Studying patterns and predictors of HIV viral suppression using A Big Data approach: a research protocol

Viral suppression is the final stage of the HIV treatment cascade, which serves as the framework for UNAIDS’ 90–90–90 goals [1]. Sustained (or durable) viral suppression permits the restoration of immune function, reduces onward secondary transmission, and indicates long-term treatment success and mortality reduction [2]. In the US, ~ 57% of all people living with HIV (PLWH) were virally suppressed according to the national surveillance data from the Centers for Disease Control and Prevention (CDC) [3], and in South Carolina (SC), 62% of PLWH were virally suppressed [4]. “Ending the HIV Epidemic (EtHE): A Plan for America” [5] federal campaign, launched in February 2019 aims to reduce the number of new HIV infections in the US by 75% and 90% by 2025 and 2030, respectively. The EtHE campaign focuses on 48 US counties that contribute to > 50% of new HIV cases and 7 states with a high rural HIV burden, including SC. Sustained viral suppression is one of four strategic areas of the EtHE. With a prolonged life expectancy of PLWH, routine monitoring of viral load (VL) status becomes more important over their life course, with the longitudinal VL information collected over time potentially adding to the predictability of subsequent virologic failure (VF) or mortality. Over the past few years, a small but increasing number of longitudinal studies have explored the dynamics of VL patterns using sustained viral suppression, viral rebound, viral blips, or low-level viremia (LLV) [6‐8]. These different VL measures are interrelated, affect each other, and also predict, to some extent, virologic failure [9]. Studies examining the association between persistent LLV and VF or viral rebound are conflicting [10‐12]. Other studies report a correlation between LLV and the risk of viral rebound [13‐15]. Furthermore, the threshold of LLV at which it would be predictive of VF varies. Some studies suggest a threshold of > 200 copies/ml as being associated with VF; yet other studies suggest a higher threshold (i.e., VL > 400 copies/ml) [16, 17].

The virological outcomes of PLWH could be affected by multiple factors from individual-level (e.g., socio-demographics, clinical characteristics, HIV care-seeking behaviors) to county-level social and environmental factors (e.g., economic environment). Socio-demographics have been frequently reported to be associated with viral suppression. According to the US CDC’s HIV Prevention Progress Report 2019, viral suppression remains lowest among persons ≤ 34 years, Blacks/African Americans, persons who inject drugs, and heterosexuals [18, 19]. The clinical indicators of HIV diagnosis (e.g., pre-antiretroviral therapy [ART] CD4 counts, pre-ART VL level) are important in determining subsequent virologic success or failure after initiation of ART [16, 19‐21]. Individuals who are more immunocompromised (e.g., low baseline CD4 counts, or opportunistic illnesses) at HIV diagnosis are more likely to develop VF [16, 19‐21]. Treatment history, including earlier ART initiation [19], no prior ART use before treatment [20], prior use of mono- or dual- antiretrovirals [22], longer duration of therapy [20], and boosted protease inhibitor (PI)-based regimen [20], also impacts success or failure of sustained viral suppression [16, 21]. Apart from individual-level factors, considerable interest remains in understanding how social and structural determinants of health affect the HIV treatment continuum, including viral suppression. For instance, the structural determinants and socioeconomic conditions of the neighborhoods or communities (e.g., county) where individual lives will profoundly impact the outcomes of the HIV continuum of care [23‐25]. A more comprehensive prediction model for virologic outcomes based on the dynamic patterns of VL, individual demographics, HIV care-seeking behavior, and social and environmental factors, could inform us on “when” and “how” to help individuals with poor viral control to achieve and sustain viral suppression.

Some critical gaps exist in the current efforts to understand the dynamics of viral suppression and the development of an optimal predictive model of viral suppression. First, most studies have focused on limited indicators of viral suppression (e.g., a single time point measure) and have failed to provide a complete picture of the dynamic process of viral suppression. Second, most studies investigating the transformations among virologic outcomes have only explored the monotonous transformation between two virologic outcomes (e.g., from viral suppression/failure to viral failure/suppression) rather than the comprehensive virologic history and dynamic viral patterns [26]. Third, most studies have examined viral suppression within a limited time window ranging from 6 to 48 months [26, 27] and were unable to provide a time-sensitive assessment of the virial suppression process. Fourth, scopes of data in most existing studies are insufficient to fully describe the viral patterns due to limited data sources (e.g., use only medical records or epidemiological surveys). The structural and socioenvironmental factors were not always taken into consideration because of either the unavailability of such data in medical records or the lack of advanced analytic tools that can model such complex data. Fifth, previous studies lacked advanced analytics to predict the VF (or other viral outcomes) using the comprehensive and longitudinal data. Most extant literature counted the presence of virologic outcomes within a limited timeframe and explored their correlates using traditional analytic approaches such as generalized estimate equation [28] and cox regression [7, 8, 29]. Most of these approaches are time insensitive and cannot make dynamic predictions based on the large representative features. Finally, most of the existing research does not go beyond the modeling phase to translate research findings into service-ready clinical tools for improved viral suppression or better viral control.

The increasing availability of electronic data, including electronic medical records, administrative databases, and public county-level data, has created a unique opportunity to expand our ability to measure HIV-related health and clinical outcomes. With the integrated data, we can examine patients’ risk factors at the individual and county levels longitudinally and generate new knowledge of HIV viral suppression. However, these integrated data sources are characterized by high volume and variation, and there are several data analytic challenges in the integrated data structure, including mismatched time scales and multilevel risk predictors. The recent developments in Big Data analytics, such as artificial neural network [50, 51], LSTM Neural Network, random forest [51, 52], support vector machine [51], and deep learning approach such as CNN [53], make it feasible to address these methodological challenges and predict virologic outcomes using data from multiple domains.

The proposed project will integrate complex yet representative population-level HIV data from multiple SC data sources at both individual and county levels and analyze the impact of historic HIV VL patterns on multiple viral outcomes considering multilevel factors. The integrated data from these multiple data sources include PLWH who were diagnosed with HIV (as early as 2005) over a period of > 15 years (to at least 2020) and thus provide us a so-called synthetic cohort which contains a complete population-based longitudinal picture of HIV VL measure, HIV treatment, HIV care-seeking behaviors, hospital diagnosis, and county-level factors. We will develop AI-based modeling accounting for multiple viral outcomes by integrating both individual- and socioecological-level factors to relate it to future viral suppression. With both extensive data integration and Big Data analytics, the proposed research is significant as it will: (1) improve the understanding of the complex inter-related effects of longitudinal trajectories of HIV viral suppressions and HIV treatment history while taking into consideration multiple factors at the individual and socioecological levels among PLWH; and (2) develop empirical public health approaches to achieve ending the HIV epidemic through translating the risk prediction model to a multifactorial decision system that enables the feasibility of AI-assisted clinical decisions.