Community-facility linkage models and maternal and infant health outcomes in Malawi’s PMTCT/ART program: A cohort study

Background In sub-Saharan Africa, 3 community-facility linkage (CFL) models-Expert Clients, Community Health Workers (CHWs), and Mentor Mothers-have been widely implemented to support pregnant and breastfeeding women (PBFW) living with HIV and their infants to access and sustain care for prevention of mother-to-child transmission of HIV (PMTCT), yet their comparative impact under real-world conditions is poorly understood. Methods and findings We sought to estimate the effects of CFL models on a primary outcome of maternal loss to follow-up (LTFU), and secondary outcomes of maternal longitudinal viral suppression and infant “poor outcome” (encompassing documented HIV-positive test result, LTFU, or death), in Malawi’s PMTCT/ART program. We sampled 30 of 42 high-volume health facilities (“sites”) in 5 Malawi districts for study inclusion. At each site, we reviewed medical records for all newly HIV-diagnosed PBFW entering the PMTCT program between July 1, 2016 and June 30, 2017, and, for pregnancies resulting in live births, their HIV-exposed infants, yielding 2,589 potentially eligible mother-infant pairs. Of these, 2,049 (79.1%) had an available HIV treatment record and formed the study cohort. A randomly selected subset of 817 (40.0%) cohort members underwent a field survey, consisting of a questionnaire and HIV biomarker assessment. Survey responses and biomarker results were used to impute CFL model exposure, maternal viral load, and early infant diagnosis (EID) outcomes for those missing these measures to enrich data in the larger cohort. We applied sampling weights in all statistical analyses to account for the differing proportions of facilities sampled by district. Of the 2,049 mother-infant pairs analyzed, 62.2% enrolled in PMTCT at a primary health center, at which time 43.7% of PBFW were ≤24 years old, and 778 (38.0%) received the Expert Client model, 640 (31.2%) the CHW model, 345 (16.8%) the Mentor Mother model, 192 (9.4%) ≥2 models, and 94 (4.6%) no model. Maternal LTFU varied by model, with LTFU being more likely among Mentor Mother model recipients (adjusted hazard ratio [aHR]: 1.45; 95% confidence interval [CI]: 1.14, 1.84; p = 0.003) than Expert Client recipients. Over 2 years from HIV diagnosis, PBFW supported by CHWs spent 14.3% (95% CI: 2.6%, 26.1%; p = 0.02) more days in an optimal state of antiretroviral therapy (ART) retention with viral suppression than women supported by Expert Clients. Infants receiving the Mentor Mother model (aHR: 1.24, 95% CI: 1.01, 1.52; p = 0.04) and ≥2 models (aHR: 1.44, 95% CI: 1.20, 1.74; p < 0.001) were more likely to undergo EID testing by age 6 months than infants supported by Expert Clients. Infants receiving the CHW and Mentor Mother models were 1.15 (95% CI: 0.80, 1.67; p = 0.44) and 0.84 (95% CI: 0.50, 1.42; p = 0.51) times as likely, respectively, to experience a poor outcome by 1 year than those supported by Expert Clients, but not significantly so. Study limitations include possible residual confounding, which may lead to inaccurate conclusions about the impacts of CFL models, uncertain gen-eralizability of findings to other settings, and missing infant medical record data that limited the precision of infant outcome measurement. Conclusions In this descriptive study, we observed widespread reach of CFL models in Malawi, with favorable maternal outcomes in the CHW model and greater infant EID testing uptake in the Mentor Mother model. Our findings point to important differences in maternal and infant HIV outcomes by CFL model along the PMTCT continuum and suggest future opportunities to identify key features of CFL models driving these outcome differences. The study was approved by the Malawi National Health Sciences Research Committee (#17/05/1812), and the institutional review boards of the University of North Carolina at Chapel Hill, United States of America (#17–1114), Brigham & Women’s Hospital, USA (Reliance Agreement #17–1114), and James Cook University, Australia (HREC #1812). All regulatory bodies approved the study protocol (S1 Protocol) and related documents, including exemption of individual informed consent for the review of existing deidentified, routinely collected data. We employed an observational cohort study design, enriched by a field survey, to evaluate the effectiveness of CFL models for improving PMTCT outcomes in high-volume health facilities in 5 districts of Malawi (S1 STROBE Checklist). Specifically, we examined the impact of 3 CFL models on accelerating ART initiation, reducing loss to follow-up (LTFU), and improving viral suppression among women diagnosed with HIV during pregnancy or breastfeeding. We also examined the impacts of these interventions on outcomes among infants exposed to HIV, including uptake of EID testing and infant LTFU over the first year of life. For the study, we purposively selected 5 districts—Lilongwe, Salima, Zomba, and Mzimba North and Mzimba South (referred to collectively here as “Mzimba”)—that together span 4 of 5 health