Idris Bachali Losada, Ph.D Adam Nassim2,3, Imran Nassim1,3
1IBM, 2Merck Institute for Pharmacometrics, Merck Serono S.A., Switzerland, a Subsidiary of Merck KGaA, Darmstadt, Germany,, 3Department of Mathematics, University of Surrey, Guildford, GU2 7XH ,Surrey, UK
1 Introduction/Objectives: Traditional mathematical models in oncology often utilize ordinary differential equations (ODEs) to describe tumor responses to treatment, linking changes in tumor volume to drug concentrations over time. These models typically incorporate tumor growth patterns—such as exponential, logistic, or Gompertzian growth—and correlate them with therapeutic agent concentrations in plasma or tumor tissues. However, this approach relies on a limited set of covariates, potentially restricting the model’s ability to fully capture the complexity of tumor dynamics. In contrast, deep learning (DL) techniques have been applied to modeling tumor growth dynamics, effectively managing heterogeneous and sparse multi-modal datasets, including omics data, immune profiling, imaging data, and electronic health records. With the advent of precision medicine, omics data provide a comprehensive patient-level overview of tumor metabolism, offering opportunities to incorporate individual-level details into tumor growth models and to better understand lesion-level behaviors. By leveraging multi-modal data, DL models can enhance our understanding of lesion-level tumor dynamics and improve individualized treatment strategies. This study discusses various DL frame- works for tumor size modeling that integrate multiple data modalities and incorporate explainable AI techniques, ultimately aiming to improve the interpretability of clinical outcomes. 2 Methods: We conducted a comprehensive review of existing DL frameworks[1] applied to tumor size modeling, focusing on those that integrate multi-modal data sources. Sources included omics data [2,3], imaging data[4], and electronic health records. We assessed the methodologies based on their ability to handle heterogeneous, multimodal data[5], model tumor growth dynamics accurately, and provide explainable insights into clinical outcomes [6]. 3 Results: Our review identified several DL frameworks capable of integrating multi-modal data for tumor size modeling. These models demonstrated improved predictive performance over traditional ODE-based models, particularly in capturing the complexity of tumor growth dynamics. Additionally, the incor- poration of omics data facilitated a better understanding of the models’ decision-making processes, thereby enhancing the interpretability of clinical outcomes at the patient-level. 4 Conclusions: Integrating multi-modal data into DL models offers a promising avenue for advancing tumor size modeling. The ability of these models to manage diverse data sources and provide explainable insights can lead to more personalized and effective treatment strategies. Future research should focus on further refining these models and validating their clinical utility in prospective studies.
1. Rodrigues, Jos´e Alberto. ”Using physics-informed neural networks (PINNs) for tumor cell growth modeling.” Mathematics 12.8 (2024): 1195. 2. Zhang, Chi, et al. ”A physics informed neural network approach to quantify antigen presentation activities at single cell level using omics data.” (2025). 3. Bansal, Hina, Hiya Luthra, and Shree R. Raghuram. ”A review on machine learning aided multi- omics data integration techniques for healthcare.” Data Analytics and Computational Intelligence: Novel Models, Algorithms and Applications (2023): 211-239. 4. Goswami, Mayank. ”Deep learning models for benign and malign ocular tumor growth estima- tion.” Computerized Medical Imaging and Graphics 93 (2021): 101986. 5. Balcerak, Michal, et al. ”Physics-regularized multi-modal image assimilation for brain tumor localization.” Advances in Neural Information Processing Systems 37 (2024): 41909-41933. 6.Hussein, Ahmad, Mukesh Prasad, and Ali Braytee. ”Explainable AI Methods for Multi-Omics Analysis: A Survey.” arXiv preprint arXiv:2410.11910 (2024). 2
Reference: PAGE 33 (2025) Abstr 11328 [www.page-meeting.org/?abstract=11328]
Poster: Methodology - New Modelling Approaches