Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
Presenter(s)
Anvi Sud, MS - Yale University, New Haven, CT
A. Sud1, J. Huang2, G. R. Hart3, J. J. Kim4, L. Tressel4, and J. Deng5; 1Department of Biomedical Data Science, Yale University, New Haven, CT, 2Department of Statistics and Data Science, Yale University, New Haven, CT, 3Division of Natural Sciences, University of Guam, Mangilao, GU, Guam, 4Department of Therapeutic Radiology, Yale School of Medicine, New Haven, CT, 5Department of Therapeutic Radiology, Yale University School of Medicine, New Haven, CT
Purpose/Objective(s):
Biology-guided radiotherapy (BGRT) for non-small cell lung cancer (NSCLC) produces high-frequency per-fraction imaging and delivered dose data that captures evolving normal tissue biology throughout treatment. Current toxicity prediction relies on static, population-derived models and constraints based on pre-treatment planning CT, which do not account for temporal dose redistribution, anatomical change, or evolving biological response. This work aims to model normal tissue toxicity as a patient-specific, time-dependent process using longitudinal multimodal data to enable clinical decision support for more individualized radiotherapy.Materials/Methods:
We developed COMPASS (COMprehensive Personalized ASsessment System), a digital-twin framework integrating per-fraction PET/CT imaging, radiomics, dosiomics, and cumulative biologically equivalent dose (BED) kinetics. Organ-specific time-series features were derived to preserve spatial dose heterogeneity and biological response for each patient. A gated recurrent unit (GRU) autoencoder was used to learn compact latent representations of evolving dose–response trajectories for critical organs, which were subsequently classified using logistic regression to predict eventual CTCAE grade =1 toxicity. Eight NSCLC patients treated with BGRT contributed 99 organ-fraction observations across 24 organ trajectories (heart, esophagus, spinal cord). Model performance was evaluated using leave-one-patient-out cross-validation.Results:
Despite the limited cohort size, dense longitudinal phenotyping enabled robust characterization of individualized normal tissue response, including treatment complications. Overall, COMPASS achieved an AUC of 0.90, with 80% sensitivity and 78% specificity. Importantly, COMPASS predicted increasing toxicity risk several fractions prior to clinical symptom onset, allowing timely intervention through early warning. Incorporation of BED kinetics and spatial dose-texture features sensitively captured transient metabolic and dosimetric perturbations not reflected by regular metrics.Conclusion:
This study demonstrates the feasibility of modeling normal tissue toxicity as a dynamic, patient-specific process using per-fraction multimodal data. COMPASS provides a physics-informed framework for adaptive radiotherapy in which toxicity risk is regularly updated based on fractional delivered dose and biological response, advancing toward individualized BGRT.