This is the explanation from here:
If you need static thresholds rather than the baseline-based alerting approach, change from automated mode to fixed thresholds, as shown below. The sensitivity controls the level of statistical confidence required to raise an event. Low sensitivity means a high confidence is required, and vice versa. For example, to see events immediately, even if only a few data points have breached the threshold, select high sensitivity.
It's confidence in the statistical sense, as to what goes into that I do not know and I have not seen it described in detail. I can forward this to some folks to see if they consider it appropriate to expand on it.