Text messaging for maternal and infant retention in prevention of mother-to-child HIV transmission services: A pragmatic steppedwedge cluster-randomized trial in Kenya

Background: Timely diagnosis of infant HIV infection is essential for antiretroviral therapy (ART) initiation. In a randomized controlled trial, we found the Texting Improves Testing (TextIT) intervention (a theory-based text messaging system) to be efficacious for improving infant HIV testing rates and maternal retention in prevention of mother-to-child HIV transmission (PMTCT) programs. Using an implementation science approach, we aimed to evaluate real-world effectiveness of the intervention. Methods and findings: In a pragmatic, cluster-randomized, stepped-wedge trial with 2 time periods of observation, we randomly allocated 10 clinics to begin implementing the intervention immediately and 10 clinics to begin implementing 6 months later. To approximate real-world conditions, inclusion criteria were broad. Women at clinics implementing the intervention received up to 14 text messages during pregnancy and after delivery and had the option to respond to text messages, call, or send inquiry text messages to a designated clinic phone. The primary outcomes were infant HIV testing and maternal retention in care during the first 8 weeks after delivery. We used modified Poisson regression with robust variance estimation to estimate the relative risk and 95% confidence intervals (CIs). Generalized estimating equations were applied on individual-level data to account for clustering by site. Between February 2015 and December 2016, 4,681 women were assessed for study participation, and 2,515 were included. Participant characteristics at enrollment did not differ by study arm. Overall median age was 27 years (interquartile range [IQR] 23-30), median gestational age was 30 weeks (IQR 28-34), 99% were receiving ART, and 87% who enrolled during intervention phases owned a phone. Of 2,326 infants analyzed, 1,466 of 1,613 (90.9%) in the intervention group and 609 of 713 (85.4%) in the control group met the primary outcome of HIV virologic testing performed before 8 weeks after birth (adjusted relative risk [aRR] 1.03; 95% CI 0.97-1.10; P = 0.3). Of 2,472 women analyzed, 1,548 of 1,725 (90%) in the intervention group and 571 of 747 (76%) in the control group met the primary outcome of retention in care during the first 8 weeks after delivery (aRR 1.12; 95% CI 0.97-1.30; P = 0.1). This study had two main limitations. Staff at all facilities were aware of ongoing observation, which may have contributed to increased rates of infant HIV testing and maternal retention in care at both intervention and control facilities, and programmatic initiatives to improve maternal and infant retention in care were ongoing at all facilities at the time of this study, which likely limited the ability to demonstrate effectiveness of the trial intervention. Conclusions: In this study, a larger proportion of infants in the intervention group received HIV testing compared with the control group, but the difference was small and not statistically significant. There was also a nonsignificant increase in maternal postpartum retention in the intervention periods. Despite the lack of a significant effect of the intervention, key lessons emerged, both for strengthening PMTCT and for implementation research in general. Perhaps most important, improving the implementation of usual care may have been sufficient to substantially improve infant HIV testing rates. This study was a pragmatic, cluster randomized, stepped-wedge trial with 2 observation time periods (S1 Trial protocol and S1 CONSORT checklist). Twenty clusters were selected for implementation. Ten clusters were randomly allocated to begin implementing the intervention immediately, while the remaining 10 began implementing approximately 6 months later. Study clusters were defined as public health facilities supported by the Family AIDS Care and Educational Services (FACES) program to provide PMTCT services in Kenya. FACES is a United States President’s Emergency Plan for AIDS Relief (PEPFAR)/Centers for Disease Control and Prevention (CDC) funded HIV prevention, care, and treatment program operated jointly by the Kenya Medical Research Institute and the University of California, San Francisco. At the time of this trial, FACES supported 136 government health facilities, spread across 3 counties (Kisumu, Migori, and Homa Bay) in the Nyanza region of Kenya. This region had the highest HIV prevalence in the country (15%) [7]. The top 20 clusters by patient volume (number of newly infected HIV-positive pregnant women in the prior 6 months) were selected for study inclusion. Potential participants included all HIV-positive pregnant women enrolled in the PMTCT program at target health facilities. To approximate real-world conditions, inclusion criteria were broad, with the intent of minimizing exclusion of subpopulations such as those without mobile phones. Women were offered the opportunity to receive text messages if they were ≥18 years old or emancipated minors and between 28 weeks’ gestation and delivery. Similar criteria were used to abstract data for women at control facilities. Women in the intervention group who reported sharing phones but had not disclosed their HIV status to the person sharing the phone did not receive the short message service (SMS) intervention but were included in the intention-to-treat analysis of outcomes. An independent biostatistician at the University of Washington’s Center for AIDS Research generated the randomization sequence and