Lasso regularization for left-censored Gaussian outcome and high-dimensional predictors

Area of Science:

• Biostatistics
• Computational Biology
• Genomics

Background:

• Biological assays often lack sensitivity, leading to left-censored data (values below detection limits).
• Human immunodeficiency virus (HIV) viral load quantification is a key example where left-censoring is common.
• Existing statistical methods struggle with left-censored data in high-dimensional settings.

Purpose of the Study:

• To develop and validate a statistical method for analyzing left-censored HIV viral load data with high-dimensional predictors.
• To improve the prediction of treatment response based on HIV viral load and genotypic mutations.

Main Methods:

• Reversed Buckley-James least squares algorithm combined with Lasso regularization for high-dimensional data.
• Incorporation of both non-parametric (Kaplan-Meier) and parametric (Gaussian) imputation methods.
• Cross-validation using a specialized loss function to handle censored and uncensored data.

Main Results:

• The proposed Lasso-regularized Buckley-James method significantly outperformed simple imputation strategies for left-censored, high-dimensional data.
• The Gaussian imputation approach with appropriate cross-validation demonstrated the lowest prediction error on simulated data.
• Analysis of real-world HIV data showed the method yields valid predictions correlating with HIV mutations.

Conclusions:

• The developed statistical approach effectively handles high-dimensional predictors and left-censored outcomes.
• This method shows significant promise for predicting HIV viral load and understanding treatment response in relation to HIV mutations.