Figure 7: Validation of the 52-gene dual-network classifier.

(a) Members of the network as a classifier of neonatal infection using leave-one-out cross-validation analysis using four independent machine-learning algorithms (red circle, random forest; green triangle, support vector machines; blue+, K nearest neighbour and black x, receiver operator characteristics). (b) Flow chart of biomarker classifier validation. Flow chart showing the samples used for classifier generation and the validation results. (c) Heat map showing hierarchical clustering of 18 infant samples (9 bacterially infected, 6 control, 3 virally infected) based on the 50 probes that were in common between the classifier gene set and the Affymetrix U219 platform. Hierarchical clustering was based on Euclidean distance. Control, blue; bacterial infected, red; viral infected, black. Classification of bacterial infection is indicated (0=non-infected, 1=infected). (d) Comparison of microarray-based classifier and expert assessment for classification of samples from patients with suspected infection. (Top) Comparison of expert assessment, CD69/FCGR1A and the 52-gene classifier on samples scored ‘high’ and ‘low’ likelihood of infection by expert assessment. (Middle) 52-gene classifier prediction and expert assessment of suspected sepsis cases (red=concordance of ‘infection’, blue=concordance of ‘control’, dark grey=discordance of microarray classifier and expert assessment). (Bottom left) Comparison of CD69/FCGR1A and 52-gene classifier of samples scored ‘medium’ likelihood of infection by expert assessment. Classification of bacterial infection is indicated (0=non-infected (pale blue), 1=infected (pink)). (Bottom right) Sensitivity, specificity and accuracy of CD69/FCGR1A and the 52-gene classifier against expert classification for suspected samples from d top. A heat map showing the 30 infant samples of suspected infection based on the 46 probes that were in common between the classifier gene set and the CodeLink platform is shown in Supplementary Fig. 7. (e) Bar plots showing clinical criteria in suspected infection cases as judged by the 52-gene classifier and expert assessment. Days on antibiotics and neutrophil counts are shown for samples based on classification of control (pale blue) and bacterial infection (pink) showing median and standard error of the median. Days on antibiotics, classifier prediction: infected: n=15 (excluding 2 missing values); control: n=12 (excluding 1 missing value). Days on antibiotics, expert assessment: infected: n=5 (excluding 1 missing value); control: n=14 (excluding 1 missing value). Neutrophil count, classifier prediction: infected, n=17; control, n=13. Neutrophil count, expert assessment: infected: n=6; control: n=15.