How to use learning curves to evaluate the sample size for malaria prediction models developed using machine learning...

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in health

Background:

• Machine learning models are crucial for predicting malaria risk, severity, and drug resistance.
• Accurate prediction models require large training datasets.
• Learning curves assess the necessary sample size for effective model training.

Purpose of the Study:

• To demonstrate the generation and interpretation of learning curves for malaria prediction models.
• To guide the sample size determination for future malaria prediction studies using machine learning.

Main Methods:

• Simulated gene expression data from Plasmodium falciparum isolates.
• Evaluated two machine learning algorithms: sPLSDA+SVMs and random forests.
• Calculated balanced error rate to measure prediction accuracy across different training dataset sizes.

Main Results:

• Balanced error rates decreased as training data size increased, reaching 14% (sPLSDA+SVMs) and 22% (random forests) with 835 samples.
• Learning curves indicated diminishing returns in accuracy beyond 835 samples.
• Initial performance was poor (50% error) with only 20 samples.

Conclusions:

• Learning curves are a valuable tool for determining minimum sample sizes in malaria prediction modeling.
• These curves are essential for optimizing machine learning applications in malaria research.
• Each specific prediction task requires a unique learning curve analysis.