This work was performed at the Manitoba Centre for Health Policy as part of a study on the epidemiology of critical illness in the Canadian province of Manitoba . It was approved by the Health Research Ethics Board of the University of Manitoba, and the Manitoba Health Information Privacy Committee.
All Manitoba residents are covered by a universal, comprehensive health insurance system. Our study included all adult Manitobans who received ICU care anywhere in the province during the eight year period from April 1, 2000 to March 31, 2008. The Manitoba population represents a virtually closed system regarding ICU care, as there are no Canadian hospitals providing ICU care within 150 miles of its borders; consequently all Manitoba residents receive their ICU care in Manitoba, except those needing such care while travelling.
Winnipeg is the capital city of Manitoba, and home to 57% of the provincial population of 1.2 million. The six acute care hospitals in Winnipeg contain 11 adult ICUs of various types, with a total of 82 ICU beds. In addition, the province has 10 other hospitals each with a single ICU, totalling 36 ICU beds.
We used two databases for this study. The hospital abstract database (HADB) includes administrative data maintained by the provincial department of health. Containing information about every admission to all Manitoba hospitals from 1974 onwards, HADB data is collected by abstractors located in each hospital. Abstractors are centrally trained, and use uniform definitions, data collection methods, and data entry software. Each hospital abstract spans the time from entry to separation in a single hospital, so a new hospital abstract is created when a patient is transferred from one hospital to another.
The Winnipeg ICU Database (WICUDB) is a clinical database. It contains detailed information about all adult ICU admissions in Winnipeg hospitals. Each WICUDB record spans the time from entry to separation in a single ICU, so a new record is created when a patient is transferred from one ICU to another. WICUDB data is obtained from bedside medical records by specially trained and dedicated personnel, all of whom are former ICU nurses. This information undergoes extensive testing of the validity of patient identification, and parameter values.
Hospitals in Manitoba occasionally care for non-Manitobans. We identified a person as being a Manitoban by the presence of a valid provincial Personal Health Identification Number (PHIN) in both databases. The current analysis was restricted to Manitoba residents, who comprise 95% of ICU patients in the province . Both databases were de-identified by removal of names and addresses, and replacement of PHINs by a unique scrambled version; for simplicity we will use the term PHIN to refer to these scrambled PHINs.
Two previously described preliminary steps for generating the data infrastructure were required . First, we linked each WICUDB record to a single hospital abstract, with a success rate of 99.2%. Second, we demonstrated that hospital abstracts alone accurately identify the presence and timing of ICU care, enabling us to go beyond the bounds of the Winnipeg ICUs and include the entire province in our analysis.
Constructing episodes of ICU and hospital care from individual hospital abstracts and ICU records requires the ability to identify which ones belong to the same individual. The presence of the PHIN in both databases made this a simple task. The fact that each ICU record was linked to a specific hospital abstract facilitated a two stage, “outside-in” approach of constructing the episodes: first we identified the hospital episodes (outside), and then, to identify ICU episodes contained within a hospital episode (inside) we only had to consider the ICU records linked to the hospital abstracts comprising that hospital episode.
Identifying episodes of ICU-containing hospital care
This process began with the creation of a list of each resident who had any ICU care during the study period. Then, for each such individual: (a) the PHIN was used to identify all hospital abstracts for that person during the study period, regardless of whether they contained any ICU care, (b) these were arranged chronologically by hospital entry date, and (c) we used Methods 1–5, detailed below, to identify and combine abstracts which were potentially part of the same hospital episode.
Hospital abstract variables used to identify inter-hospital transfers were the timing of hospital entry and separation, and information about the patient’s location before entry and after separation. In referencing two successive hospital abstracts for the same person, we call the earlier-starting one abstract#1 and the later-starting one abstract#2. The “gap” between abstracts was defined as the interval from hospital separation in abstract#1 until hospital entry in abstract#2.
The mandatory hospital abstract field called SeparationCode indicates the type of location the patient was discharged to, e.g., home or another hospital. In addition, HospitalTransferFrom and HospitalTransferTo variables are intended to identify the specific institution the patient came from or went to; unfortunately, these two data fields are optional and not always completed. Anticipating that the timing and location information may sometimes be recorded incorrectly, as with all data, a certain latitude was appropriate in using them to identify inter-hospital transfers.
We assessed differing methods for combining individual hospital abstracts into hospital episodes. Specifically, we examined the impact of: (a) allowing a ≤1-day versus ≤2-day gap between successive abstracts, and (b) using versus not using the SeparationCode, HospitalTransferFrom and HospitalTransferTo variables. We did not evaluate a gap of zero days because hospital abstracts before 2004 only recorded dates without times, so that inter-hospital transfers using a gap of zero days would be misclassified if they began before midnight and ended after midnight. The rationale for allowing gaps up to two calendar days was to allow for a degree of occasional miscoding of hospital entry and/or separation dates.
