG-PET, T-cell repertoire, and cytokines with conformal prediction.

Background: Variable response to chemoimmunotherapy in metastatic non-small cell lung cancer (mNSCLC), together with the limited discriminative accuracy of PD-L1 tumor proportion score, highlights the need for reliable, uncertainty-aware early response prediction using multimodal biomarkers. We developed and prototyped a multimodal clinical decision support (CDS) framework integrating longitudinal FDG-PET, T-cell receptor (TCR), and cytokine biomarkers with conformal prediction to deliver uncertainty-quantified, patient-level response predictions. Methods: Thirty-five patients with mNSCLC receiving first-line carboplatin-pemetrexed-pembrolizumab on the PET-BRIGHT trial (NCT04151940) underwent FDG-PET/CT and blood collection at baseline and week 3. Multimodal biomarkers included FDG-PET metrics (standardized uptake value/total lesion glycolysis), TCR diversity metrics, and inflammatory cytokines. Nested leave-one-out cross-validation with automated feature selection identified a single biomarker per modality. Class-balanced logistic regression was used for classification, with discrimination assessed by the area under the receiver operating characteristic curve (AUROC) and calibration by the Brier score. Conformal prediction (targeting 80% reliability) quantified patient-level uncertainty via prediction sets and singleton rate. Unimodal and multimodal early- and late-fusion models were evaluated using baseline-only and combined baseline plus mid-treatment biomarkers. Models were benchmarked against PD-L1 tumor proportion score, and statistical significance was assessed by permutation testing against an empirical null. Results: PD-L1 tumor proportion score, the current clinical standard, discriminated poorly (AUROC 0.58, 95% CI 0.39-0.78). At PreTx, unimodal TCR was the strongest single modality (AUROC 0.80), followed by PET (0.76) and cytokines (0.54). Late fusion of PET and TCR achieved the highest PreTx discrimination (0.85) and increased singleton predictions from 78% to 89%. After incorporating mid-treatment biomarkers, cytokines became the strongest single modality (0.78) while TCR was attenuated (0.66); late fusion of PET and cytokines achieved the highest discrimination overall (0.86). Thirteen of 22 model and timepoint combinations exceeded a permutation null at p < 0.05, and empirical coverage was 78% and 82% against an 80% target. Conclusions: A multimodal framework combining longitudinal biomarkers with conformal prediction substantially outperformed the current clinical standard while identifying patients for whom a confident prediction could not be made. A prototype CDS interface demonstrates feasibility for clinical translation. Trial registration: ClinicalTrials.gov NCT04151940, registered 26 September 2019.