Burst-method comparison on 8,746 simulator-labeled ISI intervals, including 498 Burst intervals.
Positive class: simulator Burst intervals. STPD positive class: burst-family outputs (burst, possible_burst, long_burst).
Mean-ISI and LogISI were run in their automatic support-layer mode in both panels; simulator-derived thresholds were imported only for the simulator-informed STPD detector.
The always-negative reference predicts no Burst intervals and is included to show how overall accuracy is inflated by the non-burst majority.
Because the benchmark is strongly non-burst imbalanced, precision, recall, F1, and balanced accuracy are more informative than overall accuracy.
