In the United States (U.S.), the goals of the National Human Immunodeficiency Virus and Acquired Immunodeficiency Syndrome (HIV/AIDS) Strategy 2020 are to diagnose 90% of people with HIV (PWH), link 90% of diagnosed people to care and achieve effective viral suppression for 90% of diagnosed PLW.¹ Additionally, the Ending of HIV Epidemic: A Plan for America (EHE) initiative aims to reduce HIV incidence by at least 90% by 2030 through focusing on four pillars of diagnoses, treat, prevent, and respond.² The latest estimates indicated that almost two-thirds of PWH (65%) received any care, and only 56% were virally suppressed in 2018.³ In addition to timely diagnosis, it is crucial to identify PWH who are not retained in care and/or not suppressed to design health intervention strategies at the individual and population level.

The widespread adoption of electronic health record (EHR) systems and the creation of clinical research networks with extensive collections of EHR data made it possible to develop and validate computable phenotype (CP) algorithms that search and classify patients with specific phenotypes (e.g., living with HIV) using clinical documentation, billing, laboratory data, and other EHR-related data sources. EHR-based CP algorithms to detect HIV cases can help researchers efficiently identify a large virtual cohort of PWH with longitudinal trajectories of engagement through routine care both before and after their HIV diagnosis. Further, secondary analyses of data from the identified EHR cohort can provide real-world evidence⁴ to assess gaps in the HIV care continuum, comorbidity burdens, patterns in utilization of services, and quality of care.

HIV/AIDS case detection algorithms have been developed in previous literature using Medicare and/or Medicaid claims data^5–9 or Veterans Health Administration (VHA) data.^10,11 These algorithms heavily relied on diagnostic codings, such as the use of International Classification of Disease (ICD) codes. However, the diagnosis coding is designed for administrative/billing purposes instead of clinical or research use.¹² A single record of a disease diagnostic code may not accurately reflect the actual disease that the patient had.¹³ To overcome this shortcoming, some algorithms classify HIV cases only when patients have multiple claims/encounters with a corresponding HIV diagnosis (e.g., at least one inpatient or two outpatient claims).^5,9,11

EHR-based algorithms for HIV/AIDS have improved upon previous claims-based algorithms mainly by addressing the inherent methodologic limitations of claims data via incorporating the additional rich information available in EHRs. The algorithm developed by Felsen et al¹⁴ was based on the EHR from one medical center in New York City, and it incorporated information on HIV-related ICD (ver. 9) diagnostic codes, encounters, and laboratories (including western blot, antibody test, viral load, and CD4 count). Paul et al¹⁵ designed and validated two algorithms to identify HIV patients using the EHR from the Duke University Health System. One of their algorithms was laboratory-based and detected an HIV case if the patient met one of the following two criteria: (1) having a positive HIV antibody test confirmed with a Western blot, nucleic acid test, or another positive HIV antigen test, (2) having been prescribed HIV-specific medications. Their second algorithm was ICD-9-based and classified patients as HIV cases when any HIV-related ICD-9 code was accompanied by confirmatory medication or laboratory results consistent with HIV. This second algorithm was designed to better deal with the real-world data with incomplete medical records while still relying heavily on laboratory testing results and prescription history for case detection. However, it should be acknowledged that both algorithms still possess limitations. Felsen et al’s algorithm did not incorporate HIV medication information, while the algorithms developed by Paul et al did not consider encounter patterns. Moreover, both studies were based on ICD-9 codes and developed based on EHR from one health care system, and therefore are not very generalizable.

Florida is the topmost contributor to the United States’ national cases of HIV, with an estimated total of 116,689 PWH in 2019 and more than 4,500 new HIV infections every year.^3,16 Leveraging local EHR data can help better identify gaps in the HIV care continuum to develop the targeted intervention. As one of the nine clinical data research networks (CDRNs) of the National Patient-Centered Clinical Research Network (PCORnet), the OneFlorida Clinical Research Consortium operates a statewide clinical data warehouse. It partners with 11 clinical health care systems across Florida, and all EHRs were mapped to the PCORnet common data model.¹⁷ Since patient-generated EHR data are generally customized and inconsistent across health care systems,^18,19 calibrating the EHR-based algorithm using OneFlorida EHR data are essential to detect and characterize the HIV cohort in Florida adequately. This study takes a comprehensive approach to construct and validate an algorithm that will improve previous published EHR-based phenotype algorithms by leveraging data from more domains of the PCORnet common data model. Our algorithm allows more flexible inclusion and exclusion criteria for patients with missing or incomplete laboratory test results and medication prescribing history. Additionally, this algorithm is compatible with newly adopted ICD-10 and Systemized Nomenclature of Medicine (SNOMED) diagnostic codes and can be replicable to other PCORnet Network Partners.

Around 0.6 billion encounters, 1.2 billion diagnoses, 0.3 billion prescriptions, and 0.5 billion claims were documented in OneFlorida among the 15+ million patients. Approximately 5 million patients had at least one care visit to OneFlorida partner each year. The majority (54.8%) of the sample were females, 27.6% were Hispanic/Latinx, and 22.0% were Black African American. The mean age was 30 years.

In this study, we developed and validated a CP algorithm for identifying patients with confirmed HIV diagnoses in EHRs, leveraging information from multiple health care providers and EHR domains in the PCORnet common data model. Our study extended the previously published algorithms in several important ways. First, we expanded the diagnostic codes to include ICD-10-CM and SNOMED codes in addition to ICD-9-CM-codes. Second, our algorithm allowed more flexible inclusion and exclusion criteria by leveraging information from multiple EHR domains and less reliance on complete screening test laboratory results relative to the algorithm designed by Paul et al.¹⁵ Third, in addition to considering HIV screen tests, our phenotype considered laboratory procedures commonly prescribed for patients with known HIV status (such as HIV genotyping and CD4 tests). Lastly, our algorithm was developed using EHRs that follow the PCORnet common data model and can be easily replicable to other PCORnet Network health systems with local recalibration.

