Estimating Episodes of Care Using Linked Medical Claims Data

Australia has extensive administrative health data collected by Commonwealth and state agencies. The value of the health data is how quickly and effectively the data can be converted into useful knowledge independent of its quantity and quality. Using a unique cleaned and linked administrative health dataset we address the problem of empirically defining episodes of care. An episode of care is a time interval containing medical services relating to a particular medical situation. In this paper the medical situation is a hospital admission. The medical services of interest are pathology tests, diagnostic imaging and noninvasive investigative procedures performed before or after the hospital admission, but ‘associated’ with the hospital admission. The task can be viewed as detecting a signal in a time series relating to a hospital admission, distinct from the background noise of on-going medical care. Our approach uses an ensemble (panel of experts) paradigm where we implement multiple agents (alternative predictive models) to independently estimate intervals and then combine good interval estimates using a voting scheme. The results have been used in a study for the Commonwealth Department of Health.