This bug was fixed in 10.4. You will be able to select and delete the broken arrow, and the validate button will alert you if this occurs. If you can upgrade to 10.4, that would be the easy fix. 

In general, to fix this pre-10.4, you'd have to delete the element and add it back to the process, but you would run in to the issue of terminating in-progress instances stuck on that element in this case. The workaround would be to update the process to avoid the broken element, then delete the broken element once the active instances have successfully moved off the broken element. 

To do that, copy the broken element and paste the new element right next to the broken one. Add a gateway before the broken instance that is configured to never go to the broken element and always go to the new, working element. This should fix all instances going forward. 

To deal with the active instances waiting the 2 days, I'd need to see a little more of the process. Conceptually, the goal would be to add an outflow from the broken element to the next user task which would send each active instance to both the following step AND back into the 2 day wait period. If you let this run for 2 days, all of the active instances stuck in there should move on to the next step. At that point, delete the outflow from the broken element to the next step, allowing the active instances to move on with their flow and trapping the other thread in the 2 day delay. Let the active flow continue on until the instances all complete. Once all the instances complete, you can delete that broken timer and terminate the now finished instances. Once that's done, you can clean up the process, remove the gateway, and just have the one working delay. 

It is a bit tedious, so upgrading to 10.4 and simply removing the bad sequence flow may be the best way to go.