Background: Reducing care lapses for people living with HIV is critical to ending the HIV epidemic and beneficial for their health. Predictive modeling can identify clinical factors associated with HIV care lapses. Previous studies have identified these factors within a single clinic or using a national network of clinics, but public health strategies to improve retention in care in the United States often occur within a regional jurisdiction (eg, a city or county).
Objective: We sought to build predictive models of HIV care lapses using a large, multisite, noncurated database of electronic health records (EHRs) in Chicago, Illinois.
Methods: We used 2011-2019 data from the Chicago Area Patient-Centered Outcomes Research Network (CAPriCORN), a database including multiple health systems, covering the majority of 23,580 people with an HIV diagnosis living in Chicago. CAPriCORN uses a hash-based data deduplication method to follow people across multiple Chicago health care systems with different EHRs, providing a unique citywide view of retention in HIV care. From the database, we used diagnosis codes, medications, laboratory tests, demographics, and encounter information to build predictive models. Our primary outcome was lapses in HIV care, defined as having more than 12 months between subsequent HIV care encounters. We built logistic regression, random forest, elastic net logistic regression, and XGBoost models using all variables and compared their performance to a baseline logistic regression model containing only demographics and retention history.
Results: We included people living with HIV with at least 2 HIV care encounters in the database, yielding 16,930 people living with HIV with 191,492 encounters. All models outperformed the baseline logistic regression model, with the most improvement from the XGBoost model (area under the receiver operating characteristic curve 0.776, 95% CI 0.768-0.784 vs 0.674, 95% CI 0.664-0.683; P<.001). Top predictors included the history of care lapses, being seen by an infectious disease provider (vs a primary care provider), site of care, Hispanic ethnicity, and previous HIV laboratory testing. The random forest model (area under the receiver operating characteristic curve 0.751, 95% CI 0.742-0.759) revealed age, insurance type, and chronic comorbidities (eg, hypertension), as important variables in predicting a care lapse.
Conclusions: We used a real-world approach to leverage the full scope of data available in modern EHRs to predict HIV care lapses. Our findings reinforce previously known factors, such as the history of prior care lapses, while also showing the importance of laboratory testing, chronic comorbidities, sociodemographic characteristics, and clinic-specific factors for predicting care lapses for people living with HIV in Chicago. We provide a framework for others to use data from multiple different health care systems within a single city to examine lapses in care using EHR data, which will aid in jurisdictional efforts to improve retention in HIV care.
Retention in HIV care is an important step in the HIV care continuum for both public health efforts to end the HIV epidemic and for the individual health of people living with HIV . The Centers for Disease Control and Prevention aims to reduce new HIV infections by at least 90% over the next 10 years [ ]. Out of every 10 cases of HIV transmission, 8 are associated with people living with HIV who are not engaged in regular HIV medical care [ ]. As part of the goal to reduce new HIV infections, it is important to ensure people living with HIV are receiving regular medical care. Currently, only 50.1% of people living with HIV in the United States are receiving continuous HIV care [ ]. Therefore, increasing the percentage of people living with HIV retained in care is critical to ending the HIV epidemic in the United States.
Many methods to improve retention in care are being piloted across the United States . In HIV clinical settings, the use of case managers to assist with appointment scheduling and health care navigation has been shown to improve retention [ ]. Another initiative, Data to Care, is a strategy by public health departments to use HIV surveillance data to identify people living with HIV that have fallen out of care and attempt to relink them [ ]. However, a study on the Data to Care program in Seattle and King County, Washington, found that only 50% of people living with HIV were successfully reengaged in care, suggesting a need for additional strategies to increase the number of people living with HIV in care [ ]. One upstream intervention would be to identify people living with HIV who are more likely to fall out of care prior to lapsing in care, allowing them to be prioritized for intense case management and other retention interventions.
Previous studies have created predictive models to identify people who are likely to be at risk for lapses in HIV care. Ramachandran et al  used electronic health record (EHR) data from a single urban HIV clinic to predict retention in care. Similarly, Ridgway et al [ ] used EHR data from multiple HIV clinics across the United States to predict retention in care. These studies identified factors associated with lapses in HIV care at the clinic or national level, but public health strategies to improve retention in care in the United States often occur at the city or county level. Local socio-structural factors may impact retention in care among people living with HIV within different cities in unique ways. To our knowledge, no prior studies have predicted which patients will lapse from HIV care within a single urban city (ie, using EHR data from multiple clinics within a single city or geographic region). Our study focuses on the city of Chicago, which is located in Cook County, Illinois, a priority jurisdiction for the Centers for Disease Control and Prevention’s Ending the HIV Epidemic initiative [ ]. The Chicago Department of Public Health (CDPH) is responsible for identifying and implementing HIV public health strategies, including the Ending the HIV Epidemic initiative, for both HIV prevention and HIV treatment within the city.
