Diagnostic Test Accuracy of a 2-Transcript Host RNA Signature for Discriminating Bacterial vs Viral Infection in Febrile Children

Abstract

Importance: Because clinical features do not reliably distinguish bacterial from viral infection, many children worldwide receive unnecessary antibiotic treatment, while bacterial infection is missed in others.

Objective: To identify a blood RNA expression signature that distinguishes bacterial from viral infection in febrile children.

Design, setting, and participants: Febrile children presenting to participating hospitals in the United Kingdom, Spain, the Netherlands, and the United States between 2009-2013 were prospectively recruited, comprising a discovery group and validation group. Each group was classified after microbiological investigation as having definite bacterial infection, definite viral infection, or indeterminate infection. RNA expression signatures distinguishing definite bacterial from viral infection were identified in the discovery group and diagnostic performance assessed in the validation group. Additional validation was undertaken in separate studies of children with meningococcal disease (n = 24) and inflammatory diseases (n = 48) and on published gene expression datasets.

Exposures: A 2-transcript RNA expression signature distinguishing bacterial infection from viral infection was evaluated against clinical and microbiological diagnosis.

Main outcomes and measures: Definite bacterial and viral infection was confirmed by culture or molecular detection of the pathogens. Performance of the RNA signature was evaluated in the definite bacterial and viral group and in the indeterminate infection group.

Results: The discovery group of 240 children (median age, 19 months; 62% male) included 52 with definite bacterial infection, of whom 36 (69%) required intensive care, and 92 with definite viral infection, of whom 32 (35%) required intensive care. Ninety-six children had indeterminate infection. Analysis of RNA expression data identified a 38-transcript signature distinguishing bacterial from viral infection. A smaller (2-transcript) signature (FAM89A and IFI44L) was identified by removing highly correlated transcripts. When this 2-transcript signature was implemented as a disease risk score in the validation group (130 children, with 23 definite bacterial, 28 definite viral, and 79 indeterminate infections; median age, 17 months; 57% male), all 23 patients with microbiologically confirmed definite bacterial infection were classified as bacterial (sensitivity, 100% [95% CI, 100%-100%]) and 27 of 28 patients with definite viral infection were classified as viral (specificity, 96.4% [95% CI, 89.3%-100%]). When applied to additional validation datasets from patients with meningococcal and inflammatory diseases, bacterial infection was identified with a sensitivity of 91.7% (95% CI, 79.2%-100%) and 90.0% (95% CI, 70.0%-100%), respectively, and with specificity of 96.0% (95% CI, 88.0%-100%) and 95.8% (95% CI, 89.6%-100%). Of the children in the indeterminate groups, 46.3% (63/136) were classified as having bacterial infection, although 94.9% (129/136) received antibiotic treatment.

Conclusions and relevance: This study provides preliminary data regarding test accuracy of a 2-transcript host RNA signature discriminating bacterial from viral infection in febrile children. Further studies are needed in diverse groups of patients to assess accuracy and clinical utility of this test in different clinical settings.

Figure 1. Study overview

Overall flow of patients in the study showing patient recruitment and subsequent selection for microarray analysis. HC Healthy Control; JIA juvenile idiopathic arthritis; HSP Henoch-Schönlein Purpura; SLE Systemic Lupus Erythematosus; DB Definite Bacterial; PB Probable Bacterial; U Unknown; PV Probable Viral; DV Definite Viral.

Figure 2. Classification of patients into diagnostic groups

Febrile children with infections were recruited to the Immunopathology of Respiratory, Inflammatory and Infectious Disease Study, and were classified into diagnostic groups as described in methods. CRP: C-reactive protein

Figure 3. Analysis workflow

Overall study pipeline showing sample handling, derivation of test and training datasets, data processing, and analysis pipeline including application of 38-transcript elastic net classifier and 2-transcript DRS classifier, to the test set, the validation datasets and published (external) validation datasets. DB Definite Bacterial; PB Probable Bacterial; U Unknown; PV Probable Viral; DV Definite Viral; HSP Henoch-Schönlein Purpura; JIA Juvenile Idiopathic Arthritis; SLE Systemic Lupus Erythematosus; HC Healthy Control; SDE Significantly Differentially Expressed; FC fold change; FS-PLS Forward Selection - Partial Least Squares; DRS Disease Risk Score

Figure 4. DRS and ROC curves based on the 2-transcript signature applied to Definite Bacterial and Viral groups

Classification performance and Receiver Operating Characteristic (ROC) curve based on the 2-transcript Disease Risk Score (DRS) (the combined IFI44L and FAM89A expression values), in the Definite Bacterial and Viral groups of the discovery test set (20% of the total discovery group) (A & B) and the IRIS validation dataset (C & D). Boxes show median with 25^th and 75^th quartiles; whiskers, plotted using ‘boxplot’ in R, extend ≤1 times the interquartile range. Sensitivity, specificity, and AUC are reported in eTable 4. The horizontal DRS threshold line separates patients predicted as bacterial (above the line) or viral (below the line), as determined by the point on the ROC curve that maximized sensitivity and specificity.

Figure 5. Performance of the 2-transcript DRS signature in indeterminate groups

Classification performance of the 2-transcript DRS (the combined IFI44L and FAM89A expression values) in the indeterminate groups of Probable Bacterial, Probable Viral, and Unknown of the discovery test (A) and IRIS validation (B) sets. Boxes show median with 25^th and 75^th quartiles; whiskers, plotted using ‘boxplot’ in R, extend ≤1 times the interquartile range. The horizontal DRS threshold line (threshold_{test_dataset}=-1.03; threshold_{validation_dataset}=-2.63) separates patients predicted as bacterial (above the line) or viral (below the line). It is determined by the point in the ROC curve that maximized sensitivity and specificity. For the test set, the training threshold was used.