zones nationally; include diverse urban, peri-urban, and rural catchment areas; and capture the 3 CFL models of interest. During the study period, 3 main CFL models—Mentor Mothers, CHWs, and Expert Clients—provided support to PMTCT clients in study districts. To varying degrees, all 3 models offered PMTCT clients health education, ART adherence support, and “back to care” services for women and infants who disengaged from care, among other services. Data from a CFL model survey (S1 Appendix) conducted with 15 CFL model program managers and Ministry of Health (MOH) healthcare supervisors in the 5 study districts provide a summary overview of major CFL model activities and characteristics (S1 Table). Donor-supported NGOs implemented the Mentor Mother (i.e., mothers2mothers) and CHW (i.e., the Tingathe program of the Baylor College of Medicine Children’s Foundation) models, whereas a mix of NGOs (e.g., Elizabeth Glaser Pediatric AIDS Foundation) and the Malawi MOH delivered the Expert Client model, depending on the site. All models had been in operation at the study sites for at least several months and, in most cases, a few years, during the study period. The MOH—together with donors, NGOs, and other implementing partners—determined the allocation of CFL models to study districts and individual health facilities based on public health programming priorities. In most cases, this meant that 1 CFL model was operating per site during the period of interest; however, in a few cases, particularly at the largest urban health facilities in Lilongwe district, this opened the possibility of 2 or more models concurrently operating at the same site. Additional details on the programming, policy, and implementation landscape for the CFL models studied here are described elsewhere [33,34]. The establishment and evolution of Malawi’s national PMTCT program have been well described [3,35]. Briefly, in 2002, Malawi began offering PMTCT services as part of its national HIV program, which initially included HIV testing and counseling and single-dose nevirapine. Over a period of scale-up from 2004 to 2010, the PMTCT program expanded to 454 antenatal clinics nationally offering free integrated HIV testing services, maternal combination antiretroviral prophylaxis, and referral to, and limited on-site provision of, ART services for women meeting criteria for treatment for their own health [35]. In 2011, Malawi introduced the Option B+ strategy, becoming the first country globally to offer rapid, universal, and lifelong ART to PBFW living with HIV regardless of CD4 count or stage of clinical disease. To implement Option B+, Malawi recognized and acted on the need to fully integrate its PMTCT and ART programs at all levels, including for supply chain management, health worker training and supervision, longitudinal cohort reporting of client outcomes, and decentralization of services to all facilities offering antenatal care [35]. The Option B+ experience proved foundational for the national transition to a treat all policy, which was undertaken by the Malawi MOH in mid-2016 and codified in recent national HIV management guidelines [36,37]. Under Malawi’s current integrated PMTCT/ ART program, ART for pregnant women living with HIV may be started and continued within antenatal (i.e., “ANC”)/PMTCT clinics, maternity departments, and/or ART clinics. Following delivery, follow-up for lifelong maternal ART and infant HIV exposure (including EID) is typically delivered in the same HIV care setting (usually the ART clinic) using synchronized appointments [6,38]. Despite programmatic integration, variability has been reported in how facilities organize their PMTCT/ART services, with some using referral systems between the antenatal and ART clinics [38]. The target population for this study comprised all PBFW living with HIV, and their infants, who were referred to, or newly enrolled in, PMTCT services at a high-volume health facility during introduction of national treat all policy in our 5 study districts. To reach this target population, we used electronic MOH-integrated HIV program data to identify high-volume health facilities and surrounding catchment areas (collectively referred to as “sites”) in each district. We designated a site as “high-volume” if it reported ≥30 PBFW newly diagnosed with HIV in the year prior to treat all introduction, which generally included sites in the 50th percentile and higher for new PMTCT client volumes in the district. These high-volume sites were prioritized since they were the most likely to offer CFL model services. We excluded high-volume sites serving highly transient catchment areas (reflected by high rates of “transfer out” from the PMTCT program) that would severely limit outcome ascertainment in the routine record. We visited all eligible high-volume sites in 4 of the 5 study districts (i.e., 22 total) and a 50% random sample of eligible sites (i.e., 8 of 16) in the last study district due to operational constraints. At each site, we constructed a study cohort of mother–infant pairs by collecting routine medical record data on all HIV-diagnosed