Multi-class gene expression biomarker panel identification for the diagnosis of paediatric febrile illness

Abstract

Febrile illness in children can result from infections by diverse viral or bacterial pathogens as well as inflammatory conditions or cancer. The limitations of the existing diagnostic pipeline, which relies on clinical symptoms and signs, pathogen detection, empirical treatment and diagnoses of exclusion, contribute to missed or de- layed diagnosis and unnecessary antibiotic use. The potential of host gene expression biomarkers measured in blood has been demonstrated for simplified binary diagnostic questions however, the clinical reality is that multiple potential aetiologies must be considered and prioritised on the basis of likelihood and risks of severe disease. In order to identify a biomarker panel which better reflects this clinical reality, we applied a multi-class supervised learning approach to whole blood transcriptomic datasets from children with infectious and inflammatory disease. Three datasets were used for the analyses presented here, a single microarray dataset, a meta-analysis of 12 publicly available microarray datasets and a newly generated RNA-sequencing dataset. These were used for preliminary investigations of the approach, discovery of a multi-class biomarker panel of febrile illness and valida- tion of the biomarker panel respectively. In the merged microarray discovery dataset a two-stage approach to feature selection and classification, based on LASSO and Ridge penalised regression was applied to distinguish 18 disease classes. Cost-sensitivity was incorporated in the approach as aetiologies of febrile illness vary considerably in the risk of severe disease. The resulting 161 transcript biomarker panel could reliably distinguish bacterial, viral, inflammatory, tuberculosis and malarial disease as well as pathogen specific aetiologies. The panel was then validated in a newly generated RNA-Seq dataset and compared to previously published binary biomarker panels. The analyses presented here demonstrate that a single test for the diagnosis of acute febrile illness in children is possible using host RNA biomarkers. A test which could distinguish multiple aetiologies soon after presentation could be used to reduce unnecessary antibiotic use, improve targetting of antibiotics to bacterial species and reduce delays in the diagnosis of inflammatory diseases.