Year

2024

Degree Name

Doctor of Philosophy

Department

School of Computing and Information Technology

Abstract

Background: Alcohol use disorder poses a significant health risk to the population of Australia and other countries. It can lead to severe physical and mental health complications, including damage to the central nervous system, memory loss, and an increased risk of chronic conditions and premature death. In addition, alcohol use disorder can cause serious social dysfunction, such as abandonment of social activities and hobbies, family violence, and job loss.

In Australia and New Zealand, the lifetime prevalence of alcohol use disorder ranges from 8.3% to 30.3%, making harmful alcohol consumption the sixth highest risk factor for the Australian burden of disease. Tragically, the rates of alcohol-induced deaths saw a sharp increase during the first year of the COVID-19 pandemic, not just in Australia, but also in the UK and the US. Therefore, it is crucial to understand how patients with alcohol use disorder engage with healthcare services. It helps in determining patient's healthcare needs, developing healthcare resources, training healthcare providers, and planning health promotion, prevention, or protection programs. However, research to date on the journey of patients with alcohol use disorder through health services remains limited.

Aim: This research aims to generate a comprehensive understanding of patients’ utilisation of Emergency Services, Hospital Admission and the community-based Drug and Alcohol Service for patients with alcohol use disorder within the Illawarra and Shoalhaven Local Health District, NSW, Australia. The study also seeks to identify risk factors and predictors in the electronic health records for hospital re-admissions among these patients. This aim is pursued through four specific objectives: (1) to investigate the longitudinal changes in Emergency Department (ED) presentations incurred by patients with alcohol use disorder, (2) to identify the risk factors for 28-day unplanned hospital re-admission in patients with alcohol use disorders, (3) to predict the likelihood of 28-day unplanned hospital re-admission for patients with alcohol use disorder, and (4) to explore the pathway of patients with alcohol use disorder in accessing the community-based Drug and Alcohol Service. By achieving these objectives, the research will provide valuable insights that can inform healthcare resource allocation and patient care strategies.

Methods: To address Objective 1, investigating longitudinal changes in ED presentations incurred by patients with alcohol use disorder, a retrospective quantitative analysis of data from the electronic health records was conducted to investigate differences in ED performance between patients who had interactions with community-based Drug and Alcohol Service and those who did not. The study employed statistical tests to discern specific patterns or correlations in the data. A Chi-square test was conducted to compare the categorical variables. A z-test was then used to compare which subclasses of these categorical variables had statistically significant differences, for example, comparing whether the proportions of patients aged 30-34 were significantly different between the two patient groups. For continuous variables, a Kolmogorov–Smirnov test was applied to examine their distribution. An independent t-test was used if data were normally distributed, otherwise a Mann-Whitney U test was conducted.

To achieve Objectives 2 and 3, identifying the risk factors for 28-day unplanned hospital re-admission and predicting the likelihood, two machine learning studies were conducted to predict 28-day unplanned hospital re-admission for patients with alcohol use disorder. Least Absolute Shrinkage and Selection Operator regression is a statistical method for regression analysis that achieves variable selection and regularization by adding an absolute value penalty term to the regression coefficients. It was used in the first study to screen variables for inclusion in the machine learning models - logistic regression (LR), random forests (RF) and support vector machine (SVM). The second study used univariate and multivariate LR to select input variables. It then included the selected input variables in traditional machine learning models (RF and SVM) and long short-term memory or clinical bidirectional encoder representation of transformers (Clinical Bio-BERT) - to predict 28-day re-admissions. In addition, another set of input variables was included in these machine learning models without screening, thus facilitating the comparison and selection of the most appropriate predictive models.

Finally, to achieve Objective 4, examining how patients access the community-based Drug and Alcohol Service, process mining was conducted to determine the major pathways that the patients with alcohol use disorder followed to receive the various Drug and Alcohol services. An event log was identified as a complete Drug and Alcohol encounter which included service activities, i.e. Intake, Assessment, and treatment services. A series of event logs were analysed using Disco, a process mining application, to identify pathways. Comparisons were conducted on the proportion and duration of services across different demographic groups, including various age groups and genders, as well as between individuals identified as polysubstance users and those identified as non-polysubstance users, using statistical methods including the z-test and non-parametric tests.

