Radiation pneumonitis (RP) is an early inflammatory reaction of lung parenchyma that arises following thoracic radiotherapy. Characteristically, RP can begin to manifest physiologically days to weeks before overt apparent radiological evidence manifests on CT scans. That gap of time gives rise to a diagnostic blind spot that blunts effective intervention. In this paper, we propose an entirely new multimodal AI framework that seeks to functionally identify RP in its earliest expression of manifestation-prior to anatomical observability-via an integration of CT-based radiomics together with surrogate infrared thermal indicators that are related to respiratory dynamics and thoracic surface temperature. Since there is no dataset of paired CT and IR imaging of RP patients to work with for this study, we construct a feature-level fusion dataset with CT instances being matched with thermal signs sourced from public IR databases of pneumonia, COVID-19, and respiratory function disorder. These thermal indicators simulated clinical symptoms of incipient RP stage in terms of thoracic asymmetry, impaired breathing cycles, and focal hyperthermia. This study should be interpreted as a feasibility analysis: we integrate CT radiomics with physiologically motivated surrogate infrared (IR) descriptors to support very-early functional detection of radiation pneumonitis (RP) within the subclinical, pre-radiological window. Robustness was assessed with stratified cross-validation, ablations (CT-only vs CT + IR), and two control checks—label-preserving IR permutation and leave-one-IR-dataset-out evaluation—together with probability calibration. The fused model outperformed the CT-only baseline (AUC 0.91 vs 0.79; sensitivity 0.88 vs 0.68; F1 0.86 vs 0.72). We avoid generalization claims and motivate prospective paired CT–IR acquisition.