About

David Jongwon Lee is a Boas assistant professor at Northwestern University in the Department of Mathematics. His research interests include homotopy theory, with a focus on topics such as topological Hochschild homology, algebraic K-theory, and the p-local homotopy type of spectra. He has contributed to the understanding of the Hahn-Wilson conjecture, the properties of truncated Brown-Peterson spectra, and the K-theory of the K(1)-local sphere, among other areas. He completed his PhD at MIT under the supervision of Jeremy Hahn and has since engaged in various research projects and collaborations, resulting in publications that address complex problems in algebraic topology and related fields. In addition to his research, David Lee is involved in teaching courses at Northwestern University, covering topics from calculus to algebraic topology, and has delivered talks at numerous conferences and seminars, further disseminating his work and expertise in the mathematical community.

Research topics

• Medicine
• Internal medicine
• Cardiology
• Artificial Intelligence
• Computer Science
• Biology
• Nuclear medicine
• Immunology
• Computer vision
• Radiology
• Genetics
• Biochemistry
• Cell biology

Selected publications

• Branch retinal artery occlusion time to presentation and diagnosis: a retrospective review

  International Journal of Retina and Vitreous · 2026-02-21

  articleOpen access

  Branch retinal artery occlusion (BRAO) is an acute, vision-threatening condition that often signals underlying systemic vascular disease and the need for urgent vascular risk assessment and mitigation. Although prompt evaluation is critical for accurate diagnosis and early identification of modifiable vascular risk factors, patterns of clinical presentation remain poorly characterized. This retrospective study included BRAO patients who presented to Kaiser Permanente Northern California (KPNC) within 30 days of symptom onset from 2014 to 2023. Demographic data, symptom timing, healthcare contact, and ophthalmologic evaluation were collected. The type of contact was categorized as eye care provider, call center, emergency department/urgent care, or other. Delays in presentation were defined as time from symptoms to initial contact and were analyzed across subgroups. From 2014 to 2023, 760 patients were diagnosed with acute BRAO. Mean age of the study population was 70.2 ± 12.6 years, and 330 (43.4%) were female. Initial contact most commonly occurred with eye care providers (370; 48.7%), followed by call centers (251; 33.0%), emergency/urgent care (73; 9.6%), and other providers (66; 8.7%). Only 219 (28.8%) presented within 1 day of symptom onset, while 97 (12.8%) presented after 8 days. The majority (541; 71.2%) were not evaluated by an eye care provider until > 24 h after symptom onset. Among the 153 (20.1%) with known time of symptom onset, median delay to any healthcare contact was 4.0 h. 80 (52.3%) patients presented within 4.5 h, but just 5 (6.3%) were seen by an eye care provider within that window. In a large, multi-center, community population, delays in BRAO care were common. Although no Level 1 evidence currently exists to support BRAO treatment, success of future therapies will likely require early administration. Prompt evaluation is also critical to enable risk assessment and mitigation. Most patients in this cohort presented in a delayed fashion, underscoring the need for public education and workflow improvements to support earlier recognition and care. Patients with branch retinal artery occlusion often experience delays in care. Most are not evaluated promptly by eye care providers, which may preclude timely diagnosis, stroke prevention, and potential future therapies.

  PublisherDOI

• Development of Scalable and Commercial Synthesis for GS-441524

  Organic Process Research & Development · 2026-03-31

  article

  GS-441524 is a broad-spectrum, direct-acting antiviral, and its structural simplicity and oral bioavailability make it a promising therapeutic candidate. It is the parent nucleoside of remdesivir and the active metabolite of obeldesivir. However, the lack of a scalable, impurity-controlled, and regulatory-compliant synthesis has hindered its pharmaceutical development. This study presents an optimized, high-yielding, and commercial synthesis for GS-441524, demonstrating scalability up to multikilogram production. The key steps include a Lewis acid lanthanide salt-promoted glycosylation reaction and a diastereoselective continuous flow cyanation step with a 4-stream system, enhancing safety and reproducibility while achieving up to 94:6 diastereoselectivity and consistent yields at >60 kg scale. Furthermore, the development of an effective crystallization process controls the impurities to below 0.1A%, including potential genotoxic impurities (GTIs), during the isolation of the intermediate GS-441524 HCl salt. Additionally, two polymorphs were identified for GS-441524, with the most stable polymorph isolated using an in situ massive seed generation technique, effectively generating the desired target polymorph for the API. The process was demonstrated on several tens of kilograms with a 42% overall yield starting from 2, confirming its robustness and scalability, and provides a comprehensive and practical synthetic route for the manufacturing of GS-441524 as a drug substance compliant with regulatory guidelines.

