Analysis of health trajectories using administrative data

Abstract

This thesis presents studies of health trajectories using administrative data. In the first study, we propose a new method for clustering of Medicare Benefit Schedule (MBS) claim trajectories. In the second study, we examine the extent to which the adult Australian population on lipid-lowering medications receives the level of High-Density Lipoprotein-Cholesterol (HDL-C) testing recommended by national guidelines. Finally, we study mental health services and medications usage in the MBS and Pharmaceutical Benefit Scheme (PBS) administrative data for the purpose of finding groups of patients with similar utilisation patterns. For these studies, we analyse records from seven years (2008-2014) of the 10% publicly available sample of de-identified, individual level, linked MBS and PBS administrative data. In the first study, we apply a Hierarchical Deep Belief Networks (HDBN) to cluster individuals' health trajectories in four types of services: general practitioner attendances, specialist attendances, pathology tests, and diagnostic imaging. In the second study, the PBS data is used to identify individuals on stable lipid-lowering medications. The MBS data is used to estimate the annual frequency of HDL-C testing. We develop a methodology to address the issue of "episode coning" in the MBS data, which causes an undercounting of pathology tests. We use a published figure on the proportion of unreported HDL-C tests to correct for the undercounting and estimate the probability that an HDL-C test is performed. The rate of HDL-C testing is then compared to national guidelines that people at high-risk for cardiovascular disease undergo annual testing, to determine appropriateness. For mental health study, we create individual level utilisation patterns describing the sequence of mental health services and medications, extracted from the MBS and PBS data, respectively. We propose an Extended Inter-Spike Interval (EISI) metric to estimate the pairwise distances between the individuals' utilisation patterns. Then, we develop a split-and-merge Partitioning Around Medoids (PAM) algorithm to cluster the study population and discover "interesting" utilisation patterns. In order to better understand the extent to which particular personal characteristics impact an individual utilisation pattern, we perform descriptive and multivariate analyses with gender, age, state of residence, and concessional status as covariates. For the health trajectories clustering study, we applied the proposed HDBN algorithm to cluster one million health trajectories of the New South Wales patients with the age of 45-55 years extracted from the MBS data and detected 31 clusters. For the HDL-C testing study, we estimated that approximately 50% of the population on stable lipid-lowering medications did not receive any HDL-C test in each year. We also found that approximately 19% of the same population received two or more HDL-C tests a year. These levels of underutilisation and overutilisation have been changing at an average rate of 2% and -4% a year, respectively, since 2009. The yearly expenditure associated with test overutilisation was approximately A$4.3M during the study period, while the cost averted because of test underutilisation was approximately A$11.3M a year. For the mental health study, after having excluded obvious and common utilisation patterns, we find that mental health patients can be grouped into 10 clusters with distinct and interpretable utilisation patterns. We find that patients differ in the composition of mental health services and medications and the length of use of those services. The largest cluster (27.1% of the study population) is composed of individuals who only visit general practitioners and take psycholeptics medications for a short period of time. The smallest cluster (4.4% of the study population) contains individuals that have occasional visits with general practitioners, and regularly utilise both psycholeptics and psychoanaleptics medications over long periods of time.

Date of Award	2020
Original language	English