306 - Early Prediction of Local Recurrence after SBRT in Non-Small Cell Lung Cancer Using CT-Based Imaging Biomarkers and Clinical Factors

Purpose/Objective(s): Local recurrence (LR) following SBRT for early-stage non-small cell lung cancer (NSCLC) is uncommon. However, it is clinically consequential and can be difficult to identify patients at risk of LR based only on clinical factors. Radiomic analyses show promise for recurrence prediction; however, most existing prediction models rely on pre-treatment imaging or radiographic changes observed long after therapy is completed. We aimed to develop an integrated clinical-radiomic model using early post-SBRT chest CT imaging to predict LR.

Materials/Methods: Analysis included 134 patients with T1-T3N0 (<6 cm) NSCLC who received CT chest imaging obtained within 6 months post-SBRT. Radiomics features were extracted from the primary region of interest and 5mm ring expansion region on CT scans. Feature set harmonization was performed to mitigate the effect of heterogeneity in the following image parameters on the radiomics features: contrast enhancement, manufacturer, kernel resolution, pixel spacing and slice thickness. Feature set dimensionality reduction was performed using variance thresholding to select the top 7 predictors from radiomics features and clinical variables (anatomic tumor location, histology, maximum dimension of tumor and ECOG). Predictors significant on univariate Cox regression (p<0.05) were entered into multivariate Cox models. Due to imbalance in recurrence events (recurrence event- 15, no recurrence- 119), we performed 70:30 train-test split and applied imbalance correction to the training data. Binary classification analysis for identification of two-year LR event was performed using a quadratic support vector machine. The following three models were investigated using different information for prediction: Model 1 (primary region of interest radiomics + ring radiomics + clinical), Model 2 (primary radiomics + clinical) and Model 3 (clinical).

Results: Results of Cox regression c-scores for time-to-event analysis (two-year recurrence) were: Model 1: train- 0.76 [0.72, 0.78]; test- 0.68 [0.63, 0.69]; Model 2: train- 0.74 [0.69, 0.76]; test- 0.66 [0.62, 0.68]; Model 3: train- 0.61 [0.56, 0.62]; test- 0.53 [0.51, 0.56]. Results of the two-year recurrence event classification analysis were [sensitivity, specificity, AUC]: Model 1: train- [84.2%, 87.2%, 0.89]; test- [70.3%, 75.2%, 0.78]; Model 2: train- [82.7%, 86.4%, 0.88]; test- [69.1%, 73.6%, 0.76]; Model 3: train- 73.5%, 77.2%, 0.83]; test- [63.5%, 61.4%, 0.68].

Conclusion: Machine learning models integrating non-invasive CT imaging biomarkers and clinical variables achieved effective 2-yr LR prediction within 6 months after SBRT in NSCLC patients.