EHR Phenotyping Methods for Measuring Treatment Adherence Among People Living With HIV in All of Us: Towards Disparities and Inequalities in HIV Care Continuum

Abstract

HIV treatment adherence is among the most important determinants of HIV outcomes. However, only 50% of people living with HIV in the US were retained in care. Measuring HIV treatment adherence in the clinical settings is feasible but when it comes to the growing number of multi-site Electronic Health Records (EHR), there has been a dearth of research for adequate informatics methods to handle EHR. We sought to address this gap by developing a cluster of metrics for measuring HIV treatment adherence via EHR phenotyping methods. Our methods were developed and tested in the All of Us research program. We also performed preliminary analyses to explore disparities in HIV treatment adherence and demographic factors contributing to poor adherence. This study paves the way for systematic data mining and analyses for the HIV care continuum, disparities, and inequality research on All of Us and other EHR normalized with the OMOP Common Data Model.

Introduction

Advances in antiretroviral therapy (ART), including greater effectiveness and access, have significantly extended the lifespan and improved the quality of life for people living with HIV (PLWH) ¹. As a result, treatment adherence – encompassing linkage to care, retention in care, and adherence to ART – has emerged as the most crucial determinant of treatment outcomes. According to the Centers for Disease Control and Prevention (CDC), among 1.2 million PLWH in the United States at the end of 2019, an estimated 87% were diagnosed, 66% have been initiated with HIV medical care, yet only 50% were retained in care ². CDC defines retention rate as the percentage of PLWH with diagnosis who had two or more CD4 or viral load tests, performed at least three months apart. Disparities of the HIV medical care retention rate persist and are significantly associated with sex, race/ethnicity, HIV infection risk factors, and geographical locations ³.

Patient adherence in HIV care is significantly linked with treatment efficacy, viral suppression and transmission, and disease progression ⁴. Presently, most HIV treatment guidelines recommend routine assessment of adherence to both ART and clinic appointments ⁵. Several methods have been proposed to measure ART adherence, from non-pharmacological measures such as self-reported adherence questionnaires, pill count, electronic monitoring devices, etc., to pharmacological measures involving biological-level measurements of antiretroviral drugs or their metabolites. However, no gold standard has emerged ^6–8.

Existing methods have several limitations – including the subjectivity of non-pharmacological measures, the cost and inconvenience of drug-level testing, and the high potential for inaccuracies of all approaches ^8,9. This incompleteness is evident in their inability to reliably detect viral non-suppression, hindering accurate assessment of patient adherence ¹⁰. Self-reported adherence questionnaire is an indirect measurement, and the accuracy suffers from subjectivity and recall errors which are significantly influenced by the social desirability and the patients’ memory biases ^11,12. The pill count method is another indirect method that has been commonly used but could be biased by patients’ manipulation since it does not confirm ingestion ¹³. Importantly, even when the data are accurate, self-report questionnaires and pill counts typically do not contain time-stamped data that could be useful for improving data management, recognizing patients’ behavior patterns, and supporting evidence-based interventions ¹⁴. Electronic monitoring allows health professionals to assess adherence longitudinally, but the high cost and the concern about user-friendliness of the devices make it difficult to implement ¹³. Pharmacological measures can generate accurate data on treatment adherence if the samples are extracted, stored, and transported securely and correctly, but they are also expensive because the laboratory infrastructure is required and the lab tests can be costly ¹⁵. The inability of the existing methods to accurately measure adherence to HIV treatment, especially to antiretroviral therapy (ART), remains a critical barrier to translating treatment guidelines into effective clinical practice for HIV patients.

In this light, the proliferation of large-scale Electronic Health Records (EHR) offers a promising opportunity to advance adherence measurement in HIV treatment ^16–19. By leveraging these routinely collected datasets, researchers and clinicians can actively employ multiple approaches simultaneously to measure adherence, potentially complementing the shortcomings of individual methods and leading to a deeper understanding of patient behavior.

Measuring treatment adherence through EHR is complex ²⁰. Recent studies – including but not limited to HIV – have suggested several existing challenges. For example, EHR present unique characteristics including duplicated records, missing values, and low granularity of semantic meanings, which may adversely affect the measuring of treatment adherence. Duplicated prescriptions corresponding with a single dispensation record could cause downward bias. Missing values could be imputed but a minor imputation bias can easily amplify the bias of adherence being measured. Because daily doses are often in the low single digits, distribution-based imputation methods can minimize the imputed doses. With respect to the granularity of semantic meanings, truncation of medical nomenclature is prevalent in EHR due to inappropriate data entry ²¹, and the process of harmonizing EHR with common data models. In addition to these challenges as a result of EHR data quality and interpretability, treatment adherence measurement presents a few unique challenges. Measurement of metabolite concentration is among the most accurate direct measures of medication adherence, but such laboratory tests are not widely captured by EHR. Measurement time window would critically influent accurate measurement of treatment adherence, such as proper operational definitions of start and end time of prescription and dispensation.