Goal of This Study
This study aimed to develop predictive models for determining which patients are at risk for lapses in HIV care in Chicago using the Chicago Area Patient-Centered Outcomes Research Network (CAPriCORN). CAPriCORN contains EHR data from a diverse group of 11 different health care systems in Chicago, including academic medical centers, community hospitals, and community health clinic networks . Importantly, CAPriCORN contains data for 12.8 million people, including the majority of 23,580 diagnosed people living with HIV in the city of Chicago [ , ]. Data from CAPriCORN allowed us to follow people living with HIV who obtain HIV care across the different CAPriCORN health care facilities in Chicago. We used these data to identify those engaged in HIV care and to build predictive models to identify people living with HIV at risk for lapses in HIV care within the city of Chicago.
We used deidentified EHR data obtained from 7 sites in CAPriCORN, selecting records from adults (aged 18 years and older) with encounters between January 1, 2011, and September 5, 2019. The hospitals included in the CAPriCORN research database use several different EHR systems, including Epic, Cerner, GE Centricity, eClinicalWorks, Sunquest, and Veterans Health Information Systems and Technology Architecture . CAPriCORN was responsible for all harmonization of data across different EHR systems. Data from CAPriCORN are linked by using a hash-based data deduplication method, allowing researchers to identify and follow patients across the different Chicago health care systems [ ]. All EHR data were provided at the encounter level, allowing us to incorporate changes to sociodemographic, medical diagnoses, and laboratory information over time. Data contained sociodemographic information, encounter information, diagnosis codes, laboratory data, and antiretroviral therapy (ART) medication or prescription drug orders. We used the following criteria to determine if an individual was HIV-positive: (1) a positive HIV laboratory test (confirmatory HIV antibody, p24 antigen, or HIV viral load > 20 copies/mL), (2) an HIV viral load test performed (regardless of results) and prescription for ART, excluding pre-exposure prophylaxis prescriptions, (3) an HIV diagnosis code (International Classification of Diseases [ICD] version 9 [ICD-9] codes 42, 079.53, 795.71, and V08; ICD version 10 [ICD-10] codes B20, R75, and Z21) and a prescription for ART, excluding pre-exposure prophylaxis, or (4) an HIV diagnosis code and 2 HIV viral load tests performed [ ]. Individuals were only included in the study if they met one (or more) of the above criteria and had at least two HIV care encounters. An HIV care encounter was defined as an ambulatory encounter that was either (1) with an infectious disease (ID) provider (taxonomy code “207RI0200X” or “2080P0208X”), or (2) associated with an HIV viral load test or CD4 T cell test. This definition allowed for the inclusion of both people living with HIV who receive their HIV care from their primary care physician as well as those who receive HIV care from their ID physician. To avoid double counting of care visits, we considered visits and laboratory tests within 7 days of each other to be part of the same encounter.
This study was approved by the Chicago Area Institutional Review Board (IRB #00009693).
The outcome of interest in all models was a lapse in HIV care, defined as not having a second HIV care encounter within 1 year after the original HIV care encounter. Because people living with HIV with stable viral loads may have longer time periods between their appointments (eg, from 6 to 9 months), a 1-year gap was chosen to distinguish between these patients and people living with HIV who had truly lapsed in care [, , ]. People living with HIV were censored from the study when <12 months of study time remained to ensure that all participants had the opportunity to be identified as having a lapse in care.
We initially created a list of 134 variables for inclusion in the models; however, 20 variables were excluded in the final models due to multicollinearity between these variables and those already in the model. Final models were built using 114 variables including patient demographics, history of lapse in care, site, provider specialty, insurance information, laboratory tests and results, ART medications, and patient diagnoses. Medical diagnoses were based on the presence of the ICD-9 and ICD-10 diagnosis codes. Specifically, we examined diagnoses of cancer, cardiovascular disease, diabetes, mental health disorders (including mood and anxiety disorders), obesity, sexually transmitted infections (STIs; ie, chlamydia, gonorrhea, hepatitis B, hepatitis C, and syphilis), as well as the use of alcohol, cannabis, opioids, sedatives, stimulants, and tobacco. CAPriCORN provided racial demographic information, with patients able to select more than one race. Laboratory test variables included HIV viral load, CD4 count, HIV genotype, STI tests, and other ID tests within the preceding 12 months. The complete list of variables, including those excluded due to multicollinearity, as well as their ICD-9 or ICD-10 codes, is available inand .
