Research

Research Program

My work spans three interconnected areas: statistical methodology for fairness and equity in clinical AI, missing data and survival analysis methods, and public health data science in Sub-Saharan Africa, alongside a developing interest in network and epidemic modeling. All projects use reproducible workflows in R and Quarto.

1. Algorithmic Fairness in Clinical AI

The V(S) Vulnerability-Based Bias Detection Framework

Clinical prediction models frequently perform well on aggregate metrics while failing specific subpopulations — elderly patients, racial minorities, and groups with limited data representation. I’ve developed the V(S) framework, a vulnerability-indexed approach that integrates multi-attribute utility theory and game-theoretic threshold optimization to detect and quantify this kind of subgroup bias.

Validated on two independent datasets — the UNM Trauma Registry (n = 3,132) and MIMIC-III ICU Mortality Data (n = 46,476) — the framework surfaced a striking failure mode: Random Forest showed sharp performance degradation in elderly patients (AUC 0.636 vs. 0.909 for logistic regression; p = 0.031) that standard explainability tools like SHAP missed.

Target: Statistics in Medicine

2. Missing Data Methods

MNAR Sensitivity Analysis and Tipping-Point Frameworks

Missing data are pervasive in clinical and public health research. This work focuses on:

Multiple imputation under MAR and MNAR assumptions
Tipping-point analysis for how sensitive conclusions are to departures from MAR
Two-stage multiple imputation for complex survey data, including DHS missing-hemoglobin outcomes

Applied to clinical trial data (target: BMC Medical Research Methodology) and to childhood anemia analyses using Demographic and Health Survey (DHS) data.

3. Public Health Data Science in Sub-Saharan Africa & Beyond

PHEM Capacity Building — Africa CDC

A systematic landscape analysis of Public Health Emergency Management (PHEM) competency-based training programs across African Member States, PRISMA 2020-aligned, examining training gaps, workforce capacity, and implementation fidelity.

Childhood Anemia in Sub-Saharan Africa

Prevalence and predictors of anemia among children under five across Ghana, Nigeria, and Tanzania, using pooled DHS data (n ≈ 47,919). Methods: Generalized Estimating Equations (GEE) for within-country clustering, two-stage multiple imputation for missing hemoglobin values, and country-stratified and pooled models with interaction terms.

APHA 2026 — accepted for oral presentation

Domestic HIV Surveillance — Arizona Department of Health Services

Alongside the Africa-focused work, I build Power BI dashboards and a surveillance-to-national-data-dictionary crosswalk for Arizona’s CDC-funded HIV Testing Program, covering testing volume, demographics, care linkage, and PrEP cascade metrics.

4. Network Analysis and Epidemic Modeling

Variational EM for ERGM Mixture Models

Applying variational EM algorithms for Exponential Random Graph Model (ERGM) mixture models across three datasets — the Add Health Study, the Copenhagen Networks Study, and the Colorado Springs HIV transmission network — to examine latent community structure and disease spread in social networks.

Bayesian Nonparametric Epidemic Modeling

A developing interest in Gaussian process priors for stochastic epidemic modeling, particularly for settings with sparse or irregularly observed outbreak data common in low-resource health systems.

Software & Reproducibility

Code and reproducible workflows are on GitHub. I contribute to TSGap, an open-source Python library for composable time-series missingness simulation, separating missingness mechanisms (MCAR/MAR/MNAR) from patterns for imputation-method benchmarking.

Key tools: mice, survival, randomForest, xgboost, survey, geepack, brms, ggplot2, R Markdown, Quarto, LaTeX.