The article introduces Tumor Dynamic Neural-ODE (TDNODE), an explainable deep learning framework that models tumor size dynamics and predicts overall survival (OS) in oncology patients using longitudinal clinical data. TDNODE builds upon the Neural-ODE architecture, integrating physical principles by ensuring its learned parameters represent kinetic rates with units of inverse time. The model features an encoder-decoder structure that processes pre- and post-treatment tumor size measurements to generate individualized tumor growth trajectories and patient-specific dynamic metrics. These metrics are then used in an XGBoost-based survival model to accurately predict OS, outperforming traditional tumor growth inhibition-overall survival (TGI-OS) models. The study demonstrated that TDNODE could extrapolate tumor progression from early data without bias and explained its predictions using SHAP values and principal component analysis (PCA). Applied to the IMpower150 Phase 3 trial dataset in NSCLC patients, TDNODE achieved higher accuracy and interpretability than existing approaches, supporting its potential for use in model-informed drug development and personalized therapy.

Here are five major key points from the article:

• TDNODE is a novel, pharmacology-informed neural network that leverages Neural-Ordinary Differential Equations and an encoder-decoder architecture to model tumor dynamics directly from longitudinal clinical data.

• The model achieves interpretability by imposing a time-equivariant structure, ensuring that its learned parameters (encoder outputs) represent kinetic rate parameters with physical units of inverse time, which is clinically and biologically relevant.

• TDNODE significantly outperforms traditional TGI-OS models in predicting overall survival, achieving a c-index of 0.84 on the test set using only TDNODE-derived metrics, compared to c-indices as low as 0.68 for TGI models.

• The model demonstrates strong extrapolation capability, accurately predicting tumor sizes beyond the training window from early, truncated data without systematic bias.

• TDNODE's predictions are explainable through SHAP analysis and Principal Component Analysis (PCA), which reveal how the learned tumor dynamic metrics influence survival predictions, providing transparency for clinical decision-making.

Here are three limitations of the article:

• The model was exclusively tested on non-small cell lung cancer (NSCLC) patients from a single clinical trial (IMpower150), limiting its generalizability to other tumor types or broader patient populations without further validation.

• The current TDNODE model does not incorporate crucial information such as pharmacokinetic (PK) or dosing data, which could enhance its clinical utility, mechanistic insights, and improve predictive performance in real-world applications.

• The sophisticated architecture of TDNODE, including its reliance on specific neural network architectures and ODE solvers, may require significant computational resources and tuning, potentially posing challenges for practical implementation and broad applicability.