PBFW meeting study eligibility criteria, and, for pregnancies that resulted in live births, their infant’s medical record data. Mother–infant pairs were included in the study cohort if the mother met the following criteria: newly diagnosed or had documented evidence of recent (i.e., within 90 days) HIV infection at any time between July 1, 2016 and June 30, 2017; pregnant or breastfeeding at the time of HIV diagnosis; documented enrollment in, or referral to (i.e., found to have new or recent HIV infection but not linked to care), ART services within Malawi’s PMTCT program; ≥16 years of age at time of PMTCT enrollment; and received any antenatal, maternity, or PMTCT/ART services at a study site. Because most outcomes (e.g., maternal LTFU) were measured using the MOH HIV treatment card, we excluded any pair in which the mother had a missing HIV treatment card. Next, we invited a random subset of mother–infant pairs that we sampled from the study cohort to complete a 1-time field survey. The field survey collected biomarker and questionnaire data to overcome known limitations with the routine medical record (e.g., missing data), and, thus, enable detailed ascertainment of CFL model exposure, maternal viral load (VL), and infant HIV status. All mother–infant pairs in the study cohort (regardless of whether they ever established care or started ART or were categorized as LTFU in the PMTCT/ART program) were eligible to be sampled for inclusion in the field survey. Our sampling approach is presented in the supporting information (S2 Table). Briefly, we sampled a large proportion of the study cohort for the field survey, including 100% of mother–infant pairs at 28 of 30 eligible high-volume sites, due to concerns about high rates of survey nonresponse and operational challenges with conducting community tracing and field survey activities across multiple sites (S2 Table). Survey data enriched data for the larger study cohort and were used specifically to (1) impute CFL model exposure among mother–infant pairs; (2) improve imputation of maternal VL among women missing VL data in the routine record; and (3) augment infant HIV testing data. Our primary outcome of interest was maternal retention in care in the study cohort, which we measured as LTFU over the PMTCT continuum in order to evaluate all maternal and infant outcomes on the same scale (i.e., with event probabilities increasing over time, rather than decreasing). We defined LTFU according to the MOH definition using dates reported by each facility [36,37]. With this rationale, we described the cumulative incidence of maternal LTFU from study sites at 2 years post-HIV diagnosis among all eligible women. While the primary analysis examined LTFU from study sites, we were also interested in using this outcome as a proxy for loss to care entirely. LTFU from study sites might not correspond to loss to clinical care because some “silent” transfers may not have been documented in the facility records [39]. Thus, we examined the prevalence of such silent transfers using data from the field survey. Other secondary maternal outcomes included the following: the cumulative proportion who initiated ART by 6 months after HIV diagnosis; median time from HIV diagnosis to ART initiation; and proportion having a HIV-1 VL <1,000 copies/ml (i.e., consistent with programmatic definitions of “viral suppression” [40,41]). Since maximizing cumulative time spent retained on ART with viral suppression is the ultimate goal for women living with HIV in any PMTCT program, we also developed a post hoc “longitudinal maternal suppression” metric to estimate the amount of time women stayed in this final, targeted state within the PMTCT continuum. Secondary outcomes also included the following indicators for newborns exposed to HIV: EID testing uptake (requiring documentation of a HIV-1 DNA PCR test result) from age 6 weeks; infant LTFU, defined according to the MOH definition in the routine clinic record [36,37]; and 12-month in care survival. Because relatively few infant deaths and vertical transmission events were observed, we created a post hoc infant “poor outcome” (i.e., a composite outcome encompassing a documented positive HIV test result, LTFU, or death by 1 year of age) to summarize clinical events of interest among infants exposed to HIV. Data collection took place between January 15, 2018 and May 31, 2019, and proceeded in 2 steps. In the first step, we reviewed all routinely available MOH medical record data on each woman and her infant in the study cohort from maternal HIV diagnosis through the child’s second birthday. Sources for medical record information included maternal and pediatric HIV treatment cards, the national Baobob electronic medical record, ART registers, infant follow-up cards and registers, maternity registers, and HIV VL and EID HIV-1 DNA PCR testing registers. The MOH Department of HIV and AIDS curated these routine data sources throughout the study period and ensured data quality through a robust supportive supervision protocol implemented quarterly since 2004 to collect and verify site-level data [6]. NGO