This is a retrospective data linkage study using data from Victoria, Australia. Patient-level data from primary health networks (PHNs), extracted from the Population Level Analysis and Reporting (POLAR) platform provided by Outcome Health, and admission and ED data sets from 3 metropolitan hospital networks will be linked. These data were chosen because they provide patient-level demographic, diagnosis, medication, and hospital attendance information, and they can be accessed in a timely manner to address the research questions.

The state of Victoria is the second-most populous state in Australia, with 26% of the nation’s population residing there. The current POLAR database covers approximately 30% of general practices across southeastern Victoria [26]. A total of 3 hospital networks in Victoria (Monash Health, Eastern Health, and Peninsula Health) that align with the Eastern and Southeast PHN catchments were chosen to examine opioid harms due to the high number of POLAR-participating practices in these 3 health regions. These hospital networks are located in the greater Melbourne metropolitan area and have catchments of nearly 2.6 million people, covering 40% of Victoria’s population, with their EDs managing around 500,000 presentations per year. The areas represented by these data are geographically, demographically, and socioeconomically diverse [27].

The study population will include people 18 years or older who had at least 1 attendance at general practice sites participating in POLAR data provision during the study period (2017-2022).

We will include all opioids prescribed for pain (ie, codeine, dextropropoxyphene, fentanyl, hydromorphone, morphine, oxycodone, pethidine, tapentadol, tramadol, buprenorphine [transdermal products only], and methadone [tablet formulations only]). Formulations that are only used as cough suppressants (eg, dihydrocodeine) or for opioid replacement therapy (eg, high-dose sublingual buprenorphine-naloxone and methadone liquid) will be excluded (Table S1 in Multimedia Appendix 1). The use of nonopioid medicines (eg, gabapentinoids and benzodiazepines) will be considered as covariates.

The main outcomes of interest are ED presentations and hospital admissions attributed to substance use (both opioid and nonopioid) or mental ill-health (eg, suicide, self-harm, anxiety, and depression). We focus on these outcomes as opioid- and nonopioid-related substance use, mental health crises, and suicide have all been demonstrated to be key outcomes related to changes in opioid prescribing [18-20].

Primary and secondary diagnoses related to ED presentations and hospital attendance are coded according to the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision, Australian Modification (ICD-10-AM). Opioid-related presentations will be identified where the primary or additional diagnoses include (1) F11.0-F11.9 (opioid-related disorders) and (2) T40.0-T40.4, T40.6 (poisoning by, adverse effects of, and underdosing of narcotics and psychodysleptics).

Non–opioid-related substance use will include substances such as alcohol, heroin, and cannabis. The ICD-10-AM codes F10, F12-F19 will be used to identify cases of unintentional poisoning. Cases of suicide attempts and self-harm that satisfy the following criteria will be included: (1) a principal diagnosis in the ICD-10-AM range S00-T75, T79 (injury, poisoning, and certain other consequences of external causes); (2) the first reported external cause code in the record in the ICD-10-AM range X60-X84, Y87.0 (external causes of morbidity); and (3) suicide-related behaviors (R45.81 [suicidal ideation]). A full list of ICD-10-AM diagnosis codes used to define mental ill-health–related hospital separations by the Australian Institute of Health and Welfare is provided in Table S2 in Multimedia Appendix 1.

Analyses will adjust for key covariates associated with opioid-related outcomes such as age, sex, physical and mental health comorbidity, and substance use (including alcohol and nicotine use; Textbox 1) [29-34].

Before the analysis, data will be prepared, and a quality assessment undertaken. Preparation involves (1) archiving a copy of primary data on the secure environment in original format and creating a copy for use in further analysis; (2) converting data from the received format (often .csv or .xls format) to formats used in statistical packages such as STATA (StataCorp), R (R Foundation for Statistical Computing), and SPSS (IBM Corp); (3) checking naming conventions for indicators and fields to ensure that these are consistent with data dictionaries; and (4) developing a study data dictionary and creating metadata.

Data validity and quality assurance will be conducted on the linked database. This will involve examining the missingness of individual data fields, inconsistency between data sets, inaccurate and unreasonable data values (eg, prescription quantity >200), identification and removal of duplicate records, and merging multiple episodes that occur within a single presentation (eg, where a patient may move between wards within a hospital).

Data exploration and visualization will be undertaken to enable the research team to examine the properties of the data sets, such as dispersion or distribution of values, correlation between variables, and time trends. Descriptive statistics (counts, means, medians, etc) will be calculated based on demographic data, diagnosed conditions, and medication use. Preliminary methods for defining key variables will be developed. Next, we will check the internal consistency of the linked data. This will help quantify the success of data linkage across POLAR, hospital, and ED data. We will define and examine the consistency of a range of indicators, such as patients’ age and gender.

