Ensemble Model-Driven Disease Recurrence Prediction for NSCLC Patients Treated with SBRT
Abstract
Purpose
Early prediction of distant recurrence in early-stage non-small cell lung cancer (NSCLC) patients may assist clinical decision making. Recent studies demonstrated hardly any benefit while adding systemic therapy to SBRT, and identification of patients at highest risk for developing distant relapses may be critical. Current study aims to develop a distant recurrence prediction model by combining radiomic, dosiomic, and clinical features for patients treated with SBRT.
Methods
Total 141 node-negative early-stage NSCLC patients treated with SBRT (48-60 Gy in 3-5 fractions) were included in this study. Radiomic features from planning CT images were gleaned from the gross tumor volume (GTV) zone and evaluated based on the histogram, gray level co-occurrence matrix, gray level run length matrix, and gray level size zone matrix. A novel ensemble model was developed based on: (1) Support Vector Machine, (2) Gradient Boosting, and (3) Random Forest algorithms with soft voting technique, while considering a combination of 47 radiomic features and 11 most influential dosiomic and clinical features as inputs. Output layer of the designed model utilized an optimal thresholding layer to predict distant recurrence after two-years of SBRT.
Results
Patients’ median age was 73years (range: 52-91); T-stage was T0-T2bN0; cohort had 46% female. Designed distant recurrence prediction model demonstrated superior performance with mean AUC of 0.75 (0.68-0.83), F1-score of 0.85 (0.80-0.92), sensitivity of 0.80 (0.67-0.96), and specificity of 0.70 (0.50-1). In contrast, model with the combination of radiomic features, clinical, and dosiomic variables exhibited an improvement of 8.7% in AUC compared to the model relies solely on radiomic features.
Conclusion
Leveraging radiomic, dosiomic, and clinical features has the potential for developing a distant recurrence prediction model for NSCLC patients treated with SBRT. A future study with a larger patient cohort from multiple institutes is desirable to verify the robustness of the developed ML model.