We ignored abstract#2 if its admission and discharge dates were completely contained within those of abstract#1. This could occur if a patient was sent from hospital A to hospital B for a planned short time (e.g. a procedure), and anticipating return, the abstract in hospital A was not closed at the time of the transport to hospital B. Furthermore, when using the HospitalTransferFrom
variables as part of identifying inter-hospital transfers, we tested only for the presence of codes representing transfer to or from any acute care hospital; not insisting on identification of the specific hospital was done to account for expected inaccuracy in the coding of specific hospitals. Five combination methods were used to create episodes of ICU-containing hospital care:
Method 1: No transfers. In this method, every hospital abstract was treated as though it was an entire hospital episode. This method intentionally ignored the possibility of transfers, and made no attempt to combine abstracts into episodes, no matter how brief the gap between them. Accordingly, it should over-estimate the number of episodes of care and under-estimate episode LOS, though the degree of inaccuracy is not known. Method 1 served as a baseline for comparison with the other methods, which attempted to identify inter-hospital transfers and combine corresponding abstracts.
Method 2: Allow a ≤1-day gap, and require an indication of inter-hospital transfer. Two hospital abstracts with a gap of ≤1 calendar day were considered part of the same hospital episode only if some indication of patient transfer to or from another hospital was also present, either: (A) the SeparationCode or HospitalTransferTo variable in abstract#1 indicated transfer to another hospital, or (B) the HospitalTransferFrom variable in abstract#2 indicated transfer from another hospital
Method 3: Allow a ≤1-day gap, with no requirement for an indication of inter-hospital transfer. Hospital abstracts with a gap of ≤1 calendar day were considered part of the same hospital episode.
Method 4: Allow a ≤2-day gap, and require an indication of inter-hospital transfer. Similar to Method 2 above, but the abstracts could be separated by up to two calendar days.
Method 5: Allow a ≤2-day gap, with no requirement for an indication of inter-hospital transfer. Similar to Method 3 above, but the abstracts could be separated by up to two calendar days. Method 5 should allow the most combinations of all the methods used, generating the lowest number of episodes of care, and the longest LOS.
The result for each patient was one or more hospital episodes, each constructed from one or multiple hospital abstracts. From these, the hospital episodes containing at least one ICU record (ICU-containing hospital episodes) were retained for subsequent construction of ICU episodes. The LOS of hospital episodes including multiple abstracts was calculated as the elapsed time from the beginning of the initial abstract to the end of the final abstract.
Identifying episodes of ICU care
With ICU-containing hospital episodes identified, those that included only a single ICU record necessarily contained just one ICU episode. Hospital episodes including multiple ICU records could contain one or more ICU episodes, dependent on the presence of inter-ICU transfers. However, any method to identify inter-ICU transfer had to accommodate: (i) the possibility of inter-ICU transfer with a substantial delay, such as an intervening surgery, or transport between ICUs in remote hospitals, (ii) occasional inaccuracy of recorded dates/times, (iii) temporary transfer to an ICU in another hospital (e.g., to perform specific procedures) while the ICU bed in the sending hospital is retained for the patient’s planned return, and (iv) the fact that ICU readmissions sometimes occured after only a brief time on a ward.
The data elements we considered for identifying inter-ICU transfers were the timing of ICU entry and separation, and patient location before and after ICU admission. Unfortunately, pre-ICU and post-ICU location information is problematic in both sources of ICU records. While the WICUDB contains fields for pre-ICU and post-ICU locations, these are only reliable if they are within the six hospitals included in the WICUDB. And as the ICU records for the other 10 provincial do not contain explicit information about the pre-ICU and post-ICU locations. Because of this limitation, the sole criterion used for combining successive ICU records contained within a hospital episode was the time gap between ICU records.
Five methods were used for assessing the number and length of ICU episodes. These were applied in parallel with the methods used for hospital episodes. In Method 1 every ICU record was taken to represent an entire ICU episode. The other methods assessed ICU gaps of ≤1-day or ≤2-days. When a ≤1-day gap was used for adjacent hospital abstracts in Methods 2 and 3, the same gap was allowed between adjacent ICU records; and similarly for the ≤2-day gap in Methods 4 and 5. As above, gaps between ICU records of up to two calendar days allowed for combining records in the presence of a degree of occasional miscoding of ICU entry and/or separation timing. The LOS of ICU episodes including multiple records was calculated as the elapsed time from the beginning of the initial record to the end of the final record.