JMIR Publications

ABSTRACT

Background:

Feature selection is essential in machine learning (ML) for identifying relevant variables. The Boruta algorithm and the least absolute shrinkage and selection operator (LASSO) are two widely used methods.

Objective:

This study aimed to (1) compare feature selection methods (Boruta, LASSO, their combination, principal component analysis (PCA), and non-feature selection), and (2) develop ML tools to predict cancer information seeking among U.S. adults.

Methods:

Data from 5505 individuals (2630 cancer information seekers and 2975 non-seekers) were selected from the 2022 Health Information National Trends Survey (HINTS 6). Four feature selection approaches and five ML tools (the support vector machines (SVMs) algorithms, logistic regression (LR), random forest (RF), k-nearest neighbor (KNN), and extreme gradient boosting (XGBoost)) were applied to develop ML models to predict cancer information seeking.

Results:

The cancer information seeking prevalence was 47.2% (42.8% for males and 49.7% for females). The Boruta and LASSO selected 45 and 55 variables, respectively, with 36 in common. The PCA identified 21 uncorrelated factors. RF performed best, with similar AUCs for Boruta, LASSO, and no feature selection (≈0.950) and accuracy (≈0.860). Using PCA-selected variables yielded a slightly lower AUC (0.931) but comparable accuracy (0.853). Stepwise regression confirmed 21 of 36 selected key predictors, including personal/family cancer history, health information access, education, income, social media use, smoking, alcohol beliefs, and healthcare visits.

Conclusions:

Feature selection effectively reduce dimensionality while retaining the predictive power. Boruta and LASSO performed comparably in terms of the selected variables. PCA-based selection uncorrelated s also proved useful. The key identified factors associated with cancer information seeking can guide future cancer intervention and prevention strategies.