I am looking for a list of patients in MIMIC-II or MIMIC-III who have been annotated as diabetic by experts. I.e. a gold standard, physician-validated, data set for diabetic patients in MIMIC-II/III.
Here is my take:
is it appropriate to simply ask for “diabetic” patients? Would the method of identifying the cohort vary wildly depending on the type of diabetes of interest? A: It would not be appropriate to simply ask for "diabetic" patients. Patients with diabetes are usually classified by the type (Type 1 or Type 2, which roughly correlates to whether they have diabetes as a result of pancreatic islet cell destruction or insulin resistance, respectively. This becomes significant because a Type 1 diabetic will be reliant on basal insulin otherwise they could very rapidly go into starvation ketosis, whereas a type 2 usually makes some basal insulin and they are less likely to develop ketoacidosis). Furthermore, clinically we would distinguish among type 2 diabetics, which ones receive insulin. This likely doesn't matter as much in MIMIC as critically ill diabetics will almost universally receive insulin rather than oral medications. If the only goal were to identify who has diabetes, in theory this wouldn't matter as much.
what is the best indicator of diabetes in MIMIC? Can we be confident that most cases would be labelled with the appropriate ICD-9 diagnosis code? Or should we also be cross-referencing medications/lab tests/DRG codes/CPT codes? A: ICD9 codes should work, and would likely give you the granularity to identify type 1 (likely not so common) as well as type 2 (much more common). If you wanted to cross validate you could do so with medication orders for all types of insulin (may not need administration of insulin because many times people's insulin requirement in the hospital varies, and instead is done via a "sliding scale" so if their blood sugar is controlled, they may not receive insulin which was ordered.)
what does “gold-standard” mean in this case? Does gold-standard imply that a diagnostic test for diabetes has been explicitly carried out? Would gold standard require each patient record to be retrospectively reviewed and annotated by a physician? A: The gold standard will be hard to deduce in our MIMIC cohort. The diagnostic criteria for diabetes is usually performed as an outpatient, but includes:
-Have symptoms of diabetes (increased thirst, increased urination, and unexplained weight loss) and a blood sugar level equal to or greater than 200 milligrams per deciliter (mg/dL). The blood sugar test is done at any time, without regard for when you last ate (random plasma glucose test or random blood sugar test).
-Have a fasting blood sugar level that is equal to or greater than 126 mg/dL. A fasting blood sugar test (fasting plasma glucose) is done after not eating or drinking anything but water for 8 hours.
-Have a 2-hour oral glucose tolerance test (OGTT) result that is equal to or greater than 200 mg/dL. An OGTT is most commonly done to check for diabetes that occurs with pregnancy (gestational diabetes).
-Have a hemoglobin A1c that is 6.5% or higher.
This must then be repeated on a subsequent test.
This most likely will have been performed as an outpatient and thus the "gold standard" for diagnosis diabetes would not be captured in MIMIC and you would likely rely on ICD9 codes + administering/ordering insulin as your check.
Source: American Diabetes Association (2012). Diagnosis and classification of diabetes mellitus. Diabetes Care, 35(Suppl 1): S64-S71.