A Digital Twin Framework for Per-Fraction Prediction of Normal Tissue Toxicity in Radiotherapy of Non-Small Cell Lung Cancer

Abstract

Background/Objectives: Stereotactic body radiation therapy (SBRT), often delivered in 3–5 fractions, is a highly effective treatment strategy for early-stage non-small cell lung cancer (NSCLC) in medically inoperable patients. Biology-guided radiotherapy (BGRT), a PET-guided SBRT, enables acquisition of patient-specific metabolic, anatomical, and dose information at each fraction. While prior models have focused on static, population-level prediction of radiation-induced toxicity, there remains a lack of approaches to characterize how normal tissue response evolves longitudinally within individual patients over the course of radiation treatment. We therefore hypothesize that AI-based models can quantify temporal toxicity patterns in such longitudinal data, allowing clinicians to balance tumor control against normal tissue complications to guide individualized treatment planning. Methods: In this proof-of-concept study, we developed COMPASS (COMprehensive Personalized Assessment System), a digital twin framework that generates per-fraction risk estimates that could inform future adaptive decision-making in NSCLC. With Yale Institutional Review Board approval, eight patients with NSCLC who underwent BGRT were retrospectively included in this study. A total of 99 organ–fraction observations, spanning 24 patient–organ longitudinal trajectories across three organs-at-risk (spinal cord, heart, and esophagus), were used to train a Gated Recurrent Unit (GRU) autoencoder to learn latent temporal representations, which were subsequently characterized using logistic regression to assess for CTCAE ≥ 1 toxicity. Results: In a leave-one-out patient evaluation, COMPASS achieved an area under the curve (AUC) of 0.90, with 80% sensitivity and 78.6% specificity in predicting normal tissue toxicity at the patient–organ trajectory level on a fractional basis. Conclusions: These findings support the feasibility of AI-driven digital twins for per-fraction toxicity monitoring, offering a step toward personalized radiotherapy for NSCLC patients.