If there are no deaths in those patients who had Leukemia, then the model can't estimate the 'death rate' associated with those patients - hence the unreliable estimates - you should really get a warning / error message from Strata [any time you get a ridiculous CL - your software should really tell you that parameter is unfittable]. In our Statistical Analysis plans, we also write text stating that any sub-group analyses will only be performed if a minimum number of patients belong to any subgroup. For Proportional Hazard models that text is amended to a requirement for a minimum number of events in any subgroup. Hence before fitting your model you should first summarise the number of patients and number of deaths associated with each of your comorbidities. I would suggest any comorbidity that occurs in less than 10% of your total sample size should be excluded from your model, and that ideally you require at least 20 deaths (for a reliable estimate of the death rate) for each of your comorbidities. These numbers are not fixed rules of thumb, if you had a million patients then 10% would be too high - but 20 deaths probably too low. If you had only a 100 patients then I wouldn't even bother looking at subgroups.

Very wide CIs indicate a small number of cases. This tells you that the OR estimate is highly unreliable.