Computing ROC curve confidence intervals for cardiac activation detectors

It is shown how to use the bootstrap to construct a nonparametric confidence interval for a receiver operating characteristic (ROC) curve. The method is used to perform a statistical comparison of a detectors curve with a particular point in ROC space or to compare curves derived from two similar detectors. The method is illustrated with a specific example in which a computer algorithm (AP) that was designed to detect cardiac activations in recordings taken from extracellular bipolar electrodes is evaluated. Twelve recordings of ventricular fibrillation were made on the epicardial surface of the heart in anesthetized mongrel dogs. Each recording contained between 115 and 237 activations. A cardiologist examined each recording, identified all activation events, and assigned a time of local activation (lat_C) to each event. Next, each recording was analyzed by the computer program, AP, which produced its own list of local activation times (lat _AP). The total numbers of true detections (N_TD ), false detections (N_FD), and false negatives (N_FN) were calculated for each recording. A true detection was defined as a lat_AP that fell within ±6 ms of a lat_C