David Burstein, PhD

Curriculum Vitae

Education

PhD in Mathematics (2011-2016)

University of Pittsburgh
BS in Mathematics (2007-2011)

University of Maryland

Summa Cum Laude and University Medal Finalist

Work Experience

Assistant Professor (2024 - Present)

Icahn School of Medicine at Mount Sinai, Department of Psychiatry
- Supervise and collaborate with researchers on applying machine learning approaches to electronic health record data for genetic analyses.
Principal Investigator and Data Scientist (2024 - Present)

James J. Peters VA Medical Center
- Secured $1.1 million in funding to leverage statistical approaches that mitigate genomic research disparities in the Million Veteran Program.
Senior Data Scientist (2019 - 2024)

Icahn School of Medicine at Mount Sinai
- Leveraged machine learning techniques to identify eligible prescriptions 340b pricing.
  The pipeline generated $550k revenue for Q4 2020 alone and a projected $2 million in revenue for 2021.
Senior Data Scientist (2018 - 2019)

Fifth Third Bank
- Collaborated with stakeholders to leverage causal inference to measure the success of multimillion dollar investments in bank products and services.
Visiting Assistant Professor (2016 - 2018)

Swarthmore College, Department of Mathematics and Statistics
- Taught introductory and upper-level courses in Calculus, Probability and Statistics. Supervised teaching assistants and graders.

Modeling diagnostic dropout boosts precision for schizophrenia

Dx dropout AUPRC = 73.4%, compared to phecode AUPRC = 62.7%

We assess our XGBoost diagnosis (dx) dropout model for predicting schizophrenia cases from electronic health record data, benchmarked against chart-review labels. Patients who stopped receiving schizophrenia diagnoses were likely misdiagnosed. We compare our approach with the phecode methodology, which prioritizes cases by diagnosis count. Model approaches are highlighted by different colors.

Model-derived binge-eating disorder phenotype identifies novel risk loci

Top predictors of the model, sized and colored by −log₁₀ p

Binge-eating disorder is heavily underdiagnosed: only 0.1% of Veterans carry the diagnosis, against an established US prevalence of roughly 3%. We model it from electronic health record data in the Million Veteran Program to recover those missing cases. Genome-wide analysis of the resulting phenotype identified novel risk loci and implicated iron metabolism, published in Nature Genetics.

Selected Publications

Peer-reviewed articles
Genome-wide analysis of a model-derived binge eating disorder phenotype identifies risk loci and implicates iron metabolism

Burstein D, Griffen T, Therrien K,...,Hildebrandt T, Voloudakis G, Roussos P.

Nature Genetics. September 2023
Genome-wide association study of treatment-resistant depression: shared biology with metabolic traits

Kang JE, Castro VM, Ripperger M,... Burstein D...,Walsh CG, Perlis RH, Ruderfer DM.

American Journal of Psychiatry. July 2024
Penetrance and pleiotropy of polygenic risk scores for schizophrenia, bipolar disorder, and depression in the VA health care system

Bigdeli TB, Voloudakis G, Barr PB,... Burstein D...,Harvey PD, Roussos P.

JAMA Psychiatry. September 2022
Preprints
Phecoder: semantic retrieval for auditing and expanding ICD-based phenotypes in EHR biobanks

Bennett JJR, Tomasi S, Gupta S, VA Million Veteran Program, Voloudakis G, Roussos P, Burstein D.

medRxiv. January 2026. Preprint.
Modeling diagnostic code dropout of schizophrenia in electronic health records improves phenotypic data quality and cross-ancestry transferability of polygenic scores

Burstein D, Tomasi S, Venkatesh S, Rizk M, Roussos P, Voloudakis G.

medRxiv. January 2025. Under review at Molecular Psychiatry.

Full publication list on Google Scholar

Selected Talks

Million Veteran Program (MVP) Science Meeting

March 2025, Virtual Talk.
American Society of Human Genetics (ASHG) Annual Meeting

November 2024, Denver, Colorado: USA.
PsycheMERGE Diversity Initiative

February 2024, Virtual Talk.
American Society of Human Genetics (ASHG) Annual Meeting

November 2023, Washington D.C.: USA.
PsycheMERGE Analysis Network Wide Call

March 2023, Virtual Talk.

Featured Projects

Auditable AI-Assisted Phenotyping

Phecoder uses an ensemble of pre-trained text embedding models to encode a free-text phenotype query and ICD code descriptions into a shared semantic space, ranks codes by similarity, and returns a top-k shortlist for expert review. The goal is to make phenotype curation in EHR biobanks more transparent, reproducible, and scalable. Read the preprint. View code.

Inferring eating disorders from EHR data

Binge-eating disorder (BED) is extremely underdiagnosed. Only 0.1% of Veterans have the diagnosis, in contrast to the established prevalence of 3% in the United States. We developed a machine learning model to predict BED from electronic health record (EHR) data to recover these missing cases and power a genome-wide association study of BED, yielding 3 novel genome-wide significant loci. Read the paper.

Predicting Diagnosis Code Dropout

Polygenic score performance after diagnosis dropout modeling

We model the discontinuation or dropout of diagnostic codes to identify diagnoses that suggest diagnostic uncertainty for schizophrenia. By applying our data-driven filter, we achieve a 33% increase in the association effect size when the training and target cohort both share African ancestry. Read the preprint.

Measuring Effect Size Dilution in GWAS

With the advent of genotyped biobanks, genome-wide association studies (GWAS) leverage larger sample sizes and introduce noisier phenotypic definitions. Here, we develop a statistical method, PheMED, that leverages GWAS summary statistics to quantify genome-wide effect size dilution across different phenotypic definitions and cohorts. Read the paper.

David Burstein, PhD

Curriculum Vitae

Featured Projects

Auditable AI-Assisted Phenotyping

Inferring eating disorders from EHR data

Predicting Diagnosis Code Dropout

Measuring Effect Size Dilution in GWAS

Lab Members

Current Members

Alumni

Contact