In this paper, we developed a cluster of measurements to approximate HIV treatment adherence. Our measurements are built based on phenotyping methods within the Observational Medical Outcomes Partnership (OMOP) Common Data Model (CDM). We utilized EHR from the All of Us research program as the testbed for our phenotyping methods for its robust nationwide sampling and representativeness of minorities and underserved populations, which is highly responsive to addressing the disparities in HIV treatment adherence.

Methods

Data Source

The All of Us research program has collected various types of data such as EHR, survey responses, physical measurements, wearables data, and genomic data from biospecimens from over 340 recruitment sites since 2018. As of February 2024, the program has recruited 530,920 participants who have completed the initial steps for data sharing. Of note, more than 80% of these participants are from under-represented groups in biomedical research. This study has access to the Controlled Tier data from All of Us (individual-level EHR, wearables, and survey data). Because data from All of Us are normalized with the OMOP CDM, the phenotypes we developed for measuring HIV treatment adherence were based on OMOP CDM.

On All of Us, we have previously established an EHR cohort for PLWH, described elsewhere ²⁴. See Figure 1 for the diagram of the cohort.

Figure 1.

Phenotyping diagram for establishing a cohort of PLWH in All of Us.

Clinical Measurements for HIV Treatment Adherence

We describe eight metrics that providers use to measure the states of HIV treatment adherence in the clinical setting, followed by phenotyping strategies to such measurements using EHR data. These eight metrics represent critical gaps in the HIV care cascade as described in the Joint United Nations Programme on HIV/AIDS (UNAIDS) recommended treatment initiation and the five primary stages adopted by the US CDC (HIV testing and diagnosis, linkage to care, retention in care, treatment and viral suppression) ²⁵. Figure 2 illustrates the phenotyping diagram. We adopted phenotyping methods for OMOP CDM to address missingness for all metrics development. For each metric, we assumed that class imbalance may exist (i.e., more positive examples over negative examples) and noted that our metrics/algorithms are not sensitive to such imbalance.

Figure 2.

Phenotyping diagram for the eight HIV treatment metrics.

Metric 1: Diagnosed HIV, but never initiated with ART medication. Individuals diagnosed with HIV but have not started ART were defined as those without any records of ART initiation in the EHR within the study time frame following the HIV diagnosis. Enrolment in care is commonly evaluated as the percentage of individuals who have been diagnosed with HIV that enrolled in care. For this metric, we included late enrollment to enrolment within the available time window of the data available on All of Us.

Metric 2: Retention to care. Retention to care was defined as attending at least two HIV-related visits within any twelve months, with a maximum six-month gap permitted between any two consecutive visits. HIV care visits consisted of documented records of CD4 cell counts, viral load measurements, prescribed visits, etc. within the corresponding timeframe.

Metric 3: Medication possession ratio (MPR). For this study, adherence to ART will be measured by the Medication Possession Ratio, which is defined as

$$MPR=\left(\frac{Sumofdays\text{' }suppl\text{y }forallfillsinperiod}{Numberofdaysinperiod}\right)\times 100\%$$

MPR is categorized by two levels: >90%, and <90%, where >90% is considered adequate adherence. Medication Possession Ratio (MPR) was calculated for three timeframes: 6 months, 1 year, and 2 years after the ART initiation. For each timeframe, the numerator was the sum of the days’ supply of a specific drug dispensed to a patient within that period. Individual drug MPRs were first calculated, followed by a weighted MPR. MPR calculations included only those patients with at least one record of drug exposure occurring after the end of the chosen timeframe. MPR values exceeding 100% were capped at 100%.

Metric 4: Virological Failure. Virologic failure was defined as the failure to achieve virologic suppression ²². Failure to achieve virologic suppression was defined as two consecutive HIV RNA viral load levels of ≥200 copies/mL following 36 weeks of ART commencement.

Metric 5: Late entry to care (late ART initiation) based on CD4 counts. Late entry to HIV care is defined as either having a CD4 count below 200 cells/mm³ (HIV stage 3) or a T4 helper cells to T8 suppressor cells ratio in the blood less than 0.5 before initiating ART.

