Developing a representation learning framework for high-dimensional genotype data using LD block–based variational autoencoders. Designed to overcome n << d limitations (~1K samples, 339K variants), the approach learns biologically meaningful embeddings for downstream tasks such as patient stratification and drug response modeling.

This framework improves structure discovery compared to traditional LD-pruned PCA (PC1 variance: 13% vs <1%) and is being extended to large-scale cohorts (UK Biobank) to enable scalable genotype representation learning beyond single-variant and PRS-based methods.

Analyzed Olink NPX proteomics data to identify neurocognitive subtype–specific protein signatures in pediatric Long COVID. Applied covariate-adjusted logistic regression across 87 plasma proteins, accounting for cohort heterogeneity and clinical confounding.

Extended the analysis to the UK Biobank adult cohort using PySpark and SQL-based pipelines, enabling scalable integration of large proteomics and phenotype datasets for subtype discovery.

Developed a machine learning–driven GWAS framework to model asthma heterogeneity using PCA-derived clinical endophenotypes. Patients were stratified into five ordinal subgroups (Q1–Q5) spanning mild to severe, atopy-rich disease. These subtypes were integrated into a categorical ANCOVA model, enabling simultaneous association testing across biologically coherent patient groups and improving detection of subtype-specific genetic signals.

Investigated molecular drivers of variable response to inhaled corticosteroids (ICS) in asthma using circulating microRNA (miRNA) profiles. Applied interaction modeling (Exacerbation ~ ICS × miRNA) across two pediatric cohorts (CAMP discovery and GACRS replication) to identify biomarkers associated with treatment resistance.

Identified miR-584-5p as a novel modifier of ICS response, where higher expression was associated with increased exacerbation risk specifically in treated patients, highlighting its potential as a biomarker for steroid resistance.

Developed a Random Forest Classifier to predict COVID-19 infection status in children using a dataset of 2,572 chest X-ray impressions and electronic medical records. By integrating radiological findings, clinical symptoms, and demographic variables, the model achieved an F1-score of 0.79 and an AUC of 0.85.

Developed a framework for testing context-aware mobile applications by modeling event-driven behavior using probabilistic and neural approaches. Combined conditional random fields (CRFs), neural network–based component discovery, and combinatorial optimization to capture dependencies between user actions, system states, and environmental context, enabling systematic exploration of large, constrained event spaces in Android applications.

This work established a foundation for modeling structured, dependent data and scalable search strategies—principles that extend to my current work in genomics and clinical machine learning.