Our final algorithm achieved higher sensitivity (98.5%) compared with previous algorithms (77–78% in Paul et al,¹⁵ 83% in Goetz et al²³) with comparable specificity (97.6% in our algorithm, vs. 99–100% in Paul et al,¹⁵ 86% in Goetz et al²³). The main reason for false negatives was because HIV was documented as past medical history in clinical notes, but no HIV diagnostic codes were used in the structured EHR. Misclassification of the false positives was primarily due to incomplete HIV laboratories and prescribing records. The use of two HIV-related encounter criteria may incorrectly classify a small proportion of people without an HIV diagnosis. For example, a patient may have had two HIV-related ICD codes documented in EHR for routine screening of HIV.

The high performance of EHR-based CPs offers rapid identification of patients from EHRs of different health care systems, which help the development of multicenter patient cohorts for research, clinical care, and public health initiatives.¹⁵ The HIV cases (61,313 patients) identified by our algorithm reflect the number of PWH in Florida who received some care in OneFlorida partners. In 2019, around 80% of 116,689 PWH received an HIV care,¹⁶ and OneFlorida partners provided care for 15 million patients in past years, approximately 70% of the 2019 population.²⁸ The age and race composition in our HIV cohort were similar to the characteristics of PWH estimated by the State’s Health Department, yet males (58.2% in our cohort vs. 72.7% in the state’s estimate), and Hispanics (16.6% in our cohort vs. 27.4% in the state’s estimate) were less represented in our cohort.²⁹ Patients included in the OneFlorida dataset generally have access to a major health system network in Florida. PWH who are out of care or seek care outside of the OneFlorida network, like exclusively from county health departments, may not be representative of our cohort. Females are generally more frequent users of the health care system than males^30–33; this may partially explain the overrepresentation of females relative to the state’s estimate. Moreover, research has documented lower access to care among the Hispanic/Latinx population due to lack of insurance, language barriers, and low socioeconomic status.^34–36 Additionally, 9.7% of ethnicity was unknown in our HIV cohort, which may also underestimate the true proportion of Hispanic/Latinx.

Unlike previous algorithms built solely upon the EHR of one hospital system or single-payer claims data, OneFlorida combines both all-payer claims data and hospital EHRs to capture all potential encounters of the patients, which intensify the identification of a virtual cohort with any conditions. Building on the strength of the OneFlorida network, the HIV virtual cohort identified through our CP can provide real-world evidence that helps better understand the disease burden and patterns in access to care among PWH in Florida and inform targeted intervention to improve their health outcomes. Potential future research like geo-mapping of the identified patients would be possible to examine the disparity burden by county, urbanicity, or other geolocation groups. Additionally, the rich longitudinal EHR data could also be mined to assess comorbidity and mortality burdens, gaps in the HIV care continuum, patterns in utilization of services, and quality of care among HIV patients. Another major strength of our virtual HIV cohort in OneFlorida EHR is that it captures patients’ interaction with the health care system both before and after their HIV diagnosis, whereas the other patient cohorts identified from the Ryan White HIV/AIDS Program and state HIV surveillance program, like the Enhanced HIV/AIDS Reporting System, only focused on health services after their HIV diagnosis. Examining health care utilization patterns before HIV diagnosis can help researchers to identify missed opportunities for future HIV prevention efforts. Furthermore, instead of focusing solely on HIV-related outcomes, our EHR-based virtual cohort can be used to examine the full spectrum of comorbidities along with their related engagement in care, laboratory results, and treatment. PWH often experiences comorbidities. Additionally, there is an increasing recognition that HIV may not always be a patient’s top priority or chief complaint while interacting with the health care system.^37,38 Secondary data analysis of the identified virtual HIV cohort could provide real-world evidence on their comorbidities and treatment burdens and thus provide insights into future patient-centered HIV care.

Our identified virtual cohort faces some unique challenges in assessing gaps in the HIV care continuum. Many patients in the virtual cohort do not have complete HIV laboratory results. One reason could be that OneFlorida incorporates Medicaid claim, which does not contain laboratory results. These missing laboratory results might limit the data in its potential for monitoring viral suppression over time. Another challenge is that we cannot distinguish individuals who have all of their health care documented within the OneFlorida EHR versus those who do not. As a result, we may underestimate the proportion of patients who were engaged in care as some of them may be misclassified as “not engaged in care” because they seek health care both in and out of the OneFlorida partners.

Our work has some limitations. First, not all sites and partners within the OneFlorida network have complete PCORnet common model elements, which limits the generalizability of our identified HIV cohort and the performance of our CP algorithm. Another limitation involves our validation process as it was limited to one health care system. Ideally, a random validation sample would be selected from the OneFlorida Clinical Research Consortium to extract clinical notes to perform chart reviews. However, we were not able to access medical notes from each of the partner health care systems. Additionally, the validation sample was drawn based on HIV-related inclusion clusters. This may inflate the prevalence of HIV among the validation sample and lead to a slight overestimation of PPV and underestimation of negative predictive value.

In summary, the CP we developed achieved high sensitivity and specificity in identifying patients with confirmed HIV diagnosis using EHR data. Relative to the published algorithms, our method relies less on complete laboratory results and prescribing data by leveraging the multi-provider, multi-domain EHR. Our algorithm was compatible with the PCORNet Common Data Model and can be translatable to other health care systems using the same Common Data Model. The CP will enable future researchers to identify a large cohort of PWH efficiently. The identified HIV cohort has the potential to improve research and programs in the area of HIV prevention and care and other important health outcomes of PWH in Florida.