partners operating in each district supported the MOH Department of HIV and AIDS protocol through monitoring and evaluation activities designed to promote routine data accuracy and completeness for their respective program reporting purposes. From the aforementioned routine information sources, we abstracted data on key variables necessary to reconstruct the PMTCT continuum for all mother–infant pairs in the study cohort, including the following: maternal HIV testing history; 12- and 24-month maternal PMTCT/ART program status (i.e., alive in care, transferred out, stopped ART, died, and lost to follow-up [LTFU]); 6- and 24-month maternal VL (timing per national guidelines); infant first HIV-1 DNA PCR result (from 6 weeks of age per national guidelines); 12- and 24-month infant follow-up program status (i.e., alive in care, transferred out, died, and LTFU); and 12- and 24-month infant HIV sero-status (timing of rapid testing per national guidelines) [36,37]. In the second step of data collection, study and MOH staff used detailed locator information available at the clinic to contact all mother–infant pairs sampled for the field survey. Contact with mother–infant pairs took place at routine clinic visits and through community outreach, telephone calls, and/or in-person tracing to introduce them to the study. Mother–infant pairs expressing interest, and who could be reached to complete written informed consent procedures, were asked to complete the field survey. The field survey had 2 parts: (1) a tablet-based questionnaire comprised of a series of electronic case reporting forms to collect information on participant experiences with CFL models and maternal and infant medical histories, including history of VL and EID testing in the routine PMTCT/ART program; and (2) study-initiated blood sample collection for maternal VL and HIV testing on the child from the index pregnancy (i.e., HIV-1 DNA PCR testing for infants ≤12 months and rapid testing for young children >12 months). In cases where next of kin of a deceased mother were reached, a brief verbal autopsy was conducted to confirm the maternal death and to obtain an approximate date of death. All abstracted routine medical record information was entered directly into a secure study database by trained staff and subjected to data logic constraints at the time of data entry. Field survey data were combined electronically with the routine medical record data, and the resulting merged database was subjected to weekly local and monthly central quality control checks to identify missing and discrepant data. Queries were addressed during weekly study team meetings, and review of source medical records and case reporting forms. Discrepant values for the same variable observed across 2 or more data sources were resolved by committee adjudication involving the principal investigator, senior analyst, and study coordinator. Additional details about data management processes are provided in the study protocol (S1 Protocol). Field survey participants who had not completed MOH-recommended VL or EID testing at the time of consent had samples taken for routine “catch up” testing. Field survey participants who had received MOH-recommended VL or EID testing, but who did not have a documented result within 6 months of study consent, also had specimens collected for testing. Blood sampling involved finger or heel (for infants ≤12 months) prick for dried blood spot sample collection [36,37]. All study-specific testing was done at the University of North Carolina Project Tidziwe Laboratory (Lilongwe, Malawi). Quantitative HIV-1 VL testing was done on the Abbot RealTime HIV-1 assay (Abbott Laboratories, Chicago, Illinois, USA); qualitative DNA PCR testing for EID was performed on the Xpert platform (Cephied, Sunnyvale, California, USA). Our study was powered to detect an 8% absolute difference in maternal retention in care between PBFW living with HIV who did, versus those who did not, receive CFL model support per the field survey. With a 2-sided alpha of 0.05, we estimated that we would have to assess outcomes among 894 mother–infant pairs from sites with a CFL model and 298 mother–infant pairs from sites with no CFL model (i.e., a “traditional” standard of care), thus requiring a total of 1,192 PBFW living with HIV, to have 80% power to detect the specified difference in our primary outcome. After protocol development and during programmatic mapping of available CFL models undertaken prior to study data collection, we recognized that a “traditional” standard of care in which no CFL model operated at a site was typically nonexistent. Because of this and the fact that our sample size calculation was based on estimated maternal retention under a no CFL model condition, we conferred a Study Advisory Committee meeting in November 2017 to review our study analysis plan. During that meeting, the Expert Client model was felt to represent an emerging standard of care in the national PMTCT program, since it typically involved MOH-led community outreach and peer support to enhance PMTCT services. As such, this model was considered the