Metric 6: Plasma drug (ART) concentration. Assessment of plasma ART concentration and adherence involves evaluating the levels of ART drugs and their metabolites in the bloodstream and determining how consistently a patient follows their prescribed medication regimen. (we have no data for this, so defining specifically is not possible)

Metric 7: Dropped out of care or not in care. Patients who answered ‘no’ to either of the two survey questions were defined as dropped out of care. The questions were:

1. “Are you still seeing a doctor or health care provider for HIV/AIDS?” (Form: Personal and Family Health History; concept_name: infectiousdiseases_hivaidscurrently; Choices, Calculations, OR Slider Labels: HIVAIDSCurrently_Yes, Yes | HIVAIDSCurrently_No, No)

2. “Are you currently prescribed medications and/or receiving treatment for HIV/AIDS?” (Form: Personal and Family Health History; concept_name: infectiousdiseases_rxmedsforhivaids; Choices, Calculations, OR Slider Labels: RxMedsforHIVAIDS_Yes, Yes | RxMedsforHIVAIDS_No, No)

Metric 8: Nonadherence to HIV treatment (care). Patients were classified as non-adherent or non-compliant with HIV treatment if the healthcare provider’s observations during patient interactions, as documented in the observation domain of the EHR, indicated such.

Statistical Analyses

We performed preliminary analyses to explore demographic factors contributing to poor HIV treatment adherence. We employed logistic and multinomial regression models, wherever appropriate, to explore these differences. Model outcomes are the metrics and all the metrics are binomial. Exposures are gender (male, female), age (18-49, 50-64, >64), race (Black, Which, Asian, Hawaiian or Other Pacific Islander), and ethnicity (Hispanic/Latino, not Hispanic/Latino). We dropped records with missing or unknown characteristics.

Results

The resulting OMOP concept IDs and branching logic are made available either from our project on All of Us or upon request. Due to the limited space, these OMOP concept IDs were not appended to the article.

To acquire Metric 1, we identified 5290 OMOP concepts from OMOP Domain Drug to exclude PLWH without receiving any ART medications. To acquire Metric 2, we identified 126 concepts from Domain Condition, 5290 concepts from Domain Drug, and 320 concepts from Domain Measurement to identify PLWH who do not have any HIV-related records, such as prescriptions, diagnoses, clinical findings, or laboratory tests, for a period exceeding 6 months between two adjacent HIV-related visits. To acquire Metric 3, we used the same set of concepts from Domain Drug to calculate the MPR. To acquire Metric 4, we identified 18 concepts from Domain Measurement to identify laboratory tests for HIV RNA viral load. To acquire Metric 5, we identified 5 concepts from Domain Measurement to identify laboratory tests for CD4 counts to identify late entry to ART. To acquire Metric 7, we identified two concepts from Domain Observation for clinical notes suggestive of dropping out of care. To acquire Metric 8, we identified 9 concepts from Domain Observation for records of nonadherence to HIV care.

Among the eight metrics, we found no existing data in All of Us for Metric 6. Metric 7 was not reported as well because the algorithm failed to converge due to some covariates being completely confounded. Table 1 is a summary of patients retrieved by every metric.

Table 1.

Summary of patients by Metrics 1-5, and 8.

Metric	Outcome			Total
	0	1	NA
Metric 1	6588(77.5%)	1916(22.5%)	0(0%)	8504
Metric 2	3091(46.9%)	3497(53.1%)	0(0%)	6588
Metric 3 (MPR1)	6316(95.9%)	272(4.1%)	0(0%)	6588
Metric 3 (MPR2)	6374(96.8%)	214(3.2%)	0(0%)	6588
Metric 3 (MPR3)	6446(97.8%)	142(2.2%)	0(0%)	6588
Metric 4	1650(25.0%)	653(9.9%)	4285(65.0%)	6588
Metric 5	747(11.3%)	281(4.3%)	5560(84.4%)	6588
Metric 8	5836(88.6%)	752(11.4%)	0(0%)	6588

Note: For Metric 1, the denominator for calculating the percentage is all PLWH; for Metrics 2, 3, 4, 5, and 8, the denominator is PLWH who have initiated ART.

Metric 1: 1-diagnosed HIV, but never initiated with ART medication; Metric 2: 1-ART initiated, then dropped out of care; Metric 3: 1-adequate adherence; Metric 4: 1- virologic failure; Metric 5: 1-late entry to care or late ART initiation based on CD4 count; Metric 8: 1-nonadherence to HIV treatment (care).

