Mitigating biases in feature selection and importance assessments in predictive models using LASSO regression

Abstract

Yuan et al. developed a predictive model for early response using sub-regional radiomic features from multi-sequence MRI alongside clinical factors. However, biases in feature selection and assessment may lead to misleading conclusions regarding feature importance. This paper elucidates the biases induced by machine learning models and advocates for a robust methodology utilizing statistical techniques, such as Chi-squared tests and p-values, to uncover true associations. By emphasizing the vital distinction between true and model-specific associations, we promote a comprehensive approach that integrates multiple modeling techniques. This strategy enhances the reliability of predictive models in medical imaging, ensuring that outcomes are based on objective relationships and ultimately improving patient care.