referent group for all planned comparative analyses, and we limited analyses for the few mother–infant pairs receiving no CFL model to maternal LTFU and time to maternal ART. Of note, we did not oversample mother–infant pairs for the field survey at sites found to offer a CFL model during programmatic mapping to meet the original sample size. We used responses from the field survey to identify or impute CFL model exposure. Specifically, for mother–infant pairs who participated in the field survey, we used the CFL model they reported receiving to indicate their exposure. For mother–infant pairs who did not participate in the field survey, we imputed CFL model exposure based on the CFL model most frequently cited by field survey participants as providing services at the site during the period of interest (variation in CFL model exposure as reported by field survey participants was limited at most study sites; S3 Table). Statistical analyses focused on the study cohort as the analysis population. We estimated the impact of CFL model type on primary and secondary outcomes by assessing the associations between CFL model and the cumulative incidence of each outcome. Specifically, we estimated cumulative incidence functions under each model using inverse probability–weighted Aalen–Johansen estimators [42]. We also compared the subdistribution hazard functions (which is not influenced by differences in the duration of cohort member follow-up time) for each outcome of interest, using weighted subdistribution hazard ratios computed using the Fine and Gray method [43]. In analyses of ART initiation, LTFU, and viral suppression, death was treated as a competing event. In all analyses, we applied sampling weights to account for the uneven site sampling probabilities across districts. Specifically, sampling weights had a value of 1 for all individuals in the first 4 study districts where we included all eligible high-volume sites, and had a value of 2 (i.e., 1 divided by the sampling fraction of 50%) in the last study district where only half of high-volume sites were sampled. We also applied inverse probability weights to all analyses to account for confounding by facility type/level, which standardized the distribution of health facility level between CFL exposure arms [44]. Inverse probability weights were stabilized by the marginal probability of exposure [45] such that they had the form P (A = a)/P (A = a|L), where A represents CFL exposure arm and L represents facility type. The numerator and denominator of the weights were estimated using logistic regression. We estimated our longitudinal maternal suppression metric by summarizing the percent of days over the 2 years following HIV diagnosis that women spent in HIV care, on ART, and virologically suppressed, and compared this across CFL models. The percent time suppressed was estimated for all eligible women in the cohort over the entire 2-year period by averaging the product of the probability of being retained on treatment (i.e., after starting ART and prior to, or without, becoming LTFU) at any given time point and the probability of being virally suppressed, given one was on ART. The probability of being on treatment at time t was calculated as the cumulative incidence of ART initiation minus the cumulative incidence of LTFU, where cumulative incidence functions were estimated using the Aalen–Johansen estimator, which accounted for right censoring and competing events. The probability of viral suppression, given one was on ART, was estimated at each time point using a 2-stage procedure that made use of both the VL data abstracted from the routine medical record and the VL measurements obtained from the field survey. First, we estimated the probability of having a survey-obtained VL at each time point, given that a patient did not have a routine VL documented in the medical record. Then, we assigned routinely collected VLs a weight of 1 and survey-obtained VLs a weight defined by the inverse probability that a patient had a survey-obtained VL at that time point. In the weighted VL data, we modeled the probability of viral suppression at each time point using a flexible logistic regression model. In this model, we conservatively assumed that the first 30 days on ART were all unsuppressed [46]. After the first 30 days, we allowed the probability of suppression to vary flexibly over time by modeling time since HIV diagnosis using penalized b-splines. Two subgroup analyses were conducted using data from the field survey. First, among people classified as LTFU using the facility records, we examined the proportion in the field survey who reported receiving care anywhere within the past 6 months to estimate the extent of silent transfer. Second, for infants exposed to HIV in the field survey, we estimated the prevalence and prevalence ratio (PR) of vertical HIV transmission by 18 months under each CFL model. Random sampling of sites in the last district was done in STATA (Version 14.1, College Station, Texas, USA). All statistical analyses were conducted using SAS version 9.4 (Cary, North Carolina, USA) and R 3.6.0.