ISSN: 2640-2343
Authors: Gary G*, Azmul S and Jie Z
Purpose: To investigate the impact of clinical features on model performance in CT-based Non-Small Cell Lung Cancer (NSCLC) and the potential uncertainty regarding their application in machine learning. Methods: Clinical and radiomic features were retrospectively retrieved from EMR and CT images of 496 NSCLC patients. Five feature datasets were constructed: radiomic features-only (Rad), clinical features-only (Clin), shape features-only (Shape), radiomic and clinical features (RaClin), shape and clinical features (ShClin). Five feature selection methods and seven predictive models, along with different cohort sizes, number of input features and validation methods were included for the uncertainty analysis, with two-year survival as the study endpoint. AUC values were calculated for comparisons and Kruskal-Wallis testing was performed to determine significant differences. Results: A total of 19740 distinct combinations of feature sets, feature selection methods, predictive models, cohort sizes and validation techniques are examined. Of those, 25 combinations produce an AUC > 0.7. The clinical-only feature dataset generally outperforms both the radiomic-only feature dataset and the hybrid (clinical and radiomic) feature dataset (P<0.01), which is primarily determined by the endpoint. The combination of different feature selection methods and predictive models, along with the variations in cohort size, number of input features and validation methods generate inconsistent results. Conclusion: Clinical features are a source of data that can improve machine learning model performance. However, its impact strongly depends on various factors that may lead to inconsistent results. A clear approach to incorporate clinical features to generate reliable results requires further investigation.
Keywords: Clinical Features; Radiomics; Lung Cancer; Machine Learning
