2121 - Overall Survival Prediction Utilizing Support Vector Machine with Cluster Partitioning Methodology for NSCLC Patients Treated with SBRT

K. Halder¹, T. Biswas², F. Maria-Joseph¹, Y. Zheng³, and T. K. Podder⁴; ¹Indian Institute of Technology, Roorkee, India, ²Metro Health, Case Western Reserve University, Cleveland, OH, ³University Hospitals, Case Western Reserve University, Cleveland, OH, ⁴SUNY Upstate Medical University, Syracuse, NY

Purpose/Objective(s): This study focuses on designing an overall survival (OS) prediction model for non-small cell lung cancer (NSCLC) patients following one year of stereotactic body radiotherapy (SBRT) treatment, employing the Support Vector Machine (SVM) with a cluster partitioning method. Materials/

Methods: Data was collected from a cohort of 309 NSCLC patients for predicting one-year overall survival (OS) following SBRT treatment. AI-based OS-prediction model was developed incorporating a total of twenty-one common clinical and dosimetric features: Age, Sex, Total Dose in BED, Location of tumor, PTV size, GTV size, GTV_max dose, GTV_min dose, GTV_mean dose, PTV_max dose, PTV_min dose, PTV_mean dose, Conformity index, Homogeneity index, Lung_mean Dose, aCCI, PET Max SUV, T-stage, Smoking status, History of COPD, and Performance status. The entire dataset was split into a training dataset with 243 patients and a testing dataset having 66 patients. In this study, an innovative predictor model based on SVM with cluster partition methodology was developed to enhance the prediction performance while addressing the imbalance in dataset. In this proposed process, six different sub-datasets were randomly built using a cluster partitioned method, thereby ensuring the inclusion of approximately equal instances of both classes. Subsequently, six distinct SVM models were designed to train with the same samples of minority classes and different sets of majority classes. Furthermore, in the testing phase, the abovementioned four trained models with a final decision-making module utilizing higher probability logic are combined to make predictions for survival.
Results: The proposed OS-predictor model demonstrated a significantly better performance in comparison to traditional classifiers, producing an ROC-AUC of 0.75, sensitivity of 0.85, and specificity of 0.64. In order to demonstrate the effectiveness of the proposed survival prediction model, comprehensive comparative analyses were conducted based on conventional classifier models, as illustrated in the table below. From the viewpoint of comparative analyses, sensitivity and ROC-AUC performance parameters were observed significantly low; whereas, the proposed model demonstrated a satisfactory result with an increment of 41% in ROC-AUC score.
Conclusion: The comparative analyses illustrated that conventional classification simulation exhibited favorable performance solely for the majority classes, whereas our proposed prediction model provided with significantly improved performance. The proposed AI-based OS-prediction model may potentially assist in enhancing the clinical decision making and management of the NSCLC patients. Table 1: Comparison of the proposed overall survival prediction model with existing classifier models.

Performance Parameters	SVM (Traditional)	Decision Tree	Logistic Regression	Proposed model
Sensitivity	0.15	0.15	0.07	0.85
Specificity	0.90	0.79	0.98	0.64
ROC-AUC	0.53	0.47	0.49	0.75