Our preliminary analyses of disparities revealed several notable findings. First, individuals identifying as Black or African American with HIV are significantly more likely to have never initiated ART medication (OR: 3.13, 95% CI: [1.71, 5.73]). Among individuals who initiated ART, several demographic factors were associated with an increased likelihood of dropout from care, including being male (OR: 2.03, 95% CI: [1.80, 2.29]), aged 50-64 (OR: 1.55, 95% CI: [1.35, 1.77]), and identifying as Black or African American (OR: 4.62, 95% CI: [2.98, 7.14]), or White (OR: 1.57, 95% CI: [1.02, 2.43]).

Furthermore, across all three timeframes (6 months, 1 year, and 1.5 years after the ART initiation), males were found to be more likely to exhibit adequate adherence to medication. Conversely, individuals aged 50-64 (OR: 0.75, 95% CI: [0.59, 0.95]) or over 64 (OR: 0.59, 95% CI: [0.41, 0.83]) were less likely to experience virologic failure. Additionally, those over 64 (OR: 0.58, 95% CI: [0.34, 0.99]) were less likely to have late entry to care.

Finally, individuals aged 50-64 (OR: 0.73, 95% CI: [0.61, 0.87]) or over 64 (OR: 0.42, 95% CI: [0.33, 0.55]) were less likely to exhibit nonadherence to HIV treatment. Conversely, Black or African American individuals (OR: 5.80, 95% CI: [2.34, 14.37]) were significantly more likely to experience nonadherence to HIV treatment. These findings underscore the importance of addressing the poor HIV treatment adherence of Black or African American individuals. Table 2 summarizes the statistical results.

Table 2.

Odds ratios for poor HIV treatment adherence by demographics.

	M1		M2		M3: MPR1		M3: MPR2		M3: MPR3		M4		M5		M8
	OR	95% CI	OR	95% CI	OR	95% CI	OR	95% CI	OR	95% CI	OR	95% CI	OR	95% CI	OR	95% CI
Gender	1.04	[0.92, 1.18]	2.02	[1.79, 2.28]	4.30	[2.89, 6.39]	4.99	[3.12, 7.99]	6.14	[3.29, 11.45]	1.10	[0.87, 1.40]	1.19	[0.83, 1.70]	1.15	[0.97, 1.36]
Male
Female*
Age
50-64	1.09	[0.94, 1.25]	1.54	[1.34, 1.77]	1.05	[0.76, 1.46]	1.21	[0.83, 1.76]	2.21	[1.31, 3.75]	0.74	[0.58, 0.95]	1.03	[0.72, 1.47]	0.73	[0.61, 0.87]
>64	0.95	[0.80, 1.12]	1.03	[0.88, 1.21]	0.93	[0.64, 1.36]	1.09	[0.71, 1.66]	2.55	[1.46, 4.44]	0.58	[0.41, 0.82]	0.57	[0.33, 0.98]	0.42	[0.32, 0.54]
18-49*
Race
Black	3.12	[1.70, 5.72]	4.61	[2.98, 7.13]	0.69	[0.29, 1.64]	0.47	[0.19, 1.14]	0.36	[0.12, 1.06]	1.25	[0.38, 4.08]	1.17	[0.11, 11.65]	5.79	[2.33, 14.37]
NHOPI	2.52	[0.59, 10.70]	1.68	[0.41, 6.87]	0.00	[0.00, 0.00]	0.00	[0.00, Inf]	0.00	[0.00, Inf]	0.00	[0.00, Inf]	-	-	2.80	[0.28, 27.36]
White	1.59	[0.87, 2.93]	1.57	[1.01, 2.42]	0.90	[0.38, 2.13]	0.71	[0.30, 1.70]	0.57	[0.20, 1.66]	0.41	[0.12, 1.37]	0.89	[0.08, 8.96]	1.99	[0.79, 4.96]
Asian*
Ethnicity
Not
Hispanic/Latino	0.66	[0.46, 0.96]	0.79	[0.53, 1.15]	1.54	[0.56, 4.26]	0.95	[0.38, 2.39]	0.53	[0.21, 1.36]	-	[0.40, 1.73]	1.48	[0.48, 4.56]	0.78	[0.47, 1.29]
Hispanic/Latino*

Underscore: significant

^*Reference

Discussion and Conclusions

Principal Findings

Among the eight metrics we developed to measure HIV treatment adherence, six of them were successfully implemented to retrieve adequate patient information. Metric 6 (plasma drug concentration) is theoretically valid but did not function after implementation because All of Us currently does not record laboratory testing data for drug concentration. Metric 7 (dropped out of care) was not reported because the algorithm failed to converge. This failure could be attributed to the fact that medication possession does not measure ingestion, or the data structure, where the outcome was binary which only had two levels. Regardless, the eight metrics independently measured different statuses for PLWH during HIV care, which has demonstrated the feasibility of approximating HIV treatment adherence using large-scale and OMOP CDM-normalized EHR data.

