Artificial neural network-augmented dosiomic integration for predicting distant recurrence in NSCLC patients treated with SBRT.

Objective: Stereotactic body radiotherapy (SBRT) is a standard curative treatment for inoperable early-stage non-small cell lung cancer (NSCLC) patients. However, the high rate of distant recurrence following radiotherapy remains a significant clinical challenge. This study focuses on developing a machine learning model for distant recurrence prediction using diverse dosiomic and patient-specific clinical features. The proposed model aims to assist clinicians in informed decision-making, individualized treatment decisions to improve post-SBRT outcomes. Method: This study utilized a multi-institutional dataset comprising 575 NSCLC patients who underwent SBRT. A total of 21 features, comprising 14 dosimetric and 7 clinical variables, were incorporated for developing the predictive framework. The predictive model was developed based on an artificial neural network (ANN) architecture with several dense layers. Model training and internal validation were conducted using data obtained from one institution, while external validation was performed utilizing data from an independent institution. To enhance clinical interpretability, SHAP analysis was employed to evaluate the relative importance of each feature contributing to the model's output. Results: The initial predictive model, developed using individual clinical and dosimetric features, achieved area under the receiver operating characteristic curve (ROC-AUC) in the range of 0.64 to 0.65 while validated with an external dataset, respectively. To enhance predictive performance, dosimetric features were integrated with clinical variables, resulting in improved ROC-AUC values of 0.75 for internal validation with 10-fold cross validation technique and 0.71 for external validation with 1000 bootstrap iterations. Dosiomic features enhanced performance by 9-11%, highlighting their importance in distant recurrence prediction. Additionally, to enhance the interpretability of the model's predictions, SHAP-based analysis was conducted, revealing that the number of treatment fractions, dose per fraction, and minimum dose to GTV were among the five most influential dosiomic features. Conclusion: This study introduces an ANN-based model for predicting distant recurrence in NSCLC patients followed by SBRT. This study also demonstrates the impactful dosimetric and clinical features for the designed predictive model, highlighting its potential as an assistive tool for informed and individualized treatment planning in clinical practice.