The University of Arizona Data Science · Statistics
Tucson, Arizona

Profile · 2026

Clinical medicine, public health & genomics.

I build statistical and computational methods for biomedical, population, and complex real-world data.

Huanqi Wu 吴欢祺 · FoonKei Ng · 吳歡祺

The University of Arizona Department of Mathematics Data Science · Statistics

Curriculum Vitae Google Scholar GitHub

Huanqi Lionel Wu · Tucson, Arizona

Selected work ↓

02 / Selected Work

Four projects.
Four ways of turning data into evidence.

Core Capabilities Research → Model → Evidence

01

Statistical & ML

Statistical Modeling · Machine Learning · High-Dimensional Data · Longitudinal & Panel Models

02

Biomedical Data

Computational Genomics · Epidemiology & Real-World Evidence · Mixed-Methods Research

03

Research Stack

R · Python · Stata · Git/GitHub · Data Visualization · Research Automation

Project 01 Positive-Unlabeled Learning

PU-
BOOST

PU-Boost treats potentially undiagnosed hypertension as a positive-unlabeled learning problem. Elastic Net first performs global P/U stratification, a Random Forest then refines the uncertain region, and persistently ambiguous samples are rejected before final supervised learning on a reconstructed balanced set.

81.72% Sensitivity

68.23% Balanced accuracy

729 False negatives

25,883 TRAINING SAMPLES → 16,034 RECONSTRUCTED SAMPLES · 9,849 REJECTED
VS. STANDARD XGBOOST · −508 FALSE NEGATIVES (−41.1%) · +12.75 PP SENSITIVITY

PU Learning Elastic Net Random Forest Screening

Manuscript ↗ GitHub ↗ Figures ↗

Performance / cohort design / method / explainability PU / 01

PU-Boost Figure 7 showing false positives and false negatives across models, plus specificity versus sensitivity — Figure 7 · False negatives and false positives · sensitivity vs. specificity 01

PU-Boost cohort figure showing 47,841 recruited adults, analytic cohort, and split into training, validation, and test sets — 47,841 recruited · cohort architecture 02

Two-stage reconstruction

PU-Boost stage 1 figure showing P versus U setup and the distribution of stage 1 scores — Stage 1 · N1 2,144 · U1 18,426 · P1 5,313 03

PU-Boost stage 2 figure showing random forest refinement within the uncertain region and final reconstructed labels — Stage 2 · N2 5,873 · reject 9,849 · P2 2,704 04

Explainability

PU-Boost SHAP summary plot showing global feature importance — SHAP summary · global feature importance 05

PU-Boost SHAP dependence plots for age, BMI, waist circumference, and maternal hypertension history — Age · BMI · waist circumference · family history 06

Project 02 IgAN Clinical Research Series

Clinical Nephrology / Real-World Cohort

BUDESONIDE

Part of a wider IgA nephropathy clinical research series, this study follows patients receiving targeted-release budesonide in routine care. Across nine months, we examined changes in proteinuria and kidney function, variation in individual responses, and the context of concomitant treatment. The cohort showed a substantial reduction in proteinuria, while kidney function remained broadly stable at month 9.

36 Patients · updated manuscript cohort

−62.0% Modeled proteinuria · month 9 vs baseline

0–9 mo Follow-up · four assessment points

IgA Nephropathy Targeted-Release Budesonide Proteinuria & eGFR Real-World Evidence

Manuscript Figure ↗ GitHub ↗

Clinical response / patient variation / treatment context NEPHROLOGY / 02

Latest-manuscript longitudinal mixed-effects estimates: proteinuria geometric means decreased through 9 months and eGFR estimated marginal means rose at 3 and 6 months, with 95 percent confidence intervals — Manuscript Fig. 2 · LMM estimates · 95% CI · 0–9 months 01

July exploratory R analysis: modeled proteinuria trajectories at 0, 3, 6 and 9 months, split by whether concomitant medication classes changed; both groups show decreases and the medication-by-time interaction was not significant — R reanalysis · medication × time · 95% CI 02

Budesonide chemical structural formula extracted from the July presentation, with an asterisk marking its C-22 stereocenter; budesonide is formulated as a mixture of 22R and 22S epimers, molecular formula C25H34O6 — Drug molecule · 22R / 22S epimers · PPT 03

Original July presentation waterfall plot of 32 patients with month-nine follow-up, sorted by individual proteinuria reduction; categories distinguish at least 50 percent improvement, 30 to 50 percent improvement, smaller reductions, and worsening — PPT · patient-level waterfall · n=32 04

Exploratory R violin and box plot of proteinuria reduction by IgA nephropathy disease duration, with individual patient points and an unadjusted Mann–Whitney P value of 0.013 in the July working analysis — R reanalysis · duration · violin / IQR 05

Original July presentation grouped horizontal bar chart comparing baseline and month-nine exposure to nine concomitant medication classes including ARB or ACE inhibitor, SGLT2 inhibitor, nonsteroidal MRA, ERA, HCQ, hormone therapy, MMF, TW and telitacicept — PPT · medication exposure · 9 classes 06

Project 03 Computational Genomics

Concept / In Development

cfMethyl
Former

A concept-stage transformer framework for cfDNA methylation representation learning, extending read-level sequence context toward sample-level aggregation and prediction.

