Multimodal, Ancestry-Aware Machine Learning for Equitable Risk Prediction in Admixed Populations
Date Issued
August 2026
Author(s)
Tahmin, Nahian
The University of Tennessee
Advisor(s)
Anahita Khojandi
Additional Advisor(s)
Robert Lowell Davis
Tesfaye B. Mersha
Vincenza Colonna
Abstract
Biomedical and genomic data are generated at high volume and dimensionality,
making machine learning (ML) essential for extracting signal at that scale. Disease
risk is multifactorial, and in admixed populations it reflects interacting biological,
ancestral, and social factors: genotype captures allelic variation, local ancestry (LA)
identifies the ancestral background of genomic segments, and social drivers of health
(SDOH) represent environmental and structural exposures. Conventional genomic
analyses often use genotype alone or treat ancestry globally, leaving this structure
incompletely modeled. This dissertation examines whether integrating genotype, LA,
and SDOH within interpretable ML frameworks supports accurate and equitable
disease-risk prediction while making each modality’s contribution identifiable and
interpretable. Inhaled-corticosteroid (ICS) response in African American asthma
cohorts is the case study. The first study develops a stacking framework integrating
single-nucleotide polymorphisms (SNPs) and inferred LA on whole-exome data from
the pediatric Biorepository and Integrative Genomics Initiative cohort. Regularized
feature selection, logistic regression, and random forests are combined. LA alone
performs comparably to genotype alone, while integration increases cross-validated
AUC to 0.729 (p = 0.005). SHAP analysis shows that the most important SNP and
LA loci are largely non-overlapping, indicating complementary signal. The second
study introduces X-Admix, an interpretable framework integrating genotype, LA, and
SDOH through structured pairwise cross-attention, gated fusion, and a downstream
random-forest classifier. Per-stream ablation distinguishes predictive contribution
from cross-modal representation. On the pediatric cohort, X-Admix predicts ICS
non-response with a mean AUC of 0.771, whereas classical interaction baselines
perform near chance. Removing the genotype–ancestry stream leaves accuracy
unchanged but alters the identified interactions, demonstrating that a modality’s
contribution to representation can differ from its predictive performance. This
predictive pattern is replicated in an independent All of Us adult cohort. The third
study examines how genotype and LA should be represented for a pretrained
language model. In the All of Us cohort, compact token and embedding-injection
encodings outperform verbose natural-language serialization while using several-fold
fewer tokens, showing that encoding determines how much genomic information
a model can process. Together, these studies provide ancestry-aware, interpretable
frameworks for risk prediction in populations historically underserved by genomic
medicine, emphasizing accuracy and accountable reasoning.
Degree
Doctor of Philosophy
Major
Data Science and Engineering
Embargo Date
August 1, 2029