Article Text

Download PDFPDF

1723 Diagnostic accuracy of a novel transcriptomic classifier for bacterial and viral infections – an individual patient data meta-analysis
Free
  1. Florian Uhle1,
  2. Pilot Study Investigator group, Evangelos Giamarellos-Bourboulis2,
  3. Wolfgang Bauer3,
  4. Lisa Sun4,
  5. Uan-I Chen4,
  6. Timothy E Sweeney4,
  7. Oliver Liesenfeld4
  1. 1Inflammatix Inc
  2. 2National and Kapodistrian University of Athens, Medical School, Greece
  3. 3Department of Emergency Medicine, Campus Benjamin Franklin, Charite Universitaetsmedizin Berlin, Germany
  4. 4Clinical Affairs, Inflammatix Inc., Redwood City, USA

Abstract

Aims, Objectives and Background Arising from demographic differences between healthcare systems, patients in the emergency department (ED) present with a broad range of diagnoses and clinical severities. Independent validation of a novel diagnostic tool is critical to ensure reliable and reproducible clinical performance. So far, three independent cohort studies have validated the performance of the machine-learning classifier IMX-Bacterial/Viral/Non-infected (IMX-BVN) to diagnose bacterial and viral infections; those have been combined to facilitate an individual patient data meta-analysis of performance.

Method and Design ED patients (n=1,277) with suspected infection from three international, observational studies (USA/Germany/Greece) were included. Of those, 661 had a unanimous (‘consensus’) ground truth of infection status established by panel adjudication. Quantitative expression of 29-signature mRNAs was measured on a NanoString nCounter® SPRINT system. The classifier BVN version 3 (IMX-BVN-3) was applied to generate scores, which fall into four discrete interpretation bands (very unlikely, unlikely, possible, very likely). Sensitivity, specificity, and corresponding nominal likelihood ratios were calculated with 95% confidence intervals for each interpretation band.

Results and Conclusion 360 patients (54.4%) were consensus adjudicated to have a bacterial infection (range: 37.9–81.2%) and 153 (23.1%) to have a viral infection (range: 15.3–44.1%). Pooled likelihood ratios of the interpretation bands for bacterial infections were (from ‘very unlikely bacterial’ to ‘very likely bacterial’) 0.082 (0.039–0.176)/0.333 (0.264–0.419)/2.244 (1.598–3.152)/9.459 (5.808–15.404), associated with a rule-in specificity of 0.947 (0.915–0.967) and a rule-out sensitivity of 0.981 (0.960–0.991) in the outer interpretation bands. Pooled likelihood ratios of the interpretation bands for viral infections were (from ‘very unlikely viral’ to ‘very likely viral’) 0.182 (0.102–0.324)/0.292 (0.181–0.471)/0.956 (0.593–1.540)/6.021 (4.636–7.821), associated with a rule-in specificity of 0.884 (0.853–0.909) and rule-out sensitivity of 0.928 (0.876–0.959).

The IMX-BVN-3 classifier exhibits strong performance in a combined cohort of patients from different geographies and settings to rule-in and rule-out patients presenting to EDs with suspected bacterial and viral infections.

Statistics from Altmetric.com

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.