Predicting hospital mortality among frequently readmitted patients: HSMR biased by readmission

BACKGROUND Casemix adjusted in-hospital mortality is one of the measures used to improve quality of care. The adjustment currently used does not take into account the effects of readmission, because reliable data on readmission is not readily available through routinely collected databases. We have studied the impact of readmissions by linking admissions of the same patient, and as a result were able to compare hospital mortality among frequently, as opposed to, non-frequently readmitted patients. We also formulated a method to adjust for readmission for the calculation of hospital standardised mortality ratios (HSMRs). METHODS We conducted a longitudinal retrospective analysis of routinely collected hospital data of six large non-university teaching hospitals in the Netherlands with casemix adjusted standardised mortality ratios ranging from 65 to 114 and a combined value of 93 over a five-year period. Participants concerned 240662 patients admitted 418566 times in total during the years 2003 - 2007. Predicted deaths by the HSMR model 2008 over a five-year period were compared with observed deaths. RESULTS Numbers of readmissions per patient differ substantially between the six hospitals, up to a factor of 2. A large interaction was found between numbers of admissions per patient and HSMR-predicted risks. Observed deaths for frequently admitted patients were significantly lower than HSMR-predicted deaths, which could be explained by uncorrected factors surrounding readmissions. CONCLUSIONS Patients admitted more frequently show lower risks of dying on average per admission. This decline in risk is only partly detected by the current HSMR. Comparing frequently admitted patients to non-frequently admitted patients commits the constant risk fallacy and potentially lowers HSMRs of hospitals treating many frequently admitted patients and increases HSMRs of hospitals treating many non-frequently admitted patients. This misleading effect can only be demonstrated by an analysis over a prolonged period, but occurs, in effect, every day of the year. This finding is relevant for all countries where hospitals use HSMR for monitoring and improving hospital performance. The use of 'admission frequency' as additional adjustment variable may provide a more accurate HSMR.