assigned clusters to intervention start periods. Randomization was stratified by clinic volume and experience level as shown in Fig 1. Clinic volume was measured by the average number of pregnant women newly enrolled into the PMTCT program in the 12 months prior to randomization (≥50 versus <50 enrollments per month). Facilities were considered to have prior experience with TextIT if they had been part of a 6-month pilot project run by FACES immediately after release of the initial trial’s results [5]. Facilities with prior experience did not continue to implement TextIT after the 6-month pilot project and were not implementing it at the time of randomization. Within each stratum, an equal number of clinics were randomly assigned to immediate or delayed implementation of the intervention. Due to the nature of the intervention and the need to inform facilities of their participation, it was not possible to blind clusters, healthcare providers, investigators, data analysts, or individual participants to group assignments. Clinic volume was measured by the average number of pregnant women newly enrolled into the PMTCT program in the 12 months prior to randomization (≥50 versus <50 enrollments per month). Facilities were considered to have prior experience with TextIT if they had been part of a 6-month pilot project run immediately after release of the initial trial’s results. The 4 strata included (1) experienced and high volume, (2) experienced and low volume, (3) not experienced and high volume, and (4) not experienced and low volume. Within each stratum, half of clinics were randomly assigned to begin implementing the intervention immediately, while the other half began implementing approximately 6 months later. PMTCT, prevention of mother-to-child HIV transmission. “Mentor mothers” at each health facility abstracted data from routine clinic records and registered women to receive text messages during intervention periods. Mentor mothers are a formal lay cadre of women living with HIV who have recent experience in PMTCT (during the past 6 months to 2 years.) They are deployed to maternal and child health (MCH) clinics to share best practices of PMTCT with their peers attending care as part of the Ministry of Health–approved Kenya Mentor Mother Program [8]. Mentor mothers at phase I intervention sites were trained immediately after randomization, while those at phase II intervention sites were trained at the end of their control period, immediately prior to implementing the intervention. Training was provided in 2 stages. The first stage was a 2-day, conference-based training at a central location, attended only by mentor mothers. The second stage was a 1-day, facility-based training for all MCH service providers at each health facility. The 2-day training for mentor mothers included a review of PMTCT services and national guidelines, overview of the study protocol, training on research ethics, data collection using a mobile phone-based system, and procedures for registering patients to receive SMS text messages from the automated text messaging software. At the end of the 2-day training, an evaluation of competence was conducted. Each mentor mother was required to obtain certification of research ethics training by taking an online course following the 2-day training. The second stage of training was completed on site at each health facility. During this second stage of training, all MCH service providers were given an overview of the TextIT implementation protocol, the data collection procedures, SMS-based participant registration, and the role of the mentor mother in implementing TextIT. This second stage also allowed trainers to meet mentor mothers and address any gaps in competency identified by the evaluation following the initial 2-day training. At facilities randomized to continue usual care, pregnant women did not receive any study-related intervention during the first phase of the trial. Mentor mothers at these facilities were trained, at their respective facilities, on abstracting data. They were aware of the general nature of the intervention and outcome being measured but did not receive any other training until the end of phase I of the study, immediately prior to implementing the intervention in the second phase. When health facilities implemented TextIT, both in phase I (intervention facilities only) and phase II (all facilities) of the study, mentor mothers were trained to provide a brief overview of the TextIT strategy, obtain verbal consent to receive the intervention, and provide a written information statement to women willing to receive messages. Baseline clinical and demographic characteristics were recorded only for women who provided verbal informed consent, after which they were registered to receive intervention text messages. Participants attending clinics implementing TextIT continued to receive standard care even if they did not have access to a phone and were considered as having been exposed to the TextIT intervention in the intent-to-treat analysis. Women registered in the SMS system received up to 14 text messages during pregnancy and after delivery. The message content and schedule were the same as those used in the earlier study demonstrating the efficacy of this intervention [5]. Participants at facilities implementing TextIT had the option to respond to text messages, call, or send inquiry text messages to a designated clinic phone. One nurse was assigned to respond to all calls and text messages. During the intervention period, if delivery had not occurred at the facility, the TextIT nurse called intervention