Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session D

Poster A6 - Model and parameter identifiability in predictive image-based models of high-grade glioma

Friday, October 25, 2024
3:30 PM - 4:30 PM EST
Location: Exhibit Hall E, F & G

Introduction:  High-grade gliomas are highly aggressive and invasive malignant brain tumors, and despite combination surgery, radiation, and chemotherapy, patients still have a dismal prognosis. Advances in medical imaging, especially with multi-parametric magnetic resonance imaging (mpMRI), now enable a rich series of noninvasive measurements of tumor properties including cellularity, tissue structure, blood flow, and blood volume that provide insight into tumor growth and treatment response. Furthermore, this imaging data can be used to initialize and constrain biology-based mathematical models to forecast treatment response on an individual basis. We have previously developed a family of biology-based mathematical models which describe tumor growth and response with varying degrees of complexity (i.e., number of model parameters) and assumptions on the underlying biology. Importantly, the identifiability of these models (and their parameters) in the presence of noise is under investigated. To this end, we have performed a two-part analysis to determine the ability of our model calibration framework to (1) accurately recover model parameters and (2) correctly identify the most suitable model for a given dataset.

Materials and

Methods:  We introduce 11-member model family build upon a mechanically-coupled two-species reaction diffusion model describing the enhancing and non-enhancing components of high-grade gliomas, discussed in previously published work [1]. We then introduce 10 different models of response to radiotherapy (RT) and chemotherapy (CT) which varies the uniformity of treatment efficacy (e.g., uniform efficacy, related to cell density, and related to vasculature) and the parameterization of radiotherapy and chemotherapy terms (e.g., separate vs. combined terms). Model 11 does not explicitly model RT and CT. The effects of RT and CT are assumed to result in immediate cell death at the time of delivery. For each member in the model family, we generated 21 in silico tumors (231 total tumors) whose model parameters (e.g., proliferation rates, diffusion coefficients, treatment efficacy) were sampled from our previous work. To assess our ability to recover model parameters and the model used to generate our in silico tumors, we calibrated every model to each combination of tumor and noise level (5%, 10%, and 15%), and calculated the percent error between the estimated model parameter and the model parameter used to “grow” the in silico tumor. To assess our ability to correctly identify the model that “grew” each in silico tumor, we calibrated every model to each combination of tumor, noise level, and model and then calculated the Akaike information criterion to select the model that best balances model complexity and fit.

Results, Conclusions, and Discussions: 

Results: For the first analysis, we determined that the average percent error across all parameters for all models was 2.34% at 5% noise, 3.61% at 10% noise, and 5.09% at 15% noise. The parameter with the highest noise was the alpha/beta ratio (a measure intrinsic radiosensitivity). Figure 1 shows the results of the second analysis indicating that most models were identifiable at all noise levels. Generally, as more noise was added, the models became less identifiable.

Discussion: We demonstrated that our model calibration framework was able to recover model parameters with less than 5.09% error at all noise levels. Additionally, we showed that our model framework was able to correctly identify 11 models used to generate the in silico tumors for 82% of the cases. This results indicate that our model calibration and selection framework is able to distinguish the models from each other.

Conclusion: Patient-specific mathematical models of tumor growth and response have the potential to serve as powerful tools for clinicians, enabling them to predict patient outcomes and optimize therapeutic regimens. However, without identifiable models or the ability to accurately recover model parameters, the interpretations of model predictions or optimized therapeutic regimens may be unreliable. The results of this study establish the reliability of the model calibration and selection framework to identify the most appropriate model under noisy conditions expected in the experimentally setting.

Acknowledgements (Optional):  This work is supported by CPRIT RP220225.

References: [1] Hormuth, D.A., Al Feghali, K.A., Elliott, A.M. et al. Image-based personalization of computational models for predicting response of high-grade glioma to chemoradiation. Sci Rep 11, 8520 (2021). https://doi.org/10.1038/s41598-021-87887-4