Temporal Reproducibility of a Genetic Algorithm-Derived Health Risk Score: Standardized Out-of-Fold Validation...

Area of Science:

• Data-driven health risk stratification
• Preventive health screening methodologies
• Computational health analytics

Background:

• Genetic algorithm (GA)-based scoring offers a data-driven approach for health risk stratification.
• Performance estimates require strict separation of preprocessing, optimization, and evaluation within a validation framework.
• Temporal reproducibility and uncertainty quantification are crucial for translational reliability in health screening.

Purpose of the Study:

• Evaluate the temporal reproducibility of a GA-derived composite health risk score across three annual cohorts (2021-2023).
• Assess the score's robustness to missing HbA1c data using a sensitivity analysis.
• Validate the score under a standardized out-of-fold (OOF) evaluation pipeline.

Main Methods:

• Analyzed annual health examination datasets (2021-2023) with an identical preprocessing and modeling pipeline.
• Included 13 clinical indicators and 8 lifestyle variables; outcome based on A-D grading framework.
• Employed stratified K-fold cross-validation for GA optimization and logistic regression for OOF probability prediction; quantified performance using AUC and Brier score.

Main Results:

• Out-of-fold AUC values remained stable across cohorts (0.810-0.814) with overlapping confidence intervals.
• Brier scores indicated consistent overall predictive performance (0.172-0.176).
• Excluding HbA1c minimally impacted discrimination, confirming robustness.

Conclusions:

• The GA-derived composite risk score demonstrates stable temporal discrimination and consistent predictive performance.
• Highlights the importance of standardized, prespecified evaluation frameworks for risk stratification models.
• Transparent uncertainty quantification is essential for validating models using routine health screening data.