Linking clinical and administrative data to evaluate intensive care outcomes

Abstract

There is a major focus in health reform in Australia and internationally on monitoring and reporting organisational performance measures, such as standardised mortality rates, in different clinical areas. Given the increasing reliance on existing data for measuring mortality rates, it is important that the accuracy and validity of data are high. Yet, researchers have demonstrated limitations in measuring in-hospital mortality to evaluate intensive care, as it can lead to skewed measurements depending on the discharge practices of different organisations. In Australia, the intensive care clinical registry does not currently measure survival outcomes of patients after hospital discharge but there is interest in doing so. While administrative data sources have the ability to assess outcomes of intensive care patients after hospital discharge, these data may not have sufficient clinical detail to enable robust risk adjustment. Data linkage can be used to link clinical registry and administrative data to enable the measurement of long-term outcomes while using clinical variables to enhance risk-adjustment. However, linkage must be conducted in a robust fashion so that additional error introduced from sub-optimal linkage processes will not bias results. The main aim of the thesis is to assess the utility of linked administrative and clinical data compared to administrative alone and clinical data alone for monitoring long term survival outcomes of ICU patients. The objectives of the thesis are: 1) to define key attributes of linked data for assessing the quality of study results; and 2) to compare the use of linked data to administrative data alone and clinical data alone for a) predicting survival of intensive care patients at 180 days after discharge and b) assessing systematic variation between observed and expected deaths. There were two projects involved in this thesis to address aims 1 and 2, respectively. The first project involved a Delphi consensus process including Australian experts to develop standardised reporting guidelines for assessing the quality of data linkage studies. The resulting guidelines included a list of fourteen items. The guidelines were then applied by two researchers to a stratified selection of data linkage studies to assess their inter-rater reliability (k=0.6). The second project involved the linkage of the Victorian Admitted Episodes Dataset to the Australian/New Zealand Intensive Care Society clinical database of adult critical care patient episodes in the state of Victoria. The linkage procedure was validated to determine whether sources of bias were introduced into the dataset through linkage processes. The added predictive capabilities of the full linked dataset were compared to a model using the administrative data only (C=0.85 v 0.75), a model using the clinical data only (C=0.85 v. 0.84) and a model using a limited sub-set of linked data (C=0.85 v 0.83). Variable Life Adjusted Display (VLAD) charts were developed using both of the linked, administrative and clinical predictive models to determine whether the linked data enhanced the capacity for detecting systematic variation in mortality ratios. It was found that the use of data linkage can enhance the measurement of long-term mortality indicators in intensive care by improving the accuracy of data, risk prediction models and methods for displaying systematic variation in death ratios. Yet, these benefits must be considered together with the limitations of the data, which can influence the accuracy of the linkage process. Identifying and reporting these issues will help to improve data quality and linkage in the future.