Uncertainty-Aware Multiscale Patient- and Voxel-Level Treatment Response Forecasting for Personalized Radiotherapy Using Conformal Prediction In Advanced Non-Small Cell Lung Cancer (NSCLC)
Abstract
Purpose
Multiscale treatment response prediction in advanced NSCLC enables spatially informed dose painting, yet prediction point estimates alone do not convey the uncertainty required for adaptive therapy decision support. We developed a multiscale conformal prediction framework that provides predictions and intervals for mid treatment FDG-PET response at both voxel and lesion levels.
Methods
Data were collected from two prospective clinical trials: FLARE-RT (NCT02773238, n=25 locally-advanced NSCLC) and BRIGHT (NCT04151940, n=19 metastatic NSCLC). Patients underwent FDG-PET/CT at baseline and week 3 of therapy (chemoradiation in FLARE-RT and chemoimmunotherapy in BRIGHT). A hybrid generalized least squares model was used to predict tumor voxel standardized uptake value (SUV) at mid-treatment from baseline patient-level and voxel-level covariates, building on a previously developed model (Voxel-Forecast) with re-optimized Stable variogram modeling for spatial correlation. We compared the forecast performance (cross-validated RMSE) using patient-level features only versus multiscale features. To quantify uncertainty, we applied a nested leave-one-out jackknife conformal prediction framework (targeting 80% coverage, alpha = 0.2) with spatially adaptive weighting and multiscale calibration sets.
Results
Incorporating voxel-level features with patient-level features achieved significantly lower RMSE (0.22 vs. 0.34 [FLARE-RT] and 0.44 vs. 0.64 [BRIGHT], p<0.05) than patient-level only. Voxel-level coverage of the conformal prediction intervals was 79.0% ± 22.6% in FLARE-RT and 80.5% ± 15.0% in BRIGHT, close to the target level of 80%, with mean interval widths of 0.7± 0.3 and 1.0 ± 0.2, respectively. Lesion/patient-level coverage was lower: 71.4% ± 25.1% in FLARE-RT and 68.8% ± 17.5% in BRIGHT, with narrower mean widths of 0.6 ± 0.2 and 0.9 ± 0.3.
Conclusion
Multiscale patient- and voxel-level information improved response prediction and uncertainty-aware voxel-level forecasting for personalized adaptive therapy. The proposed multiscale conformal prediction approach provides rigorous, distribution-free safety intervals for mid treatment FDG-PET response, supporting risk-aware decision support beyond point-based predictions.