Novel Feature Selection for Artificial Intelligence Using Item Response Theory for Mortality Prediction.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Feature selection is crucial for supervised machine learning classification.
• Traditional methods may introduce inefficiencies and circular errors by using end-classification results.
• A need exists for robust feature reduction techniques independent of classification outcomes.

Purpose of the Study:

• To evaluate Item Response Theory (IRT) as a feature selection method in machine learning.
• To assess the utility of IRT in identifying relevant features for classification tasks.
• To compare IRT-based feature selection with traditional machine learning approaches.

Main Methods:

• A two-parameter dichotomous Item Response Theory (IRT) model was employed.
• 18 features from an intensive care unit dataset (2520 cases) were analyzed.
• Feature utility was examined independently of end-classification results.

Main Results:

• IRT successfully identified features, demonstrating high utility.
• Features selected via IRT yielded classification results comparable to traditional methods.
• The IRT protocol proved effective in feature reduction for the analyzed dataset.

Conclusions:

• Item Response Theory provides a viable and effective alternative for feature selection in machine learning.
• IRT circumvents circular errors inherent in some traditional feature reduction techniques.
• The IRT approach offers a robust method for enhancing classification model performance and efficiency.