Open to Data Scientist, Real-World Evidence Scientist and industry Research Scientist roles in health tech and pharma. Based in St. Louis, MO; open to relocation and remote.
I own the analysis pipelines and templates a research group at Washington University School of Medicine runs on, and I manage its data. Eighteen researchers have run their analyses through them: two assistant professors, a postdoc, three PhD students, five MPH research assistants, two clinical fellows, two residents and three visiting scholars. I write the statistical methods sections, and often help design the analyses they describe. The last review before submission is mine, and what I am accountable for is that the numbers are right.
- Cited in NCCN guidelines. Lead data analyst on the JNCI 2023 red-flag signs and symptoms study, a matched case-control analysis of 5,075 cases drawn from 113 million US commercial claims beneficiaries, which the NCCN Colorectal Cancer Screening Guidelines cite (v1.2024 through v2.2026) for their recommendation on evaluating alarm signs and symptoms. The only other reference for that recommendation is a meta-analysis that pools it.
- 27 peer-reviewed papers, 1,300+ citations, h-index 17, co-first on three, in venues including Nature Medicine, Nature Genetics, Cell Metabolism and JAMA Oncology.
- Extraction, QC and processing I built myself across commercial claims (MarketScan), Medicare, UK Biobank, All of Us, Our Future Health, NHANES, the Harvard cohorts, ALSPAC, SEER and CDC WONDER, much of it inside controlled-access environments, in R, Python, SQL and SAS.
JM-omp
Rcpp/OpenMP rewrite of the JM R package for joint longitudinal-survival models, built
because upstream JM was too slow to fit a full-size NHS II cohort. Same API; at 100,000
subjects the Weibull-PH fit matches CRAN JM's log-likelihood to the last printed digit,
same iteration count, same convergence flag. The
benchmark table is on simulated cohorts, and
the repo states which methods are not accelerated and which two columns are measured against
different baselines. A collaborator is fitting a real NHS II cohort with it now; I will
quote real-cohort figures here once that lands.
GGIR performance work
Officially listed contributor on GGIR, the
standard accelerometry toolkit, with three defects filed upstream, all three fixed and
credited in GGIR's release notes. 15 further correctness and performance branches, each
benchmarked in isolation with equivalence verified at the output; one submitted, the rest
queued
(roadmap).
Per-branch measurements live in that roadmap and are not quoted here until they are merged.
docker-accelerometer
Reproducible GGIR stack for Docker, Apptainer and HPC. Image reduced 4.34 GB to 1.92 GB,
thread-capped so a scheduler allocation is never silently oversubscribed.
svyrcs
Restricted cubic splines fitted inside a complex survey design. Closes 8 documented
ways a hand-rolled script gets this wrong: Barnard-Rubin df, design-based F, t
quantiles, weighted knots. CI, tests, vignette, MIT.
More: nhanes-accel-1114 and 0306, NHANES accelerometry through a shared GGIR workflow, validated against the NCI SAS reference (97.1% MVPA, adults, unweighted) with comparability limits documented rather than glossed · gcta, COJO variance-ratio filter removed so large-effect pQTLs survive joint analysis.
Both arms built from the same source and run back to back in one job on one physical host;
seven replicates over ten UK Biobank recordings; a branch only counts if it wins every
pair. Equivalence is verified at the output rather than assumed, and the same reading
caught four defects upstream: three in GGIR, and systematic wake
overcalling in asleep on
Idle-Sleep-Mode NHANES recordings, filed with diagnostics, a GGIR comparison and a
proposed flag.
I document negative results too. Intel MKL looked like a clear win and was removed from
docker-accelerometer after measurement showed it indistinguishable from OpenBLAS at the
one thread the image actually uses, while costing 2.4 GB and one hard crash
mode.
That README carries a section called What is deliberately not claimed. I would rather
publish the negative result than carry a number I cannot defend.
The data cannot leave. The methods can.
The pipelines my group runs on live in private repositories, inside controlled-access perimeters: All of Us Controlled Tier, DNAnexus RAP, a health system's OMOP data lake, a cohort cluster holding Controlled Unclassified Information.
- EHR phenotyping on OMOP (SQL on Databricks): concept-set definitions carrying the literal rule, the coverage it produced, and an orphan review of what the rule missed. An inventory of what a definition caught cannot show you an omission.
- Polygenic risk scores at cohort scale (Hail, BigQuery): WGS against array, genome build and allele harmonisation, missingness, variant annotation.
- Multi-cohort harmonisation: pooling the Harvard cohorts, which run on SAS, through variable-level coverage matrices and impossible-entry worksheets a human signs off before anything is combined.
- Wearable phenotypes: multilevel functional PCA over one-minute actigraphy in three cohorts, cosinor and sleep-regularity metrics, physiological circadian modelling.
- Observational-design hygiene: landmark anchoring, multiple imputation with Rubin pooling, negative-control outcomes and empirical calibration.
- Review as a deliverable: verification passes that re-measure numbers already reported, defect-class audits, independent second-reader code review before sign-off.
- Independent replication of a published JAMA Network Open analysis, reproducing its estimates from the source data.
Each repository carries an explicit data red line, and a .gitignore that keeps
controlled-data paths out of Git tracking. The real control is checking what you are about
to commit.
Co-first author on three papers: Gut (2021), metabolic syndrome and early-onset colorectal cancer (260 citations, the most-cited in the record); Journal of Clinical Oncology (2023), clonal hematopoiesis and incident lung cancer; Hepatology Communications (2022), fatty liver disease and gastrointestinal cancers. Second author on Nature Medicine (2026).
Full record on ORCID.
The analysis behind most of those papers predates LLM coding assistants. I build and run coding harnesses now, and they raise throughput without moving the standard: generated code passes the same review, benchmark and equivalence gates as a colleague's, and the final sign-off remains mine.
Languages: R · C++ (Rcpp/OpenMP) · Python · SQL · SAS · Bash
Platforms: All of Us Researcher Workbench / Verily · DNAnexus RAP · Databricks · BigQuery · Hail · Docker · Apptainer/Singularity · LSF · Slurm · GitHub Actions
Cohorts & data systems: UK Biobank · All of Us · Our Future Health · NHANES · NHIS · Nurses' Health Study I/II/3 · HPFS · GUTS I/II · ALSPAC (Avon) · SEER · CDC WONDER
Data: OMOP CDM · EHR (Epic Clarity ETL) · proteomics (Olink) · genomics · microbiome (shotgun metagenomics via bioBakery, 16S via DADA2) · accelerometry · complex survey data
Methods: survival analysis · joint longitudinal–survival models · propensity scores · Mendelian randomization and colocalization · multiple imputation with Rubin pooling · negative-control outcomes and empirical calibration · polygenic risk scores · multilevel functional PCA · survey-weighted regression · restricted cubic splines · statistical genetics