Abstract
Background: Haiti's anemia prevalence in 2019 among women of reproductive age (15-49 years) is 48%, compared to 30% worldwide. Anemia jeopardizes maternal and fetal well-being and negatively impacts children's development after birth. The Haiti Health Initiative (HHI), a nurse-led non-profit organization, provides healthcare access for most rural Haitians in Timo. HHI addresses anemia through free biannual outreach clinics, supplementation of iron-rich vitamins, and education related to nutrition, water, and sanitation.
To assess the potential impact of an anemia prevention program, pregnant or lactating women ages 15-49 were followed longitudinally to evaluate change over time in anemia status. As opposed to a randomized control trial, the observational nature of the study can produce paradoxical findings when not accounting for context of care; therefore, modern causal machine learning methods were used to correct for potential confounding.
Methods: This longitudinal observational cohort study utilizes patient data collected at outreach clinics between 2011 and 2019. Inclusion criteria include (1) any pregnant/lactating woman in the clinic database who (2) attended at least three clinics and (3) had at least one blood hemoglobin test recorded in her patient medical records. Time-varying anemia states were assigned based on hemoglobin levels and pregnancy status.
As there is high potential for confounding in this observational study for both treatment and missingness (censoring) mechanisms, modern data-adaptive causal methods were used to model treatment regimens and account for complex data generation context. Targeted learning (TL) allows for the complete integration of machine learning advances while providing statistical inference for the target parameter(s) of interest. Longitudinal targeted maximum likelihood estimation (LTMLE) is applied to understand the impacts of these interventions when data are longitudinal or patients are followed over time. LTMLE for the Longitudinal Average Treatment Effect and working marginal structural models (MSM) for the trend over time were used to estimate treatment effects.
Results: Between 2011 and 2019, 597 pregnant (n=388) or lactating (n=215) women aged 15-49 attended the HHI clinic and met study inclusion criteria. The overall mean hemoglobin of participants at baseline was 10.87 g/dL, with average hemoglobin of 10.44 g/dL for pregnant women and 11.34 g/dL for non-pregnant women, suggesting mild anemia. Both pregnant (58.0%) and non-pregnant women (63.0%) were more likely to be anemic than non-anemic at the next clinic visit, reflecting a higher prevalence rate than Haiti's national average (48%).
Linear mixed effects with random intercept indicated a modest improvement in anemia status (β = -0.17/ treatment) and in hemoglobin levels (β = 0.21 g/dL/treatment). Multistate modeling suggested improvement in women's anemia category, transitioning from a high-risk category (moderate or severe anemia) to a low-risk category (normal or mild anemia) (transition probability = 0.68 over 5 visits) and staying in the low-risk category once achieved (self-transition probability = 0.68 for 5 visits). Initial results from longitudinal TMLE while adjusting for age, clinic, residence, season, pregnant/lactating status and time since last visit indicated a slight protective effect of having multiple treatments (Figure 1, β = 0.18 g/dL/ treatment, p=0.035); the results seems to be sensitive to the severity level of anemia, with a more protective effect in severe anemia.
Discussion: Longitudinal patient data in low- and middle-income countries, such as Haiti, are scarce but essential to improving healthcare access and quality for vulnerable populations. Modest hemoglobin increases suggest the need for nurses to develop innovative and culturally appropriate approaches to deliver evidence-based interventions to hard-to-reach populations. For example, anemia interventions should consider integrating iron-rich vitamins with vitamin C supplementation to elicit a greater hemoglobin increase. Future analysis should collect prospective qualitative data to understand ecological factors influencing quantitative results.
Figure 1: Longitudinal TMLE via Marginal Structural Model for additional treatment/visit (getting vitamins: MVI/PNV/Vitamin A) on mean counterfactual anemia episodes (psi) (p-value via tmle = 0.0355).
Disclosures
No relevant conflicts of interest to declare.
