RP-Fuse: Development of an Adaptive Dose-Gated Multimodal Model for Predicting Radiation Pneumonitis in Non-Small Cell Lung Cancer

Predicting radiation pneumonitis (RP) in non-small cell lung cancer (NSCLC) remains challenging due to complex dose–tissue interactions. Traditional dose-volume histograms (DVHs) are widely used but do not capture spatial dose distributions. In contrast, applying 3D deep learning to raw scans is prone to overfitting, particularly with small single-center datasets. We present RP-Fuse, a multimodal architecture featuring an adaptive dose gate that learns a continuous dose-response curve from the 3D dose map and uses it to weight spatial features from a computed tomography (CT) encoder. Independent expert networks process clinical records, dosimetrics, and a set of positron emission tomography (PET) uptake ratios. An attention layer combines these expert networks using only their out-of-fold (OOF) predictions. RP-Fuse was developed and internally validated using 5×5 nested cross-validation on a retrospective cohort of 87 NSCLC patients (30 developed grade ≥2 RP within six months). Our results show that integrating a dose-based prior facilitates spatial feature learning from raw imaging in small cohorts. With only CT and dose volumes, the adaptive gate increased the area under the receiver operating characteristic curve (AUC) from 0.550 to 0.704. Our final architecture recorded an AUC of 0.822 (95% confidence interval [CI] 0.74–0.91) with a positive Brier skill score of +0.218. Under an identical evaluation protocol, these were the highest values among fifteen reimplemented reference models; the margin over the strongest tabular model is +0.052 AUC (95% CI −0.03 to +0.13). Individually, the scalar and dose-gated configurations achieved identical pooled AUCs of 0.781. Averaging the two yielded a modest gain (+0.041 AUC, 95% CI −0.02 to +0.11) as their predictions exhibit complementary error patterns across patients. Given the relatively small cohort size and the fact that 92% of the participants are male, the RP-Fuse framework should be considered an exploratory proof of concept.