KPNC serves 4.3 million members, comprising about one-third of the population in Northern California. The membership is diverse and highly representative of the US population with access to care [20]. Membership includes enrollees from Medicaid (12%), Medicare (16%), employer-based plans, and health insurance exchanges. KPNC members have direct access to specialty care clinics, including addiction medicine and psychiatry [21].

In June 2013, KPNC implemented Alcohol as a Vital Sign, a systematic alcohol screening, brief intervention, and referral to treatment initiative, in adult primary care [19]. While the initiative is primary care-based, the EHR screening tools are available for use in outpatient medical departments. KPNC has maintained an average 87% screening rate systemwide in adult primary care. As part of the screening, patients are asked three questions about their alcohol use, including a modified version of the evidence-based National Institute on Alcohol Abuse and Alcoholism (NIAAA) single-item screening question [16] (tailored to the patient’s age and sex)—“How many times in the past three months have you had 5 or more drinks containing alcohol in a day?” (for men aged 18-65 years), or “4 or more drinks” (for all women and for men aged ≥66 years)—and two questions that are used to calculate average drinks consumed per week—“On average, how many days per week do you have an alcoholic drink?” and “On a typical drinking day, how many drinks do you have?” The EHR issues a best practice alert during a primary care visit when screening is required (ie, first visit, annually, or every six months if unhealthy alcohol use was previously reported). The medical assistant may skip these questions for a variety of reasons (eg, late appointment arrivals, forgetting), and patients may decline to respond.

Registry data are leveraged from two existing data sources: Clarity (GridApp Systems, Inc) and the Virtual Data Warehouse (VDW). Clarity is the back-end database of EHR data collected in KP HealthConnect, which we used to extract alcohol screening data. The VDW is a distributed data model developed by the Health Care Systems Research Network (HCSRN) to maintain single extract, transform, and load processes that efficiently create relational tables useful for research [22]. The VDW gathers data from various EHR-based sources, including Clarity, and legacy systems that were used before the implementation of KP HealthConnect in 2005. The VDW data has been developed over many years with standardized data definitions and formats, and rigorous quality assurance.

In collaboration with NIAAA, we defined the objective of the registry and target population and formed research aims to frame the registry’s scope. The purpose of the registry is to study the full course of alcohol problems with the flexibility to address many research questions, such as the escalation of unhealthy drinking to development of AUDs and alcohol-related health problems, and the ability to be updated with new data. The target population for the registry is adult patients diagnosed with an alcohol problem and those at risk of developing one.

We developed a protocol for building the registry (available upon request), following recommendations from the US Agency for Healthcare Research and Quality [4] and other disease registries [23,24], and received approval by the Institutional Review Board at KPNC. We benchmarked our approach to that of other disease registries at KPNC (eg, HIV [25], diabetes [26], cancer [27], opioid use [28]), to determine feasibility, data storage, and access. We surveyed the literature and involved KPNC physicians in psychiatry and addiction medicine to help select key data elements and clarify data definitions. We established a plan for leveraging available data by characterizing eligibility criteria for inclusion, defining the structure of the registry, and identifying core data elements and variables needed to address the research aims. We developed codebooks to define the scope of the registry (eg, diagnosis codebook of International Classification of Diseases, 9th Revision, Clinical Modification [ICD-9] and 10^th Revision, Clinical Modification [ICD-10] codes), which can easily be updated to extend the breadth of data that the registry captures.

Implementation of the protocol took about 10 months with 50% programmer time effort. We wrote programs using SAS software, version 9.4 of the SAS System for Unix (SAS Institute), to build the registry, which were designed to minimize user interaction and could be used again to refresh the registry data (eg, using macros and macro variables). We minimized data cleaning to allow future studies to make their own decisions regarding the use of the data. We created a data dictionary to describe the files and variables that comprise the registry. We also developed queries for quality control, such as identifying missing data and characterizing data storage requirements. Last, we created reporting tools to display trends of the registry data over time.

Since the EHR is a constantly changing data environment, refreshing the registry with new data requires programs and documentation to be updated. For example, source variables and tables may be renamed or become deprecated during upgrades of data systems. The amount of time required to refresh the registry depends on the quantity and types of changes needed (eg, adding more ICD codes versus editing SAS programs), but may take anywhere from an hour to several days. Receiving ongoing feedback of the registry as research staff use it for their projects is also critical to ensuring the registry’s validity and usefulness.

