Have you updated the statistics recently and if so, have you tried updating with a full scan? There is a bit of a discrepancy between the estimated rows & the actual rows. This could be caused by out of date statistics (there are other reasons too, but that's one).
How selective is the data in the TIMESTAMP_NC column?
As bad as the key lookup is for performance, the LOOP join just a little later to the QuestionKeyword table is just as bad. It's doing 4000 seeks on that table. I'm just wandering if updated statistics could help the optimizer to use a hash instead.