Figure 1. Breath test outcome in healthy women (normal screening mammogram) versus women with breast biopsy positive for cancer.

Identification of breath biomarkers (upper panel): A list of candidate breath biomarkers of disease was obtained by segmenting chromatograms into a time series of alveolar gradients, where the alveolar gradient comprised detector response in breath minus corresponding detector response in room air. The diagnostic accuracy of each candidate biomarker was quantified as the area under curve (AUC) of its associated receiver operating characteristic (ROC) curve. This figure displays the number of candidate biomarkers (y-axis) as a function of their diagnostic accuracy (x-axis). The “correct” curve employed the correct assignment of diagnosis (normal mammogram or cancer on biopsy). The “random” curve employed multiple Monte Carlo simulations comprising 40 random assignments of diagnosis in order to determine the random behavior of each candidate biomarker. The horizontal separation between the “correct” and “random” curves varies with the amount of diagnostic information in the breath signal. Where the number in the “random” curve declines to <1, its vertical distance from the “correct” curve identifies the excess number of candidate biomarkers that identified the disease group with greater than random accuracy. The number of apparent biomarkers with greater than random accuracy exceeded 30, but several segments were closely adjacent in the time series, consistent with approximately 10 biomarker peaks in the chromatogram. Similar analyses were also performed in normal versus abnormal screening mammograms and cancer versus no cancer on breast biopsy in order to develop separate algorithms. Diagnostic accuracy of the breath test (lower panel): The ROC curve displays the breath test's accuracy in distinguishing healthy women with a normal screening mammogram from women whose breast biopsy was positive for cancer. The breath test employed a multivariate predictive algorithm derived from the biomarkers with greater than random accuracy that were identified with the Monte Carlo simulations in the left panel. Sensitivity and specificity values were determined from the point on the ROC curve where their sum was maximal. Cross-validation of predicted outcomes is shown in red ROC curve. A repeated leave-one-out bootstrap method was employed to estimate the prediction error (method described in text).