recipients weekly, beginning at 37 weeks’ gestation, to determine whether they had delivered elsewhere. These calls continued until delivery status was ascertained. For both intervention and control periods, if no record of delivery existed at the facility at the end of the follow-up period, clinic staff obtained this information at the first postnatal contact with the mother (for women who visited the clinic). Primary outcome measures were infant HIV virologic testing—defined as obtaining a dried blood spot sample for HIV polymerase chain reaction (PCR) testing within 8 weeks after birth—and the proportion of women retained in postpartum care—defined as documented return for at least one visit at the PMTCT, postnatal, or general HIV care clinic within 8 weeks after delivery. We assessed infant HIV testing by abstracting information from medical records, including the HEI follow-up register, HEI card, and laboratory registers. We assessed maternal postpartum retention by abstracting information from patient charts and clinic records. As part of routine care, women who failed to attend clinic appointments were traced to attempt to reengage them in care. We relied on this well-established tracing program to ascertain birth and other outcomes for women who failed to return to clinic [9,10]. Women who could not be contacted were considered to have failed to obtain HIV testing for their infants. Women who voluntarily transferred out to facilities not supported by FACES were excluded from analysis, because we had no mechanism to trace them outside the FACES network. For pregnancies resulting in more than one infant, our analysis was restricted to the firstborn infant. Infants of women who died during follow-up, stillbirths (fetal loss after 28 weeks of pregnancy) and infant deaths before 8 weeks were excluded from the analysis. The primary outcome for sample size determination was infant HIV virologic testing. This outcome was binary and measured at the individual level. We estimated a relative risk comparing the proportion of infants who were tested for HIV in the intervention period compared to proportion tested within the control period. Power was determined using the method described by Hussey and Hughes [11]. We assumed a coefficient of variation between clusters of 0.25 and calculated the power for a two-tailed test with α = 0.05. Assuming that 40% of infants in the control group would undergo HIV testing (based on observed testing rates within the Nyanza region prior to the TextIT study), with 20 clusters and 2 time points, we estimated that a harmonic mean of 57 individuals per cluster per time interval would achieve ≥90% power to detect an increase to 53% tested or greater in the intervention group. Presuming 10% loss to follow-up, we proposed a final sample size of 2,508 women at 20 clinics overall. All analyses were performed using Stata software (StataCorp, College Station, TX). We described participant characteristics at baseline in control and intervention groups. Inferential analysis for the primary outcomes followed the intent-to-treat principle. The predictor of interest was treatment assignment at the clinic/cluster level (i.e., intervention versus control clinics). Time period and stratum of randomization were included as covariates [12]. Our analysis used modified Poisson regression with robust variance estimation to estimate the relative risk and 95% confidence intervals (CIs). We applied generalized estimating equations on individual-level data with working exchangeable correlation structure to account for clustering by site [13]. All tests were two-sided and conducted at the 5% significance level. We conducted secondary analyses to better understand potential changes in the effect of the intervention over time. We also explored whether adjustment for time period resulted in smaller effect sizes due to infant testing and maternal retention increasing over time, independent of the intervention. First, we analyzed the intervention effect moving from period 1 to period 2 in the 10 control clusters that were not assigned to begin receiving the intervention until period 2 (before-after analysis). Second, we analyzed the intervention effect in period 1 alone, comparing infant testing and maternal retention between intervention and control clusters (equivalent to analysis of a parallel cluster randomized study). Finally, we analyzed the change in outcome proportions between period 1 and period 2 in the clusters assigned to begin receiving the intervention in period 1 (intervention clusters). This analysis would allow us to identify a change in outcomes over time in the presence of a consistent intervention, which could be due to a secular trend or a time-dependent change in the intervention effect. County health officials provided written permission for the health facilities in their respective jurisdictions to take part in the study. All text message recipients provided verbal informed consent to receive the messages. For women at health facilities not implementing the TextIT intervention, we had approval from the relevant institutional review boards to collect study data from clinical data that were already being captured routinely for evaluation without obtaining individual-level consent. Ethical approval was obtained from the Kenya Medical Research Institute’s Scientific and Ethics Review Unit, the University of Washington’s Human Subjects Division, and UCSF’s Committee on Human Research. The trial is registered in ClinicalTrials.gov ({"type":"clinical-trial","attrs":{"text":"NCT02350140","term_id":"NCT02350140"}}NCT02350140).