Encounter data were randomly split into a training set (70% of total data) and a validation set (30% of total data). The split was performed while holding a constant ratio of the outcome variable. We created a baseline logistic regression model using only patient demographics and previous history of lapses in care allowing us to quantify how much more predictive capability was provided by including additional data related to laboratory tests and diagnoses. We then created 4 different predictive models: logistic regression, elastic net logistic regression, random forest, and XGBoost. The hyperparameters for the elastic net, random forest, and XGBoost models were tuned using a grid search and can be found in. The hyperparameters were optimized using 5-fold cross-validation on the metric of the area under the receiver operating characteristic curve (AUC).
The performance of these 4 models was compared with each other and with the baseline logistic regression model. Model performance was evaluated using AUC via DeLong’s  method. Variable influence for each model was evaluated using the absolute value of the t statistic for the baseline and full logistic regression models, the value of coefficients for the elastic net logistic regression, the Gini index for the random forest, and the relative influence for the XGBoost model [ , ]. Each model can provide the association of variables; however, unlike the baseline regression, full logistic regression, and elastic net logistic regression models, the random forest and XGBoost models we created cannot provide the directionality of association. All analyses were performed using R (version 4.0.3; R Core Team).
The study population contained 16,930 people living with HIV with a total of 191,492 HIV care encounters during the study period. When examining the demographic information for participants based on information provided at their last HIV care encounter, people living with HIV in this study were predominantly male (75.2%, 12,733/16,930), Black (61.4%, 10,395/16,930), non-Hispanic or -Latino (84.2%, 14,249/16,930), and had all HIV care encounters at a single site within CAPriCORN (95.3%, 16,135/16,930). Based on each participant’s last 2 HIV care encounters, 25.1% (4257/16,930) of participants had a lapse in care (). Over the study period, 42.1% (7135/16,930) of participants had a change in their lapse in HIV care status (either from engaged to lapsed in care or vice versa). The different care sites served a wide range of number of patients; the smallest site served 224 unique people living with HIV, and the largest site served 10,035 unique people living with HIV. The proportion of lapses in care varied among sites. The worst-performing site was Site F, with 21.4% (61/292) of encounters resulting in a lapse in care. The best-performing site was Site G, with 3.7% (5283/141,864) of encounters resulting in a lapse in care. Over the study period, 94.4% (180,777/191,492) of HIV care encounters were followed by a second HIV care encounter within 1 year.
|Variable of interest and values||Frequency, n (%)|
|Lapses in HIV care|
|No lapse in care||12,673 (74.9)|
|Lapse in care||4257 (25.1)|
|African American or Black||10,395 (61.4)|
|Hispanic or Latino||2681 (15.8)|
|Non-Hispanic or -Latino||14,249 (84.2)|
|Less than 25||868 (5.1)|
|25 to less than 35||3080 (18.2)|
|35 to less than 45||3520 (20.8)|
|45 to less than 55||4587 (27.1)|
|55 or older||4875 (28.8)|
|Primary payer type|
|Single site||16,135 (95.3)|
|Mental health||5757 (34.0)|
|Mood disorder||3914 (23.1)|
|Gonorrhea laboratory test in the past year|
|Chlamydia laboratory test in the past year|
Using our validation data set containing 57,447 encounters, all models created using 114 variables outperformed the baseline logistic regression model. The greatest improvement in AUC was seen with the XGBoost model (AUC 0.776, 95% CI 0.768-0.784 vs 0.674, 95% CI 0.664-0.683 for the baseline model; P<.001;).
The XGBoost model had a significant difference in performance between the full logistic regression (AUC 0.776, 95% CI 0.768-0.784 vs 0.752, 95% CI 0.743-0.760; P<.001), elastic net logistic regression (AUC 0.776, 95% CI 0.768-0.784 vs 0.754, 95% CI 0.746-0.762; P<.001), and random forest (AUC 0.776, 95% CI 0.768-0.784 vs 0.751, 95% CI 0.742-0.759; P<.001) models. There was a significant difference in performance between the elastic net logistic regression model and the full logistic regression model (AUC 0.754, 95% CI 0.746-0.762 vs 0.752, 95% CI 0.743-0.760; P<.001); however, the magnitude of the difference between the 2 models was slight and may not be clinically meaningful. There was no significant difference in performance between the elastic net logistic regression model and the random forest model (AUC 0.754, 95% CI 0.746-0.762 vs 0.751, 95% CI 0.742-0.759; P=.43;).
