3587 - Integrating Clinical and Radiomic Features for Enhanced Prognostic Modeling for Lung Cancer Survival

Purpose/Objective(s):

Traditional prognostic models for lung cancer patients remain primitive and lack robustness and accuracy. Current outcome prediction models rely solely on clinical factors. Radiomic features extracted from computed tomography (CT) imaging have been shown to improve outcomes prediction for lung cancer patients. This study integrates multimodal data from multiple institutions, combining clinical and radiomic features to develop a survival prognostic model for lung cancer patients.

Materials/Methods:

The study included two datasets: (1) an Institutional dataset of 207 lung cancer patients treated with radiotherapy (2019-2022) and (2) the TCIA NSCLC-RADIOMICS dataset with 422 non-small cell lung cancer (NSCLC) patients from The Cancer Imaging Archive (TCIA). Clinically used gross tumor volume were analyzed, and radiomics features were extracted from simulation CT scans. CT images were preprocessed, and the data was split 70%/30% for model training/evaluation. Feature reduction involved the Log-Rank Test for relevance assessment and Agglomerative Clustering for collinearity removal. Cox Proportional Hazards models were built for Overall survival (OS) prediction using three approaches: a traditional Cox (baseline) model, a Least Absolute Shrinkage and Selection Operator (LASSO) Cox model, and a Cox model using optimal features selected through LASSO regression. Model Performance was assessed via the concordance index (C-index).

Results:

The baseline Cox model achieved training C-index values of 0.83, 0.70, and 0.68, for the institutional dataset the TCIA NSCLC-RADIOMICS dataset and combined data set, respectively, with corresponding evaluation C-index values of 0.69, 0.55, and 0.58. Using the LASSO Cox approach, training C-index values were 0.64, 0.70, and 0.63 and the evaluation C-index values were 0.59, 0.55, and 0.59, respectively. Finally, with the optimal features, the training C-index values were 0.74, 0.64, and 0.63, while the evaluation values were 0.62, 0.57, and 0.58, respectively. After feature selection, the features found to be most critical in predicting OS were 13 radiomics texture features for the institutional dataset, 2D diameter and 4 radiomics texture features for the NSCLC dataset, age, gender, and 5 radiomics texture features for the combined dataset. There were no common features among the datasets. The integrated clinical and radiomics model included only age and gender along with many radiomics features, achieving a significantly higher C-index (clinical-only: 0.50–0.56 and integrated: 0.57–0.69).

Conclusion:

This study evaluates a novel radiomics-clinical algorithm for OS prediction in a multi-institutional 629 patient lung cancer dataset. It lays the foundation for automated, data-driven prognostic modeling. Future work includes deep learning-based feature extraction, synthetic data augmentation, and multi-modal data integration for early prediction and personalized treatment planning.