We calculated the prevalence of alcohol problems among all adult KPNC patients who had at least one encounter with the health system between June 1, 2013, and May 31, 2019. We conducted descriptive analyses to describe demographic, clinical (eg, medical and mental health conditions), and insurance characteristics of the registry cohort. We included only key variables in the current analysis to compare the registry cohort to those in other published studies. All characteristics, such as age, were based on the patient’s index date. We estimated patients’ household income and education using US Census data that has been geocoded to patients’ closest residential addresses in the year prior to and including the month of their index date. If the index date was before January 1, 2017, we used the 2010 US Census data; otherwise, we used 2017 data, since census block boundaries can change over time [31]. We used the median household income of the census block to estimate patients’ household income and categorized patients into groups used in prior epidemiologic studies of the general US population [18,32]. The education level with the highest proportion of households in the census block was used to estimate patients’ education. To identify smoking status, we used the closest screening in the year prior to and including the month of the index date. We calculated the Charlson comorbidity score, which estimates the 1-year mortality risk based on a weighted score of 17 medical conditions [33], and identified chronic medical and mental health conditions and substance use disorder diagnoses in the year prior to the index date. All analyses were conducted using SAS software version 9.4.

EHR-based registries enable observational studies of “real-world” settings (eg, comparative effectiveness research), an alternative to randomized controlled trials, which may not be feasible; however, using secondary data for research has limitations, including the omission of essential variables and potential for bias (eg, selection bias, information bias, confounding). For example, clinicians in addiction medicine and psychiatry assess AUD symptoms based on the DSM-5, but detailed data are not entered in the EHR. Instead, clinicians record ICD codes to indicate AUD diagnoses, which we use in the registry. While ICD codes for AUDs are not given lightly in other departments, they do occur, and it is not clear what guidelines are used. Therefore, a future validation study of alcohol screening results and the use of these ICD codes is warranted. We also do not have direct measures of individual socioeconomic status (eg, income, education), which are important factors associated with unhealthy alcohol use [18], or social functioning (eg, the Psychosocial Functioning Inventory [45]), an important recovery outcome [46]. Though not only an issue with secondary data analysis, missing data can create bias in a study if it is not missing completely at random; therefore, future studies utilizing the registry data should check for missingness and apply proper statistical methods to address issues as needed [47]. Reliance on accurate reporting of alcohol use and other measures is also a concern; however, it is not a unique problem of EHR data and shared by other studies that collect self-report data. While novel statistical methodologies can be applied to deal with issues of confounding [48] (eg, the counterfactual framework [49]), temporality may remain an issue. Measures of alcohol use and diagnoses are recorded in the EHR when patients seek care rather than when alcohol-related issues emerge, similar to other disease-based registries that rely on data collected during care (eg, diagnostic tests for cancer) and survey-based studies that gather data on past-year or lifetime alcohol problems without specific dates.

Sex and gender variables in the EHR can change and are not collected longitudinally, so their values in the registry reflect what was present at the time of the data extraction rather than historical values, for example, at the time of alcohol screening. We are also not certain which variables are used to determine the appropriate screening questions and risk thresholds (especially for gender minorities), which a future study could evaluate. Therefore, some alcohol screening results may have been misclassified in the registry, affecting eligibility; however, we expect this to have a minimal impact on future studies.

While we included only core data elements that were necessary to address our research aims, the registry could be extended to include other types of data, including provider information, family members of patients with alcohol problems, and medications prescribed off-label to treat AUD (eg, gabapentin [50]). Other health systems in the HCSRN with harmonized VDW data [22] may also want to create their own registry of alcohol problems, enabling the potential for multi-site studies [51-53]. The registry’s utility may also extend beyond research to clinical or administrative purposes, for example, to manage care or evaluate performance, which would require additional support from KPNC organizational stakeholders and institutional review boards to protect patient privacy and confidentiality.

We demonstrate that EHR-based data collected during routine clinical care within an integrated health care system can be leveraged to develop a registry of patients with alcohol problems that is flexible and can be easily refreshed and extended. The registry can be used to address critical public health questions related to the full spectrum and course of alcohol problems, which will complement other methods used in alcohol research. Future analyses will aim to provide insight on how to strengthen efforts in the prevention of alcohol-related disability and mortality and improve patient-centered health care delivery. We hope that other researchers and health systems interested in assembling a similar registry can take advantage of the time we invested in developing this protocol.