500 bp Read context

Read Methylation-aware encoding

Sample Multi-read aggregation

Transformer cfDNA Methylation Representation Learning

Concept Note ↗ GitHub ↗

Architecture / representation learning GENOMICS / 03

GSE149438 plasma cfDNA methylation UMAP by clinical diagnosis, disease stage, sequencing coverage, and CpG feature completeness — GSE149438 · 300 plasma samples · 2,500 CpGs · PCA + UMAP 04

cfMethylFormer architecture — Methylation-aware encoder → transformer → aggregation 01

cfDNA methylation read encoding or attention visualization — Sequence context · masked methylation sites 02

cfMethylFormer sample-level aggregation — Read embeddings → sample-level prediction 03

Project 04 Epidemiology · Nonlinear Statistical Modeling

Cross-sectional Study / Manuscript

Obesity
Paradox

Does greater body size correspond to fewer depressive or anxiety symptoms? In a cross-sectional sample of 43,079 Chinese adults aged 45+, this study tested six anthropometric measures using sex-stratified logistic regression and restricted cubic splines. The results challenge a universal “jolly fat” effect: selected measures showed inverse associations in men, whereas women's anxiety associations were often positive and, for BMI and WHR, nonlinear.

43,079 Eligible adults aged 45+ · model N varies

7 / 24 Primary associations significant after FDR

4 knots 5th · 35th · 65th · 95th percentiles

Restricted Cubic Splines Sex-Stratified Models Depression & Anxiety FDR Correction

Research Repository ↗ RCS Figure ↗

Sex-specific patterns in adiposity & mental health EPIDEMIOLOGY / 04

Four restricted cubic spline plots for BMI and waist-to-hip ratio versus depressive and anxiety symptoms, showing female and male odds ratios, 95 percent confidence bands and exposure distributions — Figure 2 · RCS odds ratios relative to the median · 95% CI 01

Sex-stratified forest plot for six anthropometric measures, presenting adjusted odds ratios and 95 percent confidence intervals for depressive and anxiety symptoms — Figure 1 · M2 logistic OR per 1-SD increase 02

Covariate-standardized predicted probabilities of depressive and anxiety symptoms across BMI and waist-to-hip ratio from separate linear-exposure M2 logistic models — Figure 3 · Marginal probabilities from linear M2 models 03

07 / Additional Research

Beyond the four.
Population health, applied statistics & policy.

Additional work applying statistical modeling to cardiovascular risk, digital health, healthcare-service utilization and regional economic structure.

01 · Population Health / Statistical Modeling ↗

Systolic Blood Pressure Among Middle-Aged & Older Chinese Adults

Models physiological and lifestyle correlates of systolic blood pressure using LASSO, transformations and nonlinear regression.

43,088 participants LASSO · Regression · Risk Modeling

02 · Public Health / Health Policy ↗

Healthcare-Seeking Behavior in China

Examines how socioeconomic and clinical factors shape choices between primary care and Grade-A hospitals in China.

Frontiers in Public Health · 2024 9,098 person-times · Published Research

03 · Digital Health / Mixed Methods Manuscript

Digital Therapeutics for Dementia Care in China

Mixed-methods study of clinician adoption, barriers and implementation strategies for DTx in dementia care.

388 respondents · 22 interviews 196 cognitive centers · Ordinal Logistic · Thematic Analysis

04 · Econometrics / Public Policy ↗

Local Government Financing & Industrial Structure

Panel-data study of how LGFV versus statutory debt structures relate to provincial industrial structure.

2016–2025 panel TWFE · Panel Data · Data-Driven Index

08 / Experience & Research

Industry experience.
From healthcare investment to computational oncology.

01

Genetron Health

Research & Development · Bioinformatics Internship

Worked on AI methods for liquid-biopsy cancer early detection, including cfDNA sequence representation and classification, and developed a Python workflow for automated NCBI GEO retrieval, screening and evidence organization.

↗

02

Inn Capital

约印医疗基金 · Healthcare Investment Internship

Healthcare investment research in an industry environment spanning biotechnology, medical devices, in-vitro diagnostics and intelligent medicine, with emphasis on clinically driven innovation.

↗

Selected Publication ↗ Google Scholar ↗ Curriculum Vitae ↗ LinkedIn ↗

Huanqi Wu Data Science · Biomedical Research

09 / Explore

HUANQI
WU.

Data Science · Statistics
Computational Research

The University of Arizona
Department of Mathematics
Tucson, Arizona

Curriculum Vitae ↗ Google Scholar ↗ GitHub ↗ ResearchGate ↗ LinkedIn ↗ huanqiwu@arizona.edu ↗

Clinical medicine, public health & genomics.

Four projects. Four ways of turning data into evidence.

PU- BOOST