theonesp / hr_vs_lr_repos Goto Github PK

Can we put together a Table 1 for the NIS and eICU-CRD cohorts? It is crucial that we have comparable cohorts for the 2 databases. If not, we have to come up with a way to select comparable cohorts from the 2 databases.

Sample Size
Age (mean)
Gender (% males)
Co-morbidities (% of total, e.g. Congestive Heart Failure, Chronic Kidney Disease, etc.)
Illness severity score (we have to figure out the mapping of the administrative data-based score with APACHE-IV)
Proportion of patients on mechanical ventilation
Proportion of patients who receive new renal replacement therapy
Hospital mortality
Hospital length-of-stay of survivors

To reiterate, we will be confining the study to patients with sepsis based on Angus criteria admitted in 2014 and 2015. We are excluding patients who have ESRD on hemodialysis.

Columns

• impute missing data on those where missing <= 30%
• exclude those with Missing > 30%
• Should I try to fine tune one of the columns excluded in case they are fundamental for the study?

Modeling

• Run a random forest
• Run RF OR table

• restrict the analysis to mech. vent only
• >= 16 yo
• sepsis on apache dx
• drop those with no AKI
• include patients from 2015
• stratify by mortality

1. Seems like some outliers in the height and weight - looking at the upper limit
   I have improved the extraction query and normalized it, it's looking better now.

2. Need to discuss if we are keeping OSH admissions- granted small number 1.5% but sometimes difficult to tell if they received any RRT there- I am not tied to any particular approach (incl/excl)
   Please let me know about this. What do you think @leo Anthony Celi ?

3. What is the distinction between hospital admit source and unit admit source?

   • hospital admit source = location from where the patient was admitted to the hospital e.g.: Direct Admit, Floor, Chest Pain Center. etc.
   • unit admit source = picklist location from where the patient was admitted e.g.: Emergency Room, Recovery Room, Direct Admit, etc.

4. Depending on the definition: ICUs could be categorized into medical/surgical/mixed medical surgical
   Please let me know about this.

5. Wasn't sure why we were looking at specific unit discharge times? were we trying to tie in with shift change? If the idea is to look at discharges at night versus day- could dichotomize- again would need to know the rationale.
   I don't think we need discharge times, I just extracted it since it was in the demographics table. @leo Anthony Celi can you please confirm?

6. Hospital ID: Are we considering all hospitals or ones which have been contributed data for some period of time (say 3 years + or so) either ways its fine. We have used such criteria to from some ACC AHA GWTG database studies. However reasonable to keep as is if numbers diminish.
   We are studying 2014 only since this is the data Barret is using, we are not taking Hospital ID into account in the exclusion criteria.

7. Vent days (has an upper limit of 347)- recheck if you can- might be one outlier or so- might want to consider how to deal with it.
   There were less than 3 patients with outliers so I removed those data points.

8. Chronic dialysis : We are excluding these patients I suppose - since we are looking at de novo RRT
   Yo are correct. we have two ways of knowing who is getting rrt, we might want to be less strict about this since if we use both criteria we lose too many patients receiving rrt, we have only ~ 300.
   The two ways are:
   apacheapsvar.dialysis = 1 -- chronic dialysis prior to hospital adm
   (treatment.treatmentstring) LIKE ANY ('{%rrt%,%dialysis%,%ultrafiltration%,%cavhd%,%cvvh%,%sled%}') AND (treatment.treatmentstring) LIKE '%chronic%'
   So as you can see one is coming from the apache table and the other one from treatment.

9. APACHE is this II or III? Think it is 3. please look at the lower limit (negative) usually starts at 0
   Contains the variables used to calculate the Acute Physiology Score (APS) III for patients.

10. AKI offset- could you please elaborate
    The nephrologist suggested not to use this variable since we can not infer when AKI was developed from it. Removing it to avoid confusion

11. UOP - negatives for days--- would need to be looked into
    There were less than 10 patients with outliers so I removed those data points

12. Similarly dialysis output- Negative needs to be looked into
    There were less than 10 patients with outliers so I removed those data points

13. DM is listed twice with 2 different %s
    diabetes = diabetes from apachepredvar table: variables underlying the APACHE predictions.
    dm = Diabetes from pasthistory table: Provides information related a patient’s relevant past medical history. Links to: PATIENT on patientUnitStayID Important considerations Providing detailed Past History is not common, but items such as AIDS, Cirrhosis of the Liver, Hepatic Failure, Chronic Renal Failure, Transplant, and Pre-existing Cancers / immunosuppression are more reliable because of their importance in severity outcome scoring.

14. Is it SOFA on day 1
    You are correct

15. Mech vent score- could you elaborate?
    Used to calculate OASIS:
    -- This query extracts the Oxford acute severity of illness score in the eICU database.
    -- This score is a measure of severity of illness for patients in the ICU.
    -- The score is calculated on the first day of each ICU patients' stay.
    -- Reference for OASIS:
    -- Johnson, Alistair EW, Andrew A. Kramer, and Gari D. Clifford.
    -- "A new severity of illness scale using a subset of acute physiology and chronic health evaluation data elements shows comparable predictive accuracy*."
    -- Critical care medicine 41, no. 7 (2013): 1711-1718.
    -- Variables used in OASIS:
    -- Heart rate, GCS, MAP, Temperature, Respiratory rate, Ventilation status
    -- Urine output
    -- Elective surgery
    -- Pre-ICU in-hospital length of stay
    -- Age
    -- Note:
    -- The score is calculated for all ICU patients, with the assumption that the user will subselect appropriate ICUSTAY_IDs.
    -- For example, the score is calculated for neonates, but it is likely inappropriate to actually use the score values for these patients.

Other things

These could be looked into/pulled of course within the limits of the database:

1. Hypotensive minutes : duration of hypotension (can choose <65mmHg); use of vasopressor; 
2. History of CKD on admission
3. Other risk factors such as acidosis ie pH<7; use of nephrotoxic agents (common ones like Abx/contrast) and anemia