Flow Cytometry-Based Classification in Cancer Research: A View on Feature Selection.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Machine learning models are increasingly used for cancer detection and prediction.
• Effective feature selection is crucial for building accurate and stable models, especially with limited sample sizes.
• Previous research indicated high accuracy with sufficient data, but reliability with small datasets remains a challenge.

Purpose of the Study:

• To evaluate feature selection methods for cancer-related machine learning tasks using simplified pipelines.
• To identify reliable feature selection approaches for extremely small sample sizes.
• To compare traditional cross-validation with a Bayesian error estimator for model selection in logistic regression.

Main Methods:

• Investigated the accuracy and stability of various feature selection techniques.
• Utilized simplistic machine learning pipelines to focus on feature selection performance.
• Compared cross-validation and a Bayesian error estimator for model selection within the ℓ 1 regularization path of logistic regression classifiers.

Main Results:

• Demonstrated that up to 50% of features can be discarded without compromising prediction accuracy.
• Showcased the reliability of certain feature selection approaches even with extremely small sample sizes.
• Found comparable performance between cross-validation and the Bayesian error estimator in specific model selection scenarios.

Conclusions:

• Feature selection is robust in cancer machine learning, allowing significant data reduction without accuracy loss.
• The findings support the use of simplified models and effective feature selection for reliable cancer prediction with limited data.
• Both cross-validation and Bayesian error estimation are viable for model selection, offering flexibility in pipeline design.