Predictive Digital Twins with Quantified Uncertainty for Patient-Specific Decision Making in Oncology

TL;DR

This work presents an end-to-end Bayesian data-to-decisions framework for predictive digital twins in oncology, integrating patient-specific MRI-derived geometries with a high-dimensional reaction-diffusion model of glioma growth and treatment effects. It employs Gaussian-process-inspired priors and a low-rank Laplace approximation to perform scalable Bayesian calibration, enabling tractable uncertainty quantification and forward propagation to clinically relevant quantities. The methodology is demonstrated first on a virtual UPENN-GBM patient and then on IvyGAP clinical data, showing improved predictive accuracy and reduced uncertainty compared to priors, while highlighting model inadequacy and opportunities for refinement. The approach, implemented with HPC-accelerated forward/adjoint solves and publicly available software, lays a foundation for digital twins that can inform personalized treatment planning and optimal data acquisition strategies in oncology.

Abstract

Quantifying the uncertainty in predictive models is critical for establishing trust and enabling risk-informed decision making for personalized medicine. In contrast to one-size-fits-all approaches that seek to mitigate risk at the population level, digital twins enable personalized modeling thereby potentially improving individual patient outcomes. Realizing digital twins in biomedicine requires scalable and efficient methods to integrate patient data with mechanistic models of disease progression. This study develops an end-to-end data-to-decisions methodology that combines longitudinal non-invasive imaging data with mechanistic models to estimate and predict spatiotemporal tumor progression accounting for patient-specific anatomy. Through the solution of a statistical inverse problem, imaging data inform the spatially varying parameters of a reaction-diffusion model of tumor progression. An efficient parallel implementation of the forward model coupled with a scalable approximation of the Bayesian posterior distribution enables rigorous, but tractable, quantification of uncertainty due to the sparse, noisy measurements. The methodology is verified on a virtual patient with synthetic data to control for model inadequacy, noise level, and the frequency of data collection. The application to decision-making is illustrated by evaluating the importance of imaging frequency and formulating an optimal experimental design question. The clinical relevance is demonstrated through a model validation study on a cohort of patients with publicly available longitudinal imaging data.

Figures (16)

• Figure 1: Illustration of our digital twin workflow for a cancer patient. Observational data are integrated with mechanistic models of tumor growth to update the computational representation of the dynamics with quantified uncertainty. The calibrated model is used to make probabilistic forecasts of tumor progression accounting for patient response to therapy. In turn, these forecasts guide clinical decision making. The model may be used to assess what-if scenarios for alternative interventions or to optimize therapy directly by (for example) adjusting dose level or schedule.
• Figure 2: The computational pipeline: anatomic segmentation and mesh generation, cellularity estimation, and longitudinal registration. Together $T_1$- and $T_2$-weighted scans are used to generate a computational domain tailored to the patient's anatomy. Tumor cellular density is estimated by combining ADC imaging with tumor segmentations. When radiologist defined segmentations are not available, automated tools utilizing $T_1$ pre- and post-contrast imaging along with $T_2$-weighted and FLAIR modalities are used to develop appropriate regions of interest.
• Figure 3: Snapshots of synthetic tumor progression for UPENN-GBM subject 101. Note the heterogeneous initial state and the retreat while under therapy (weeks 2-8). When therapy ends, tumor recurrence is swift and the tumor extent is larger at the prediction time (week 16) than during the imaging window (up to week 12).
• Figure 4: First column: axial slices showing white and gray matter segmentation with tumor state at final observation (week 12). Second column: map reconstruction of the log-diffusion field. Third column: absolute error in the reconstructed log-diffusion field. Fourth column: map reconstruction of the log-reaction field. Fifth column: absolute error in the reconstructed log-reaction field. Each log parameter field has discretized dimension $137,261$. Note the inferred heterogeneity in the reconstructed log-diffusion field as well as the spatial structure of the error where the parameters are well informed where the tumor is active.
• Figure 5: Spectral decay of the prior-preconditioned Hessian for various imaging frequencies. The larger eigenvalues associated with more frequent imaging indicate more information gain in the corresponding eigenvectors.