3591 - When AI Explains Itself: Regional Recurrence Prediction in Early-Stage NSCLC Patients Using Interpretable DNN Framework

Purpose/Objective(s): Accurate identification of regional recurrence in patients with early-stage non-small-cell lung cancer (NSCLC) is critical for optimizing treatment management and clinical decision-making. This study aims to develop an interpretable deep neural network (DNN) model to support prediction of disease recurrence following Stereotactic Body Radiation Therapy (SBRT). Primary objective of this work is to determine the most influential clinical and dosiomic features contributing to regional recurrence risk prediction.

Materials/Methods: The study cohort included 240 patients with early-stage NSCLC treated with SBRT. Among them, 47% had T1b-stage disease. The median age was 76 years (range: 49-91 years), and 51% were female. Model inputs consisted of 10 clinical variables and 15 dosiomic features derived from dose-volume histograms (DVH). The proposed DNN architecture comprised nine hidden layers with Rectified Linear Unit (ReLU) activation functions and a sigmoid activation function in the output layer for binary classification. Model training was performed using the Adaptive Moment Estimation (ADAM) optimizer with weight decay regularization and a focal loss function to address potential class imbalance. To enhance interpretability, SHapley Additive exPlanations (SHAP) were employed to quantify feature importance and provide insight into the model’s predictive behavior.

Results: The DNN model using 10 clinical features achieved a mean AUC of 0.73 (±0.06), while the 15-feature dosiomic model achieved 0.75 (±0.09). The combined clinical–dosiomic model showed superior performance (AUC 0.81 ±0.07), with 0.89 sensitivity and 0.74 specificity. SHAP analysis identified the top predictors of regional recurrence as mean GTV dose (0.082), dose per fraction (0.036), PTV size (0.025), minimum GTV dose (0.025), and T-stage (0.020).

Conclusion: An interpretable DNN framework integrating clinical and dosiomic features improved prediction of regional recurrence after SBRT in early-stage NSCLC. Key predictive features were identified, supporting potential risk stratification. External multicenter validation is needed before clinical implementation.