Databases used to study the care of patients in hospitals and Intensive Care Units (ICUs) typically contain a separate entry for each segment of hospital or ICU care. However, it is not uncommon for patients to be transferred between hospitals and/or ICUs, and when transfers occur it is necessary to combine individual entries to accurately reconstruct the complete episodes of hospital and ICU care. Failure to do so can lead to erroneous lengths-of-stay, and rates of admissions, readmissions, and death.
This study used a clinical ICU database and administrative hospital abstracts for the adult population of Manitoba, Canada from 2000–2008. We compared five methods for identifying patient transfers and constructing hospital episodes, and the ICU episodes contained within them. Method 1 ignored transfers. Methods 2–5 considered the time gap between successive entries (≤1 day vs. ≤2 days), with or without use of data fields indicating inter-hospital transfer. For the five methods we compared the resulting number and lengths of hospital and ICU episodes.
During the study period, 48,551 hospital abstracts contained 53,246 ICU records. For Method 1 these were also the number of hospital and ICU episodes, respectively. Methods 2–5 gave remarkably similar results, with transfers included in approximately 25% of ICU-containing hospital episodes, and 10% of ICU episodes. Comparison with Method 1 showed that failure to account for such transfers resulted in overestimating the number of episodes by 7-10%, and underestimating mean or median lengths-of-stay by 9-30%.
In Manitoba is it not uncommon for critically ill patients to be transferred between hospitals and between ICUs. Failure to account for transfers resulted in inaccurate assessment of parameters relevant to researchers, clinicians, and policy-makers. The details of the method used to identify transfers, at least among the variations tested, made relatively little difference. In addition, we showed that these methods for constructing episodes of hospital and ICU care can be implemented in a large, complex dataset.||