Optimization of Genomic Classifiers for Clinical Deployment: Evaluation of Bayesian Optimization to Select Predictive...

Area of Science:

• Computational biology
• Machine learning in healthcare
• Genomics and transcriptomics

Background:

• Acute infections pose significant risks, including sepsis, organ failure, and death.
• Current methods for detecting acute infections and assessing illness severity are suboptimal.
• Gene expression profiling offers a promising avenue for timely and precise diagnostics.

Purpose of the Study:

• To compare hyperparameter optimization (HO) approaches for developing diagnostic classifiers of acute infection and in-hospital mortality.
• To evaluate the effectiveness of Bayesian optimization against grid search and random sampling in this context.
• To assess classifier performance using a deployment-centered approach with multi-study patient cohorts and external validation.

Main Methods:

• Utilized gene expression data from 29 diagnostic markers.
• Developed and compared diagnostic classifiers using grid search, random sampling, and Bayesian optimization for hyperparameter tuning.
• Employed a deployment-centered analysis, including dataset partitioning and cross-validation across multiple studies.
• Validated selected classifiers both internally and externally.

Main Results:

• Classifiers selected via Bayesian optimization demonstrated superior performance for in-hospital mortality prediction compared to grid search or random sampling.
• Bayesian optimization did not consistently prove more efficient than other methods in all scenarios.
• Marginal performance improvements were observed only in specific circumstances with a common Bayesian optimization variant (automatic relevance determination).

Conclusions:

• While Bayesian optimization shows potential for improving mortality prediction classifiers, its advantages are not absolute.
• The efficiency and performance benefits of Bayesian optimization require careful consideration within specific healthcare applications.
• Further practical, deployment-centered benchmarking of HO methods is crucial for reliable implementation in clinical settings.