Repeated Sieving for Prediction Model Building with High-Dimensional Data.

Area of Science:

• Biostatistics
• Machine Learning
• Personalized Medicine

Background:

• Accurate patient outcome prediction is crucial for personalized medicine.
• High-dimensional data (genomics, EHRs) require effective variable selection for prediction models.
• Existing methods like LASSO and Elastic Net can over-select features, impacting model accuracy and cost.

Purpose of the Study:

• To introduce and evaluate a novel machine learning method, repeated sieving, for variable selection in high-dimensional data.
• To compare the performance of repeated sieving against established methods like LASSO and Elastic Net.
• To assess the impact of variable selection on prediction accuracy and future data collection costs.

Main Methods:

• Proposed a repeated sieving machine learning method, extending regression with stepwise variable selection.
• Compared repeated sieving with LASSO (L1-norm penalty) and Elastic Net (L1/L2-norm penalties).
• Evaluated methods using extensive numerical studies and real-world data examples.

Main Results:

• Repeated sieving selected significantly fewer features compared to LASSO and Elastic Net.
• The proposed method demonstrated higher prediction accuracy than existing machine learning approaches.
• Numerical studies and real data confirmed the superior performance of repeated sieving.

Conclusions:

• The repeated sieving method offers superior performance in both variable selection and prediction accuracy for high-dimensional data.
• This approach effectively addresses the over-selection issue common in other machine learning methods.
• Repeated sieving reduces the cost associated with future data collection for prediction models.