Predictor Variable Importance
The top 10 most influential predictors in each model are listed inand . Although each model had slightly different results, common predictors included no lapse in care in the prior year, being seen by an ID provider (vs a primary care provider), and Hispanic or Latino ethnicity. Influential predictors for the full logistic regression model were no lapse in care in the prior year (adjusted odds ratio [aOR] 0.42, 95% CI 0.39-0.46; P<.001), being seen by an ID provider (vs a primary care provider; aOR 0.54, 95% CI 0.49-0.58; P<.001), age of 55 years or older (vs 25 years to less than 35 years; aOR 0.56, 95% CI 0.51-0.61; P<.001), Hispanic ethnicity (aOR 0.57, 95% CI 0.52-0.63; P<.001), age of 45 years to <55 years (vs 25 years to less than 35 years; aOR 0.68, 95% CI 0.63-0.73; P<.001), and site F (aOR 6.39, 95% CI 3.89-10.47; P<.001).
|Model and variable||aORa (95% CI)||P value|
|No lapse in care in the prior 12 months||0.34 (0.32-0.36)||<.001|
|Hispanic ethnicity||0.32 (0.29-0.34)||<.001|
|Age of 55 years or older||0.53 (0.49-0.57)||<.001|
|Black race||0.41 (0.37-0.46)||<.001|
|Age of 45 years to less than 55 years||0.68 (0.63-0.73)||<.001|
|Unknown race||1.72 (1.52-1.94)||<.001|
|White race||0.63 (0.56-0.70)||<.001|
|Asian race||0.40 (0.31-0.51)||<.001|
|Age of 35 years to less than 45 years||0.86 (0.80-0.93)||<.001|
|Multiracial race||0.64 (0.41-0.99)||.047|
|Full logistic regression|
|No lapse in care in the prior 12 months||0.42 (0.39-0.46)||<.001|
|Infectious disease provider||0.54 (0.49-0.58)||<.001|
|Age of 55 years or older||0.56 (0.51-0.61)||<.001|
|Hispanic ethnicity||0.57 (0.52-0.63)||<.001|
|Age of 45 years to less than 55 years||0.68 (0.63-0.73)||<.001|
|Site F||6.39 (3.89-10.47)||<.001|
|CD4 T cell percentage result greater than or equal to 28 in the prior 12 months||1.36 (1.25-1.48)||<.001|
|Gonorrhea laboratory test in the prior 12 months||0.35 (0.26-0.47)||<.001|
|Primary insurance unknown||0.56 (0.47-0.66)||<.001|
|Cancer diagnosis ever||0.78 (0.72-0.83)||<.001|
aaOR: adjusted odds ratio.
|Model and variable||Variable importance score|
|Elastic net logistic regression|
|No lapse in care in the prior 12 months||–0.835|
|Chlamydia laboratory test in the prior 12 months||0.815|
|Gonorrhea laboratory test in the prior 12 months||–0.769|
|Infectious disease provider||–0.583|
|Primary insurance unknown||–0.506|
|Emergency medicine provider||0.504|
|Primary insurance Medicaid||216.8|
|No lapse in care in the prior 12 months||205.9|
|Hypertension diagnosis ever||200.0|
|Age 35 years to less than 45 years||195.0|
|Tobacco diagnosis ever||194.5|
|Age of 45 years to less than 55 years||193.0|
|Mental health diagnosis ever||190.5|
|CD4 T cell percentage result less than 28 in the prior 12 months||184.1|
|Mood disorder diagnosis ever||172.5|
|Syphilis diagnosis ever||171.4|
|No lapse in care in the prior 12 months||0.065|
|Infectious disease provider||0.045|
|Previous viral load result suppressed||0.031|
|Age of 55 years or older||0.019|
|Primary insurance Medicaid||0.019|
|CD4 T cell percentage result greater than or equal to 28 in the prior 12 months||0.018|
|HIV viral load laboratory result unsuppressed in the last 12 months||0.017|
Our study found that 25.1% (4257/16,930) of patients experienced a lapse in care based on their last eligible HIV care encounter, which is lower than the CDPH estimate of 33.8% of people living with HIV in Chicago not accessing HIV care . Our lower percentage is likely due to two factors: (1) a slightly different patient care population and (2) different definitions being used to determine who was not in care. We defined a lapse in HIV care as not having a second HIV care encounter within 1 year after the original HIV care encounter; however, CDPH defines a lapse in HIV care as not having at least one HIV laboratory test (CD4 or viral load) within a year [ ]. Additionally, we only calculated lapses in care among patients with at least two HIV care encounters, rather than among all people living with HIV within the jurisdiction of Chicago identified in the enhanced HIV/AIDS Reporting System used by CDPH [ ]. By limiting our study population to people with at least two HIV care encounters, we may have selected patients with a history of retention in care, resulting in a patient population less likely to experience a lapse in care.
