Best Practice to Extract In-hospital Mortality Label

[oezyurty]

Hi, I want to be working on in-hospital mortality prediction based on a variety of lab measurements provided in the dataset.

For this task, I would like to kindly ask what would be the best practice to extract high-quality in-hospital mortality labels.

I know that dateofdeath and dischargedat in admissions table could be used for such purpose. However, as stated in this issue dateofdeath is coming from a different source. Therefore, I want to ask you about how to handle the time difference between dateofdeath and dischargedat . To make it more concrete, would you recommend the following: if (dateofdeath - dischargedat) is less than one day (i.e. dateofdeath is later but not more than 1 day), it still can be counted as in-hospital mortality?

PS: The time diff (dateofdeath - dischargedat) didn't look to have a clear cut-off point for the above scenario (i.e. 12 hours, 1 day, 2 days etc.). That's why I need your expert opinion before I get to start.

Best regards,

[T3hlo]

Not sure if this helps but if you have look at the wiki you could maybe use the destination string as this indicates if the patient died:

department the patient has been discharged to or 'Overleden' if the patient died during the ICU/MCU admission

[patrickthoral]

Technically, you cannot reliably determine in hospital mortality, because the in-patient stay is not documented in the database, only the ICU stay. As @T3hlo mentions, a non-'Overleden destination means that the patient did not die on the ICU. For patients that have a non-null dateofdeath value that is relatively close to dischargedat, it is reasonable to assume that they died in the hospital, but it could also mean that the patient died at home, nursing home, etc. Depending on your use case, you may be able to use 30-day mortality instead of in-hospital mortality.