Selected Projects & Demos
Selected source code and related resources are available on our GitHub. Trained models are available on Hugging Face for research and development use.
Clinical Decision Support
A decision aid portal for cancer-associated venous thromboembolism (VTE), designed to support guideline-concordant discussions between treating oncologists and patients. A companion SQL script for data extraction from Epic Caboodle is available upon request.
Resources:
Publications:
Trial Pre-Screening
A pre-screening tool designed to identify cancer patients at risk for venous thromboembolism (VTE). The current version supports an ongoing randomized controlled trial evaluating statins for VTE prevention. A companion SQL script for data extraction from Epic Caboodle is available upon request.
Resources:
VTE-BERT
A fine-tuned BERT NLP model for detecting VTE events in clinical text.
Resources:
Demo • GitHub • Hugging Face
Publications:
Bleed-BERT
A fine-tuned BERT NLP model for detecting bleeding events in clinical text.
Resources:
NLPMed-Portal
A specialized platform that supports database integration, clinical note preprocessing (including section detection and keyword highlighting), user annotation and adjudication, and NLP model development and testing. It also integrates with AWS to enable database agent deployment and large language model (LLM) capabilities.
Resources:
Publications:
Cosmos Hematology Research Collaborative
We lead a consortium supporting thrombosis, hematology, and oncology researchers across the U.S. and Canada. Our team provides expertise in study design, data science, and biostatistics, leveraging the Epic Cosmos database to drive impactful research. Explore our recent publications.
Publications:
Jama Open (2025) • AJH CAT (2026) • AJH CRB (2026) • JNCI ATE (2026)
Cancer Cohort Database
Cancer Cohort Database is a custom clinical data model designed to support cancer research and longitudinal cohort studies. Inspired by established clinical data warehouse architectures and common data models, it organizes key domains such as demographics, encounters, diagnoses, medications, laboratory results, procedures, imaging, and clinical notes into a research-focused relational structure. The model is designed to simplify cohort development and downstream analytics while preserving clinically meaningful relationships across data domains.
ER Diagram:
Publications:
AJH (2022) • Jama Open (2023) • JREHD (2024) • JTH (2024) • JTH (2025) • JTH (2025) • JTH (2026)
Health-DeID
Health-DeID is an open-source Python library for auditable and reproducible de-identification of clinical text and structured health data. It supports multiple PHI detection approaches, including AWS Comprehend MEdical, Claude models through Amazon Bedrock, and configurable local rules, with flexible policies for redacting, generalizing, data-shifting, or replacing sensitive information. The library provides a web interface, command-line tools, and a Python API, along with resumable processing, optional LLM-based validation, human review, and consistent transformation of both free text and structured fields, making it suitable for preparing clinical datasets for research and machine leaning workflows.