In this study, we developed predictive models for lapses in care among people living with HIV using linked data that incorporated information from multiple health care systems across Chicago. Our XGBoost model outperformed our baseline model, which included only information about demographics and past lapses in care. Across the different models, we assessed various HIV health care and STI-related factors in predicting the likelihood of lapses in HIV care. In all models, prior lapses in care were among the most influential predictors of future lapses in care. This observation aligns with other studies that have found previous retention in care to be important in predicting future retention [, , ]. Our models also showed that higher CD4 T cell counts are associated with an increased risk of lapsing in care, supporting previous findings [ - ]. A previous suppressed HIV viral load laboratory result and an unsuppressed HIV viral load laboratory result in the prior 12 months were found to be associated with lapsing in care; however, the directionality could not be determined from the XGBoost model. A gonorrhea laboratory test in the last year was predictive of a decreased risk of lapsing in care in the logistic regression model and the elastic net logistic regression model. A chlamydia laboratory test in the last year was predictive of an increased risk of lapsing in care in the elastic net logistic regression model.
In addition to HIV-related factors, we also examined the ability of sociodemographic characteristics to predict lapses in HIV care. Since CAPriCORN covers the majority of people living with HIV in Chicago, this is a unique opportunity to view how these factors impact lapses in care in Chicago. We found that younger age is a predictor for lapses in care, which has been seen in other studies [, , , - ]. Hispanic or Latino ethnicity was associated with a decreased risk of lapsing in care in this study, a finding supported by some previous literature but contradicted by others [ , , , - ]. Another predictor for not lapsing in care was the female sex, consistent with previous studies [ , ].
In terms of general medical and health care factors, we found that people living with HIV with chronic comorbidities, particularly cancer, are less likely to lapse in care, as others have also found . This lower risk of lapsing in care may result from people with other comorbidities being more likely to follow up with both their chronic disease health care provider as well as their ID provider [ , ]. The comorbidity of hypertension (having hypertension or being hypertension-free) was also associated with lapsing in care [ , ]. Unfortunately, this result comes from the random forest model, so we cannot assess the directionality of the association. Finally, despite examining data from a single city, we found that the HIV care site was associated with a higher risk of lapsing in care. The HIV care sites included in our study ranged from a small percentage to almost a fifth of encounters resulting in a lapse of care, which could be due to a myriad of clinic-level factors, such as different patient demographics, acceptance of multiple insurance types, provider familiarity and trust, different neighborhood locations, and varying levels of resources available for identifying patients at risk for lapses in care and reengaging them. Additionally, structural racism and disinvestment in certain neighborhoods in Chicago have disproportionately impacted marginalized communities and the clinics that serve these communities, which may also be reflected in clinic site differences [ ]. Our findings regarding the impact of chronic comorbidity and clinic-level factors on risk for lapsing in care suggest that public health departments may benefit from using data from city-wide research networks like CAPriCORN to better understand which people living with HIV have access to HIV care and are engaged in care.
Consequently, our model could allow public health agencies to identify which groups of people living with HIV within Chicago are at the highest risk of lapsing in care, allowing resources to be targeted toward these patients before they lapse in care. The integration of research data with city- or county-based HIV surveillance data can provide a more complete picture of access to and engagement with HIV care than either data source alone. For instance, CDPH, like most city public health departments, can follow the entire population of people living with HIV in Chicago due to mandatory laboratory reporting of HIV tests. CDPH also has precise knowledge of who enters or leaves the jurisdiction. This information was unavailable to us in the CAPriCORN health network. However, CAPriCORN contains a wide range of information derived from the EHRs of patients, encompassing their entire range of medical illnesses and test results, instead of the public health departments’ relatively limited HIV-related medical information. City health departments typically use viral load and CD4 count laboratory tests as proxies for HIV care visits, whereas we gathered information from both laboratory and clinic visits, allowing us to capture the full range of HIV care encounters and providing a unique way to look at the city-wide continuum of care. For instance, using the CAPriCORN data, we were able to determine that most patients receive care from a single site; only 4.7% (795/16,930) of patients received care from more than one site within the research network. This is less than a study that found that 8% of people living with HIV in Philadelphia received HIV care from multiple sites . Having the ability to determine this important information about the care continuum demonstrates the benefits of incorporating data from EHRs through research networks like CAPriCORN into public health surveillance for HIV care.
