Towards the Efficient Inference by Incorporating Automated Computational Phenotypes under Covariate Shift

Area of Science:

• Biostatistics
• Computational Biology
• Health Informatics

Background:

• Manual phenotype data collection is slow and labor-intensive.
• Automated computational phenotypes (ACPs) provide a faster alternative but differ from gold-standard data.
• Integrating ACPs requires careful methods to avoid biased results in downstream analyses.

Purpose of the Study:

• To develop robust and efficient estimators for incorporating ACPs into semi-supervised learning.
• To leverage both labeled and unlabeled data under a covariate shift framework.
• To quantify the efficiency gains from using ACPs, particularly for unlabeled data.

Main Methods:

• Developed doubly robust and semiparametrically efficient estimators.
• Utilized a semi-supervised learning setting with labeled and unlabeled data.
• Analyzed efficiency gains by comparing scenarios with and without ACP inclusion.

Main Results:

• Identified that ACPs from unlabeled data significantly enhance efficiency gains.
• Demonstrated the practical advantages of the proposed method through synthetic experiments.
• Validated the approach on multiple real-world datasets.

Conclusions:

• The proposed estimators effectively leverage ACPs in semi-supervised learning.
• Incorporating ACPs, especially from unlabeled data, leads to substantial efficiency improvements.
• The method offers a valid and practical approach for utilizing ACPs in complex data analyses.