I already explained the feature request in the forum: http://www.zabbix.com/forum/showthread.php?t=20490

One could implement 'median' exactly as its definition here: http://www.zabbix.com/forum/showthread.php?t=20490, but in the past I had very good results with the average value with 35% of top and bottom values discarded (with 3 samples this would be equal to the median). Maybe this value (35%) should be variable too depending on the amount of samples.

I want to keep it open how exactly the function should be implemented, but I believe there's a need in some applications to have a 'fuzzy' approach to prevent false positives. With only max, min, avg and last it's just not enough.