Proposal Abstract
Marothi Peter Letsoalo 
Danielle Jade Roberts
Nonhlanhla Yende-Zuma
PhDDesk
Focus
Person-centered latent phenotype modeling for individualized severe maternal outcome risk stratification in obstetric HIV under real-world, irregular longitudinal follow-up.
Proposal Abstract
The peripartum period for women living with HIV represents a complex intersection of biological, clinical, and structural stressors that are frequently obscured by population-average modeling techniques. Low- and middle-income countries account for approximately 94% of all maternal deaths globally (World Health Organization 2024). Compared with HIV-negative women, HIV-positive women show approximately two to three times higher odds of adverse pregnancy outcomes and intrauterine death, and nearly two times higher odds of low birth weight (Tukei et al. 2021). Traditional epidemiologic approaches often assume population homogeneity, which fails to account for unobserved heterogeneity across clinical settings in sub-Saharan Africa (Naicker et al. 2022; Serunjogi et al. 2025). This proposal uses high-resolution longitudinal data from a multi-national pregnancy cohort with sites in Durban and Soweto, Kampala, Blantyre and Lilongwe, and Harare, Seke-North, and St Mary’s, and develops a person-centered predictive framework for severe maternal outcomes (Taha et al. 2022).
Maternal risk estimation in obstetric HIV also requires explicit attention to hierarchical structure because participants are nested within sites with different service environments. Population averages can mask correlated stressor pathways that vary by both individual and context (Naicker et al. 2022). The proposal therefore combines latent class methods, causal adjustment, multivariate continuous-time growth modeling, and joint modeling to produce clinically useful risk stratification (Serunjogi et al. 2025).
Aim 1 establishes the methodological baseline through a structured audit of latent trajectory studies in maternal health. Class extraction can be spurious when distributional misspecification is ignored (Bauer and Curran 2003), so this risk is evaluated directly together with reporting quality standards including GRoLTS (Van de Schoot et al. 2017) and SMART-oriented latent class reporting guidance (Lissa et al. 2024). The audit also examines temporal binning in obstetric data, where fixed trimester grouping can distort longitudinal signal under asynchronous visit schedules (Asparouhov and Muthen 2024; Proust-Lima et al. 2023).
Aim 2 identifies latent social risk phenotypes and estimates their association with severe maternal outcomes. A multilevel latent class framework separates individual-level vulnerability from site-level context, and classification uncertainty is handled using BCH-based correction rather than naive class assignment (Asparouhov and Muthen 2014; Bakk and Vermunt 2014). To strengthen causal interpretation, inverse propensity weighting is integrated with latent class modeling so that baseline imbalance is explicitly addressed (Le et al. 2024). Aim 3 then models multivariate physiological burden using continuous-time growth mixture methods with subject-specific timing, avoiding bias from forced interval alignment and preserving irregular antenatal measurement structure (Asparouhov and Muthen 2024; Wickrama et al. 2016).
Aim 4 translates these findings into individualized dynamic prediction. A latent class joint modeling framework combines longitudinal marker history and time-to-event processes to produce updated risk probabilities at clinical landmarks while handling informative dropout (Rizopoulos et al. 2017). Predictive discrimination is complemented by decision-curve analysis to evaluate clinical net benefit in triage and surveillance decisions (Heagerty and Zheng 2005; Vickers and Elkin 2006). Taken together, the dissertation contributes an integrated biostatistical framework for precision maternal epidemiology in obstetric HIV that links methodological rigor, causal inference, and implementable prediction in high-burden settings.
Research Questions
- How can latent trajectory methods be audited and improved so class solutions are methodologically robust in maternal-health research?
- Which latent social risk phenotypes are present in obstetric HIV populations, and how are they associated with severe maternal outcomes?
- How do multivariate physiological burden patterns evolve over irregular antenatal follow-up when modeled in continuous time?
- How can latent class joint modeling support dynamic, individualized prediction of severe maternal outcomes at clinical decision timepoints?
Core Aims
- Audit and strengthen latent trajectory methodology and reporting practice in maternal health research.
- Derive latent social risk phenotypes and estimate associations with severe maternal outcomes using causal adjustment.
- Model dynamic multivariate physiological burden with continuous-time latent growth mixture methods.
- Deliver individualized dynamic risk prediction for obstetric HIV care pathways.