There are several limitations to this study. In order to deidentify the data set and protect patient privacy, CAPriCORN removed specific dates of encounters. Therefore, we used patient age in days to determine the relative time between encounters. The lack of dates also means that we cannot account for study-wide temporal trends. Some of our sites also had incomplete data. For example, 1 site was missing HIV laboratory data, resulting in no eligible HIV care encounters, meaning that we were unable to incorporate this site into our models. This incompleteness of data shows a larger limitation of city-wide research networks. It is critical for cities using our framework to focus on incorporating high-quality data, ensuring the most accurate and complete data are available for tracking engagement in HIV care. Additionally, we determined that an HIV care encounter included primary care visits in which an HIV viral load test or CD4 test were performed. However, it is possible that some primary care encounters were for HIV care but did not include performing an HIV viral load test or CD4 test leading to misclassifications. Similarly, we focused on 1 definition of a lapse in HIV care as our outcome, but there are multiple standards for defining a lapse in HIV care . Future studies should compare different standards, such as the Health Resources and Services Administration HIV/AIDS Bureau measure [ ]. As we did not have access to individual patient records for review of medical records, we used ICD-9/10 codes to identify diagnoses, which might lack sensitivity and specificity in identifying patients with medical comorbidities. However, ICD-9/10 codes provide a ubiquitous way for researchers to build predictive models using easily accessible data even if they lack the resources to carry out a review of medical records for every patient. Previous studies have shown that using ICD-9/10 codes can be effective at scaling up the rapid identification of people living with HIV [ , ]. Future studies should include social history variables and unstructured fields or free-text notes to see if they improve overall model performance. This model is possible because of the existence of the CAPriCORN research network. For this work to be extended in other settings, a deidentified, centralized data source from multiple hospital systems would need to be created in other jurisdictions. Our study is retrospective, so the performance of the model in a prospective setting should be tested in future studies prior to implementation.
We built predictive models for lapses in HIV care for people living with HIV from 7 different health care systems across Chicago and identified health care factors found in structured EHR data that are important for predicting lapses in care. Our models also serve as a proof of concept to show that differences in site-based practices (ie, laboratory tests ordered, diagnosis codes, different EHR systems, etc) do not prevent the ability to develop predictive models for lapses in care within a city. To our knowledge, this is the first time predictive models of lapses in HIV care have been built using EHR data from a large research network of health care systems within a major urban city with a high HIV prevalence rate. By focusing on data from patients served by a single public health department jurisdiction, findings from this study may inform HIV public health strategies for the city of Chicago. This model framework can also be adapted and used in clinics to identify patients most at risk for lapses in care who would benefit from tailored interventions or resources, thereby mitigating the number of people living with HIV who fall out of care and supporting efforts to end the HIV epidemic.
The authors wish to thank CAPriCORN for providing the data used in this publication. This work was supported by the National Institutes of Health (K23 MH121190).
Conflicts of Interest
JPR reports receiving fees from Gilead Sciences for consulting. EEF receives funding from the Gilead Sciences Focus grant as well as the Ryan White grant.
Clinical variables included in prediction models of lapses in care for people living with HIV in Chicago, Illinois, 2011-2019.XLSX File (Microsoft Excel File), 54 KB
ART codes included in prediction models of lapses in care for people living with HIV in Chicago, Illinois, 2011-2019.XLSX File (Microsoft Excel File), 14 KB
Hyperparameter values used in prediction models of lapses in care for people living with HIV in Chicago, Illinois, 2011-2019.DOCX File , 15 KB
