Not sure if you ever got an answer... but I have seen this.
Mostly on hight latency networks with bandwidth limitations. QoS policy has prioritized voice and video and left everything else best effort.
The hello packets used to keep EIGRP neighbor states up are ICMP. During high traffic periods, these ICMP messages were getting dropped. In order to keep the tunnel states up, we had to extend the hold-down timers and hello-intervals.
config# interface tunnel <tunnel#>
config-if# ip hello-interval eigrp <as#> 15
config-if# ip hold-time eigrp <as#> 45
15 and 45 worked for us. Of course you can play with the time settings to meet your specifications.