Sample size requirements for learning to classify with high-dimensional biomarker panels

Area of Science:

• Biomedical Research
• Statistical Modeling
• Machine Learning

Background:

• Determining adequate sample size for biomarker-driven classifiers is a common challenge in biomedical research.
• High-dimensional biomarker panels require robust methods for sample size calculation.

Purpose of the Study:

• To present a simple Gaussian approximation method for calculating the required sample size to train a classifier.
• To assess the impact of biomarker panel characteristics and optimal predictive performance on sample size requirements.

Main Methods:

• Utilized a Gaussian approximation to model classifier performance.
• Derived sample size calculations based on the C-statistic (Copt) and the sparsity of biomarker effect sizes.
• Assumed identical correlation structures for biomarker effect sizes and biomarkers.

Main Results:

• The required sample size is independent of biomarker correlations, depending only on Copt and effect size sparsity.
• For 80% predictive information extraction, sample size ranges from 0.1 to 9 cases per variable.
• Lower sample sizes are needed for high Copt (0.9) and sparse effects (1% nonzero), while higher sizes are needed for lower Copt (0.75) and diffuse effects.

Conclusions:

• The proposed method simplifies sample size calculation for biomarker panels.
• Sparsity of effect sizes and optimal predictive performance are key determinants of sample size.
• This approach aids in efficient study design for biomarker-based classification.