Sensitivity and specificity values at different cut-off points for prediction models of lapses in care for people living with HIV in Chicago, Illinois, 2011-2019.DOCX File , 16 KB
- Li Z, Purcell DW, Sansom SL, Hayes D, Hall HI. Vital Signs: HIV Transmission Along the Continuum of Care - United States, 2016. MMWR Morb Mortal Wkly Rep 2019 Mar 22;68(11):267-272 [FREE Full text] [CrossRef] [Medline]
- Ending the HIV epidemic: HIV treatment is prevention. Centers for Disease Control and Prevention. 2019. URL: https://www.cdc.gov/vitalsigns/end-hiv/index.html [accessed 2022-03-24]
- Monitoring selected national HIV prevention and care objectives by using HIV surveillance data United States and 6 dependent areas, 2019. Centers for Disease Control and Prevention. 2021. URL: https://www.cdc.gov/hiv/library/reports/hiv-surveillance/vol-26-no-2/index.html [accessed 2022-03-24]
- Department of Health and Human Services. Ending the HIV epidemic. HIV.gov. 2021. URL: https://www.hiv.gov/federal-response/ending-the-hiv-epidemic/overview [accessed 2022-03-29]
- Davila JA, Miertschin N, Sansgiry S, Schwarzwald H, Henley C, Giordano TP. Centralization of HIV services in HIV-positive African-American and Hispanic youth improves retention in care. AIDS Care 2013;25(2):202-206. [CrossRef] [Medline]
- Sweeney P, DiNenno EA, Flores SA, Dooley S, Shouse RL, Muckleroy S, et al. HIV data to care-using public health data to improve HIV care and prevention. J Acquir Immune Defic Syndr 2019;82(suppl 1):S1-S5 [FREE Full text] [CrossRef] [Medline]
- Chang EJ, Fleming M, Nunez A, Dombrowski JC. Predictors of successful HIV care re-engagement among persons poorly engaged in HIV care. AIDS Behav 2019;23(9):2490-2497 [FREE Full text] [CrossRef] [Medline]
- Ramachandran A, Kumar A, Koenig H, De Unanue A, Sung C, Walsh J, et al. Predictive analytics for retention in care in an urban HIV clinic. Sci Rep 2020;10(1):6421 [FREE Full text] [CrossRef] [Medline]
- Ridgway JP, Ajith A, Friedman EE, Mugavero MJ, Kitahata MM, Crane HM, et al. Multicenter development and validation of a model for predicting retention in care among people with HIV. AIDS Behav 2022;26(10):3279-3288 [FREE Full text] [CrossRef] [Medline]
- Jurisdictions: ending the HIV Epidemic. Centers for Disease Control and Prevention. 2021. URL: https://www.cdc.gov/endhiv/jurisdictions.html [accessed 2022-05-18]
- Kho AN, Hynes DM, Goel S, Solomonides AE, Price R, Hota B, CAPriCORN Team. CAPriCORN: Chicago area patient-centered outcomes research network. J Am Med Inform Assoc 2014;21(4):607-611 [FREE Full text] [CrossRef] [Medline]
- CAPriCORN executive summary February 2022. CAPriCORN. 2022. URL: https://www.capricorncdrn.org/wp-content/uploads/2022/02/CAPriCORN-executive-summary-february2022.pdf [accessed 2022-07-27]
- HIV/STI surveillance report. Chicago Department of Public Health. 2020. URL: https://www.chicago.gov/content/dam/city/depts/cdph/statistics_and_reports/2019_HIV_STI_REPORT_Final.pdf [accessed 2023-04-26]
- Ridgway JP, Mason JA, Friedman EE, Devlin S, Zhou J, Meltzer D, et al. Comparison of algorithms for identifying people with HIV from electronic medical records in a large, multi-site database. JAMIA Open 2022;5(2):ooac033 [FREE Full text] [CrossRef] [Medline]
- Edwards JK, Cole SR, Westreich D, Moore R, Mathews C, Geng E, CNICS Research Network. Loss to clinic and five-year mortality among HIV-infected antiretroviral therapy initiators. PLoS One 2014;9(7):e102305 [FREE Full text] [CrossRef] [Medline]
- Laboratory testing schedule for monitoring people with HIV before and after initiation of antiretroviral therapy. National Institutes of Health. URL: https://clinicalinfo.hiv.gov/en/table/table-3-laboratory-testing-schedule-monitoring-people-hiv-and-after-initiation-antiretroviral [accessed 2022-06-16]
- DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988;44(3):837-845 [FREE Full text] [CrossRef] [Medline]
- Breiman L. Random forests. Mach Learn 2011;45(1):5-32 [FREE Full text] [CrossRef]
- Friedman JH. Greedy function approximation: a gradient boosting machine. Ann Statist 2001;29(5):1189-1232 [FREE Full text] [CrossRef]
- Pence BW, Bengtson AM, Boswell S, Christopoulos KA, Crane HM, Geng E, et al. Who will show? Predicting missed visits among patients in routine HIV primary care in the United States. AIDS Behav 2019;23(2):418-426 [FREE Full text] [CrossRef] [Medline]