To ensure comprehensive and accurate demographic information, a cross-check will be conducted to address missing data in both hospital records and the POLAR data set. When key demographic information is absent in hospital data, the gaps will be filled by leveraging the variable field from the POLAR data. Similarly, in instances where the POLAR data lack essential demographic details, an imputation process will be used based on the available hospital data. In cases where a discrepancy in demographic information is identified between the hospital data and the POLAR data set, priority is given to the values recorded in the hospital records. The demographic details present in the hospital data are considered more reliable and representative of the specific patient population being studied. In addition, where key prescription-related information in the POLAR data is missing (eg, quantity prescribed or dose), these will be imputed based on standard pack sizes and the most common formulations used. Missing data for all variables will be checked and quantified. Multiple imputation methods will be considered where there are more than 5% missing data within variables.

The process of collection, transfer, and storage of general practice data on the POLAR platform was granted ethics approval in August 2017 (NREEC RACGP Protocol ID: 17-008). Consent for the use of deidentified patient data for research extracted and processed on the POLAR platform occurs as follows: practices that contribute data to the POLAR system enter into an agreement with their local PHN. The agreement articulates the responsibility of practices to inform their patients that data are being collected and of their potential use for research. As an additional layer of assurance, practices contributing data to the POLAR platform adhere to the RACGP Accreditation Standards, which stipulate requirements for responsible management of health data. Patients are informed at the point of collection and are presented with the opportunity to opt out. POLAR data are deidentified at the point of collection, according to Outcome Health’s Data Deidentification Decision-Making Guide, developed with guidance from the Office of the Australian Information Commissioner. Hospital data are proposed to be used without the consent of individuals. A multisite ethics application was approved by Monash Health as the lead Human Research Ethics Committee (ID 76744), with site-specific applications approved by Monash Health (RES-22-0000-026A), Peninsula Health (SSA/76744/PH-2022), and Eastern Health (S22-032-76744). A waiver of consent was granted by the respective Human Research Ethics Committees for linkages between the POLAR data and hospital data. Privacy-preserving encrypted linkage software was used to link the data, and deidentified content data were submitted into a secure e-research environment by each of the data custodians for analysis by the research team. Ethics approval has been obtained from the Monash University Human Research Ethics Committee (ID 76744), Monash Health (RES-22-0000-026A), Peninsula Health (SSA/76744/PH-2022), and Eastern Health (S22-032-76744).

Policies restricting access to medications such as opioids are commonly implemented, with limited evidence of their effect on reducing opioid-related or other harms. This study will add to the limited evidence base to help understand the impact of these policies in Australia, including whether intended or unintended outcomes are occurring as a result. By closely monitoring and analyzing ED presentations and hospital admissions as markers, policy makers can assess the impact of interventions, identify areas for improvement, and drive evidence-based policy changes that effectively address the complex challenges associated with opioid use and promote healthier communities.

This study will link primary care and hospital data to investigate opioid exposure and adverse outcomes. POLAR data contain rich patient-level data, including diagnoses and comorbidities, that are often missing from existing administrative data such as Medicare Benefits Schedule and PBS data in Australia. The linked data set will also capture both publicly subsidized and nonsubsidized prescriptions, making it distinct from the more commonly used PBS (government subsidy) database. Moreover, the linkage key is consistent throughout POLAR participating practices, which will allow for a more comprehensive capture of covariates, particularly coprescribing and hospital readmission.

There are several limitations to note. First, the use of prescribing data does not enable confirmation that prescribed medicines were dispensed or taken by the patient. Second, as both primary care and hospital data are drawn from 3 specific geographical areas, there is potential for selection bias. This concern is lessened by the inclusion of 3 PHNs whose average opioid-prescribing patterns are consistent with national opioid-prescribing patterns [45]. Further, there is established geographic and sociodemographic diversity across these health areas that limits concerns about generalizability [46]. Third, patients who visited non-POLAR GP practices and were admitted to hospitals other than the selected ones are not captured in this study; however, it is noted that the 3 large hospital networks serve the majority of the population in each of the catchment areas. Fourth, a list of ICD-10 codes will be used to identify study outcomes. However, there are limits to the completeness of the coding. For instance, suicide in EDs is commonly recorded using a single diagnostic code [47], and opioid-poisoning presentations can be coded under the more generic code T50: “Poisoning by diuretics and other and unspecified drugs, medicaments and biological substances” [48]. Therefore, our opioid-related poisoning and suicide and self-harm estimates are likely to be an underestimate of the true prevalence. In addition, deaths that occur before reaching the hospital and harm that occur outside the hospital setting (eg, ambulance attendances) would not be captured unless the patient was later transferred to a hospital.