Optimization of Genomic Classifiers for Clinical Deployment: Evaluation of Bayesian Optimization to Select Predictive Models of Acute Infection and In-Hospital Mortality.

• DOI: 10.7490/f1000research.1118405.1
• Publication Date: 2020-12-06
• Authors: Michael B. Mayhew; Elizabeth Tran; Kirindi Choi; Uros Midic; Roland Luethy; Nandita Damaraju; Ljubomir Buturović
• Keywords: Bayesian Optimization; Hyperparameter Optimization; Hyperparameter

Acute infection, if not rapidly and accurately detected, can lead to sepsis, organ failure even death. Current detection of acute infection as well assessment a patient's severity illness are imperfect. Characterization immune response by quantifying expression levels specific genes from blood represents potentially more timely precise means accomplishing both tasks. Machine learning methods provide platform leverage this host for development deployment-ready classification models. Prioritization promising classifiers is dependent, in part, on hyperparameter optimization which number approaches including grid search, random sampling Bayesian have been shown be effective. We compare HO the diagnostic in-hospital mortality gene 29 markers. take deployment-centered approach our comprehensive analysis, accounting heterogeneity multi-study patient cohort with choices dataset partitioning objective assessing selected external (as internal) validation. find that outperform those search or sampling. However, contrast previous research: 1) efficient selecting all instances compared sampling-based 2) we note marginal gains classifier performance only circumstances when using common variant (i.e. automatic relevance determination). Our analysis highlights need further practical, benchmarking healthcare context.