The Univariate Flagging Algorithm (UFA): An interpretable approach for predictive modeling.

Area of Science:

• Data Science
• Machine Learning
• Bioinformatics

Background:

• Many data classification methods yield similar accuracy, but lack interpretability for non-experts.
• Domain experts understand input variables but not their complex interactions.
• Identifying thresholds in variable ranges is crucial for practical applications.

Purpose of the Study:

• To introduce the Univariate Flagging Algorithm (UFA) for automatic threshold detection.
• To develop interpretable classification approaches using UFA.
• To evaluate UFA's performance and identify optimal application conditions.

Main Methods:

• The Univariate Flagging Algorithm (UFA) searches for data separations optimizing outcome differences and support.
• Performance evaluated on six datasets, comparing UFA-based classifiers to traditional methods.
• Analysis of UFA's effectiveness in scenarios with missing/noisy data, high input-to-observation ratios, and low target incidence.

Main Results:

• UFA identified thresholds aligning with published results and known physiological boundaries.
• UFA-based classifiers achieved comparable or superior performance on unseen data versus traditional classifiers, considering confidence intervals.
• UFA demonstrated robustness in datasets with significant missing or noisy data.

Conclusions:

• UFA provides an interpretable and robust classification method suitable for non-expert users.
• Its ability to detect thresholds and handle data imperfections makes it valuable for clinical applications.
• UFA offers a simple yet effective approach for identifying low-incidence adverse outcomes.