Surrogate Assisted Semi-supervised Inference for High Dimensional Risk Prediction.

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Risk modeling using electronic health records (EHR) is hindered by unobserved disease outcomes and high-dimensional predictors.
• Existing methods struggle with incomplete outcome data and complex predictor sets.

Purpose of the Study:

• To develop a robust semi-supervised learning approach for risk prediction using EHR data.
• To address challenges of missing outcomes and high-dimensional predictors in EHR-based risk modeling.
• To leverage both labeled and unlabeled data for improved statistical inference.

Main Methods:

• A surrogate-assisted semi-supervised learning framework is proposed.
• Unobserved outcomes are imputed using a sparse imputation model with outcome surrogates and high-dimensional predictors.
• A one-step bias correction is applied for valid interval estimation in risk prediction.

Main Results:

• The proposed method demonstrates superiority over existing supervised methods in extensive simulation studies.
• The inference procedure remains valid even with misspecified imputation and risk prediction models.
• The approach enables high-dimensional statistical inference for dense risk prediction models.

Conclusions:

• The novel approach effectively utilizes unlabeled EHR data for enhanced disease risk prediction.
• This method offers a powerful tool for genetic risk prediction, as demonstrated in a type-2 diabetes mellitus cohort.
• The technique provides a valid and robust framework for statistical inference in challenging EHR data settings.