Integrating classifiers across datasets improves consistency of biomarker predictions for sepsis. Issue 26 (2016)

• Record Type: Journal Article
• Authors: Saraiva, João Pedro; Oswald, Marcus; Biering, Antje; Assmann, Cora; Klassert, Tilman; Blaess, Markus; Czakai, Kristin; Claus, Ralf; Löffler, Jürgen; Slevogt, Hortense; König, Rainer
• Is Part Of: IFAC-PapersOnLine. Volume 49:Issue 26(2016)
• Page Start: 95
• Page End: 102
• Publication Date: 2016
• Subjects: Machine Learning -- Consistency -- Biomarker -- Feature Selection -- Sepsis -- Infection; Automatic control -- Periodicals; 629.805
• DOI: 10.1016/j.ifacol.2016.12.109
• Languages: English
• ISSNs: 2405-8963

Abstract

Abstract: Systemic infection can cause multiple organ failure leading to severe sepsis and often death. Hence, early diagnosis is mandatory. Several transcriptomics studies were performed resulting in biomarker lists for diagnosis. This lists, however are very inconsistent. We developed Mixed Integer Linear Programming based classifiers (Support Vector Machines), trained them separately with different datasets, and combined them by constraining them to use the same sets of features. Strikingly, this improved the consistency of the predicted biomarkers across datasets by 42%. Our approach is generic; it enabled to integrate diverse datasets and, with this, improved the consistency of predictions.