All of Us is unique as it contains a considerable number of individuals from historically underrepresented groups in biomedical research. This advantage is important for studying the HIV medical care continuum because various sources of disparities have been well reported among PLWH who do not adhere to HIV medical care. The present study, though in its pilot state, has discovered racial disparities which is consistent with findings using other datasets and methods ³. For example, African Americans consistently show poor adherence to HIV medical care compared to other race/ethnicity groups. These findings support the validity of the EHR phenotyping methods we developed for approximating HIV treatment adherence, as well as using data from All of Us to study health disparities, underrepresented populations, and their contributing factors.

Clinical Implementation

Accurate and timely measure of adherence to HIV treatment indicates the effectiveness of medical treatment and patient responses to treatment (e.g., viral suppression) and thus can provide appropriate feedback for evaluating and/or adjusting therapeutics and developing proactive strategies to reduce the risk of progression and transmission of HIV. Given the proliferation of EHR in the US, adequate methods are desperately needed for EHR data. This study successfully demonstrates a feasible approach to approximate HIV treatment adherence using existing protocols and experience from clinic-based measurements. Our approach holds the promise of developing rapid screening tools on large populations and datasets, which is a cost-effective solution, compared to clinic-based measurements, for a cross-state implementation. Our approach also promotes personalized healthcare in HIV care in the sense that contributing factors for poor adherence could be identified and tailored for each patient based on one’s health records and social determinants of health (SDoH).

With accurate methods to measure HIV treatment adherence, providers can use cost-effective triage to deliver target adherence support interventions for at-risk individuals. This advantage is particularly valuable for resource-limited clinics and community-based settings, which typically align with geographic and racial disparities. Although this is theoretically promising, we acknowledge that infrastructure and other resources are often limited for regions and communities with poor adherence rates for implementing such an EHR phenotyping tool ²³. Therefore, a cost-effective solution for implementing health service delivery would be of value for further discussions. Regardless, our approach demonstrates the feasibility of better allocating and optimizing healthcare resources for resource-limited settings and communities.

Contributions to Informatics

There has been a desert of research in phenotyping HIV treatment adherence based on EHR. This study made unique contributions as it offers a phenotyping approach based on OMOP CDM. Over recent years, there has been an increasing number of data repositories for multi-site multi-format EHR. This trend responds closely to the Feasible, Accessible, Interoperable, Reusable (FAIR) principles and, more importantly, suggests the increasing demands for clinical information extraction and data mining methods tailored for OMOP CDM-normalized EHR.

All of Us is unique as the data collection was designed to include and promote minorities and underrepresented populations in biomedical research. In response to it, our study made use of All of Us data to explore disparities and their contributing factors among PLWH with poor HIV treatment adherence. Successful experience from our effort would pave the way for informatics researchers to contribute to inequality and disparities research in the fields of data mining, clinical decision support, and artificial intelligence in its generic definition.

Limitations and Future Directions

This study presents several limitations. First, the methods we tested in this study only addressed some of the challenges discussed in the scientific communities for assessing medication/treatment adherence in EHR ²⁰. However, recognizing some of the challenges inherent in treatment adherence assessment, future studies will aim to revalidate the reported methods on different datasets and patient cohorts. Second, methods for Metrics 3, 7, and 8 all present some degree of subjectivity, making them prone to underestimating or overestimating true adherence. The algorithm converging issue could be mitigated with the accumulation of the data. Third, we noted data missingness and class imbalance issues, which could have introduced biases to the results. Future studies will include validation with clinicians to identify and reduce possible biases.

Conclusions

This study developed and tested phenotyping methods to assess HIV treatment adherence using EHR from All of Us. After the methods were implemented on All of Us, we performed preliminary analyses to explore health disparities and contributing factors for PLWH with poor adherence which resulted in consistent findings from existing studies, which further validated the reported methods and confirmed the potential to use our methods for systematic health disparities and inequalities research for HIV care continuum on All of Us.