- Giordano TP, Hartman C, Gifford AL, Backus LI, Morgan RO. Predictors of retention in HIV care among a national cohort of US veterans. HIV Clin Trials 2009;10(5):299-305 [FREE Full text] [CrossRef] [Medline]
- Fleishman JA, Yehia BR, Moore RD, Korthuis PT, Gebo KA, HIV Research Network. Establishment, retention, and loss to follow-up in outpatient HIV care. J Acquir Immune Defic Syndr 2012;60(3):249-259 [FREE Full text] [CrossRef] [Medline]
- Cunningham CO, Buck J, Shaw FM, Spiegel LS, Heo M, Agins BD. Factors associated with returning to HIV care after a gap in care in New York state. J Acquir Immune Defic Syndr 2014;66(4):419-427 [FREE Full text] [CrossRef] [Medline]
- Catz SL, McClure JB, Jones GN, Brantley PJ. Predictors of outpatient medical appointment attendance among persons with HIV. AIDS Care 1999;11(3):361-373 [FREE Full text] [CrossRef] [Medline]
- Rebeiro P, Althoff KN, Buchacz K, Gill J, Horberg M, Krentz H, North American AIDS Cohort Collaboration on Research and Design. Retention among North American HIV-infected persons in clinical care, 2000-2008. J Acquir Immune Defic Syndr 2013;62(3):356-362 [FREE Full text] [CrossRef] [Medline]
- Israelski D, Gore-Felton C, Power R, Wood MJ, Koopman C. Sociodemographic characteristics associated with medical appointment adherence among HIV-seropositive patients seeking treatment in a county outpatient facility. Prev Med 2001;33(5):470-475 [FREE Full text] [CrossRef] [Medline]
- Sheehan DM, Fennie KP, Mauck DE, Maddox LM, Lieb S, Trepka MJ. Retention in HIV care and viral suppression: individual- and neighborhood-level predictors of racial/ethnic differences, Florida, 2015. AIDS Patient Care STDS 2017;31(4):167-175 [FREE Full text] [CrossRef] [Medline]
- Myers K, Li T, Baum M, Ibanez G, Fennie K. The individual, interactive, and syndemic effect of substance use, depression, education, and ethnicity on retention in HIV care. Int J STD AIDS 2021;32(2):184-193 [FREE Full text] [CrossRef] [Medline]
- Pecoraro A, Royer-Malvestuto C, Rosenwasser B, Moore K, Howell A, Ma M, et al. Factors contributing to dropping out from and returning to HIV treatment in an inner city primary care HIV clinic in the United States. AIDS Care 2013;25(11):1399-1406 [FREE Full text] [CrossRef] [Medline]
- Smith JD, Davis P, Kho AN. Community-driven health solutions on Chicago's South Side. Stanf Soc Innov Rev 2021;19(3):A27-A29 https://ssir.org/articles/entry/community_driven_health_solutions_on_chicagos_south_side [FREE Full text] [CrossRef] [Medline]
- Yehia BR, Schranz AJ, Momplaisir F, Keller SC, Gross R, Frank I, et al. Outcomes of HIV-infected patients receiving care at multiple clinics. AIDS Behav 2014;18(8):1511-1522 [FREE Full text] [CrossRef] [Medline]
- Mugavero MJ, Westfall AO, Zinski A, Davila J, Drainoni ML, Gardner LI, Retention in Care (RIC) Study Group. Measuring retention in HIV care: the elusive gold standard. J Acquir Immune Defic Syndr 2012;61(5):574-580 [FREE Full text] [CrossRef] [Medline]
- Mugavero MJ, Davila JA, Nevin CR, Giordano TP. From access to engagement: measuring retention in outpatient HIV clinical care. AIDS Patient Care STDS 2010;24(10):607-613 https://www.liebertpub.com/doi/10.1089/apc.2010.0086 [FREE Full text] [CrossRef] [Medline]
- Paul DW, Neely NB, Clement M, Riley I, Al-Hegelan M, Phelan M, et al. Development and validation of an electronic medical record (EMR)-based computed phenotype of HIV-1 infection. J Am Med Inform Assoc 2018;25(2):150-157 [FREE Full text] [CrossRef] [Medline]
|aOR: adjusted odds ratio|
|ART: antiretroviral therapy|
|AUC: area under the receiver operating characteristic curve|
|CAPriCORN: Chicago Area Patient-Centered Outcomes Research Network|
|CDPH: Chicago Department of Public Health|
|EHR: electronic health record|
|ICD: International Classification of Diseases|
|ID: infectious disease|
|STI: sexually transmitted infection|
Edited by Y Khader; submitted 27.09.22; peer-reviewed by P Han, D Gopukumar; comments to author 23.01.23; revised version received 01.03.23; accepted 11.04.23; published 17.05.23Copyright
©Joseph A Mason, Eleanor E Friedman, Samantha A Devlin, John A Schneider, Jessica P Ridgway. Originally published in JMIR Public Health and Surveillance (https://publichealth.jmir.org), 17.05.2023.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Public Health and Surveillance, is properly cited. The complete bibliographic information, a link to the original publication on https://publichealth.jmir.org, as well as this copyright and license information must be included.