Results: The retrospective quantitative analysis for Objective 1 included data from a total of 2,519 individual patients with alcohol use disorders, who made a combined total of 21,715 presentations to ED. Among these patients, 75.4% did not have interactions with the community-based Drug and Alcohol Service. Compared with those who had, these patients were older, more likely to be diagnosed with abdominal pain (26.9% vs 12.0%, p < 0.001) and chest pain (16.2% vs 8.6%, p < 0.001) and had a longer mean length of ED stay (7 hours and 41.7 minutes vs 6 hours and 25.6 minutes, p < 0.001). For the patients who had interactions with the community-based Drug and Alcohol Service, their 28-day re-presentation rates decreased from 55.5% (2013-14) to 45.1% (2017-18); however, these rates were higher than that of those who had no interactions (41.1% to 32.8%). Overall, 21.9% - 24.5% of the patients were frequent ED presenters (i.e., more than 4 visits per year). Frequent ED presenters were proportionately higher among the patients who had interactions with the community-based Drug and Alcohol Service, consistently over the relevant years. Although patients with alcohol use disorders frequently presented to EDs, their alcohol use disorders were only identified in 8.9% of their presentations.

The first study for Objectives 2 and 3 included data from 1,591 patients with alcohol use disorders who had a total of 8,114 hospital admissions. The 28-day all-cause hospital re-admission rate of alcohol use disorder patients was 33.8%. A comparison of the performance indexes of the three models (LR, RF, SVM) revealed that the LR model had the best performance with a higher area under the receiver operating characteristic curve than the other two models, suggesting that its use is more suitable for making re-admission predictions.

The second study for Objectives 2 and 3 included data from 869 patients with alcohol use disorders who had a total of 2,254 hospital admissions. Patients aged 45~49, 70~74 or 75~79 were 4~5 times more likely to be re-admitted than those in other age groups; males were 36.4% more likely than females; polysubstance users were 3.3 times more likely to be re-admitted than otherwise. Patients with “respiratory system disorders” or “hepatobiliary system and pancreas disorders” had 60% higher risk than otherwise. Patients who had interactions with ED or Drug and Alcohol Service between the index admission and the unplanned re-admission had 70.6% and 78.5% reduced risk of being re-admitted within 28 days than those without these interactions, respectively (Odds Ratio [OR]=0.294, 95% Confidence Interval [CI]=0.197~0.439, p

A total of 571 patients with alcohol use disorders were included in the patient access to the community-based Drug and Alcohol service analysis. They incurred 1,447 encounters, comprising 13,974 activities. The top three actual pathways with at least three activities were (1) pathway ‘Intake->Withdrawal Management->Withdrawal Management’ occurring in 170 (11.7%) encounters and experienced by 130 (22.8%) patients; (2) pathway ‘Intake->Adult Counselling->Adult Counselling’ occurring in 161 (11.1%) encounters and experienced by 126 (22.1%) patients; and (3) pathway ‘Intake->Assessment->Withdrawal Management’ occurring in 155 (10.7%) encounters and experienced by 129 (22.6%) patients. However, 313 (21.6%) encounters did not proceed beyond the Intake stage. Patterns uncovered from the actual pathways showed that when the patients started their pathways from Intake, their immediate next activities were frequently Withdrawal Management, Assessment or Adult Counselling. There were differences in the direct care activities and pathway patterns between polysubstance users and non-polysubstance users.

Conclusion: Through data mining of electronic health records and using machine learning, this doctoral project delves into the presentation of alcohol use disorder patients in ED, the risk factors and prediction of hospital re-admission, and the connection pathways with the community-based Drug and Alcohol Service. These studies demonstrated the crucial role of comprehensive and integrated approaches within the healthcare system in effectively addressing the needs of patients with alcohol use disorders. Key aspects to enhance the utilization of healthcare systems among patients with alcohol use disorder were identified, including enhancing identification and engagement strategies, offering post-discharge support, addressing specific risk factors, and customizing interventions to cater to diverse patient populations.

FoR codes (2008)

0806 INFORMATION SYSTEMS, 0801 ARTIFICIAL INTELLIGENCE AND IMAGE PROCESSING

Download

Share

COinS
 

Unless otherwise indicated, the views expressed in this thesis are those of the author and do not necessarily represent the views of the University of Wollongong.