Comparison of Feature Selection Methods in Machine Learning Models of Cancer Information Seeking Among United States...

Feature selection is the process of identifying the most informative and relevant features from a larger set of candidate features in machine learning (ML) models. The Boruta algorithm and the least absolute shrinkage and selection operator (LASSO) are 2 widely used methods.

This study aimed to (1) compare several feature-selection strategies, including Boruta, LASSO, their intersection, principal component analysis (PCA), and a no-feature-selection baseline, and (2) evaluate ML models to predict cancer information-seeking behavior among US adults.

Data from 5505 individuals (2630 cancer information seekers and 2875 nonseekers) were selected from the 2022 Health Information National Trends Survey. The Boruta algorithm, LASSO, and PCA were used to perform feature selection of 73 variables. Five ML tools (the support vector machine algorithms, logistic regression [LR], random forest [RF], k-nearest neighbor, and extreme gradient boosting) were applied to develop ML models to predict cancer information-seeking. The area under the receiver operating characteristic curve (AUC) and the DeLong test were used to evaluate and compare the performance of the models. Stepwise LR analysis was performed to estimate the odds ratios and their 95% CIs for the associations of potential variables selected in ML analyses with the outcome.

Overall, 47.8% (2630/5505) of respondents reported seeking cancer information (949/2189, 43.4% of men; 1681/2189, 50.7% of women). RF achieved the highest AUC (0.781) and second-highest accuracy (0.714) using LASSO-selected variables, while the support vector machine with linear kernel and LR models using all 73 features yielded the highest accuracy (0.717). Notably, RF produced comparable AUCs when using Boruta-only features, LASSO-only features, or no feature selection yet (all 73 features); these AUCs were significantly higher than those derived from PCA components or from the 20 PCA-loading-based variables. Stepwise LR confirmed that 19 of the 27 shared variables selected by both Boruta and LASSO were independently associated with information seeking (P<.05). The top predictors included a personal history of cancer, greater worry about developing cancer, a family history of cancer, non-Hispanic White race, higher household income, awareness of genetic testing, viewing health-related videos on social media, interest in cancer screening, being offered access to an online medical record, and knowledge of human papillomavirus.

Boruta and LASSO demonstrated strong and consistent performance in feature selection for predicting cancer information seeking, whereas PCA provided a dimension-reduced yet less predictive alternative. Findings offer actionable insights for tailoring public health communication strategies and improving engagement in cancer information resources among US adults.