OncoTwin: A Living, Probabilistic, Multimodal Digital Twin of a Cancer Patient with Calibrated Uncertainty

Digital twins are computational models kept in step with a specific real system as new data arrives are a promising route to personalized oncology, yet a clinically meaningful twin must forecast uncertain tumor dynamics honestly and update as new scans appear. Outcome predictors that emit a single number, with neither a trustworthy sense of confidence nor a mechanism to revise beliefs over time, are poorly suited to this setting. We ask whether an interpretable, continually-updated twin can produce forecasts whose uncertainty is well-calibrated, and we evaluate it candidly against real patient data. We present OncoTwin, a living, probabilistic, multimodal patient twin that fuses medical imaging with clinical data to forecast tumor progression and to compare treatment strategies as counterfactuals. A machine-learning estimator maps patient features to a distribution over biologically meaningful parameters of a mechanistic Gompertz growth model that includes treatment effects, acquired chemotherapy resistance, and linear-quadratic radiotherapy. Beliefs are recalibrated from each new scan through a particle-filter data-assimilation loop; an imaging pipeline segments the tumour and extracts radiomic features fused with clinical variables. Forecasts are reported as predictive intervals rather than point estimates. Our central finding concerns calibration. Back-testing on a synthetic cohort showed forecasts were badly over-confident: nominal 90% predictive intervals covered the ground truth only 31–37% of the time. Introducing process noise that grows with the forecast horizon raised coverage to 95% with no loss of accuracy, yielding intervals that are honest rather than merely narrow. Evaluated against real outcomes from the public TCIA NSCLC-Radiomics cohort (399 patients), the twin's biology-derived risk attains a survival concordance index of 0.505 at chance a negative result we report transparently and attribute to the model not being trained on survival and to per-patient imaging features not yet being used at cohort scale. Prior radiomics work shows that imaging features predict survival in this cohort, and classical growth models capture tumor kinetics; our contribution is to combine them in a single continually-updated twin and, distinctively, to foreground calibrated uncertainty and an honest negative result rather than a favorable headline number. Ongoing work extracts expert-delineated radiomics across the full cohort, trains a survival model on real imaging, and pursues fitting to longitudinal data, toward the prospective, clinician-in-the-loop validation a clinical tool would ultimately require.