  PublisherDOI

• Cardiovascular MRI–based Biventricular Perfusion Assessment in Two Patients with Chronic Thromboembolic Pulmonary Hypertension Undergoing Pulmonary Thromboendarterectomy

  Radiology Cardiothoracic Imaging · 2025-11-20

  articleOpen access

  Chronic thromboembolic pulmonary hypertension (CTEPH) can lead to right ventricular (RV) ischemia and dysfunction due to chronic pulmonary artery obstruction and increased afterload. While cardiovascular MRI (CMR) enables noninvasive assessment of myocardial perfusion, its role in CTEPH remains unclear. The authors report adenosine stress perfusion CMR findings from two patients with CTEPH before and after pulmonary thromboendarterectomy (PTE). Both showed reduced biventricular perfusion before PTE; one demonstrated post-PTE improvement. Perfusion findings aligned with invasive hemodynamics, suggesting that CMR-derived myocardial perfusion reserve may serve as a valuable tool for assessing treatment response and RV pathophysiologic characteristics in CTEPH. Keywords: Cardiac, Pulmonary Arteries, Chronic Thromboembolic Pulmonary Hypertension, Pulmonary Thromboendarterectomy, Quantitative Perfusion Cardiovascular MRI, Myocardial Blood Flow, Myocardial Perfusion Reserve Supplemental material is available for this article. © RSNA, 2025

  PublisherOA PDFDOI

• Team ACK at SemEval-2025 Task 2: Beyond Word-for-Word Machine Translation for English-Korean Pairs

  ArXiv.org · 2025-04-29

  preprintOpen access1st authorCorresponding

  Translating knowledge-intensive and entity-rich text between English and Korean requires transcreation to preserve language-specific and cultural nuances beyond literal, phonetic or word-for-word conversion. We evaluate 13 models (LLMs and MT models) using automatic metrics and human assessment by bilingual annotators. Our findings show LLMs outperform traditional MT systems but struggle with entity translation requiring cultural adaptation. By constructing an error taxonomy, we identify incorrect responses and entity name errors as key issues, with performance varying by entity type and popularity level. This work exposes gaps in automatic evaluation metrics and hope to enable future work in completing culturally-nuanced machine translation.

  PublisherOA PDFDOI

• Rewind and Render: Towards Factually Accurate Text-to-Video Generation with Distilled Knowledge Retrieval

  Proceedings of the AAAI Conference on Artificial Intelligence · 2025-04-11

  articleOpen access1st authorCorresponding

  Text-to-Video (T2V) models, despite recent advancements, struggle with factual accuracy, especially for knowledge-dense content. We introduce FACT-V (Factual Accuracy in Content Translation to Video), a system integrating multi-source knowledge retrieval into T2V pipelines. FACT-V offers two key benefits: i) improved factual accuracy of generated videos through dynamically retrieved information, and ii) increased interpretability by providing users with the augmented prompt information. A preliminary evaluation demonstrates the potential of knowledge-augmented approaches in improving the accuracy and reliability of T2V systems, particularly for entity-specific or time-sensitive prompts.

  PublisherOA PDFDOI

• Systematic Review on Large Language Models in Orthopaedic Surgery

  Journal of Clinical Medicine · 2025-08-20 · 3 citations

  reviewOpen accessSenior author

  Background/Objectives: Since ChatGPT was released in 2022, many Large Language Models (LLM) have been developed, showing potential to expand the field of orthopaedic surgery. This is the first systematic review looking at the current state of research of LLMs in orthopaedic surgery. The aim of this study is to identify which LLMs are researched, assess their functionalities, and evaluate their quality of results. Methods: The systematic review was conducted using PubMed, Embase, and Cochrane Library databases in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Results: A total of 60 studies were included in the final review, all of which included ChatGPT versions 3.0 or 4.0. There were five studies that included Bard and one article each for Perplexity AI and Bing. Most studies assessed orthopaedic assessment questions (23 studies) and their ability to correctly answer free ended questions (31 studies). Outcome measures used to assess the accuracy of LLMs in most of the included studies were the percentage of correct answers on multiple-choice questions or expert-graded consensus to open-ended responses. The accuracy of ChatGPT 4.0 in orthopaedic assessment questions ranged from 47.2 to 73.6% without images, and 35.7–65.85% with images. The accuracy of ChatGPT 3.5 was 29.4–55.8% without images and 22.4–46.34% with images. The accuracy of Bard ranged from 49.8 to 58%. Orthopaedic residents consistently scored better than LLMs in the range of 74.2–75.3%. Conclusions: ChatGPT 4 showed significant improvement over ChatGPT 3.5 in answering orthopaedic assessment questions. When comparing performances of orthopaedic residents to LLMs, orthopaedic residents scored higher overall. There remains significant opportunity for development of LLM performance on orthopaedic assessments as well as image-based analysis and clinical documentation.

  PublisherOA PDFDOI

• Abstract 4369181: Extracellular Volume and T1 Mapping by Cardiac MRI Predict Outcomes in HFpEF: A Systematic Review and Meta-Analysis

  Circulation · 2025-11-03

  articleSenior author

  Background: Heart failure with preserved ejection fraction (HFpEF) accounts for nearly half of all cases of heart failure (HF) and carries a similar burden as heart failure with reduced ejection fraction (HFrEF). While outcomes in HFrEF have improved with evidence-based therapies, adverse events in HFpEF remain unchanged. Imaging biomarkers such as extracellular volume fraction (ECV) and native T1 relaxation time, assessed by cardiac magnetic resonance imaging (MRI), may provide valuable prognostic information. We aimed to systematically evaluate the clinical utility of MRI-derived native T1 and ECV in risk stratification of patients with HFpEF using updated data. Methods: A systematic search of PubMed, Embase, Web of Science, and Scopus was conducted in March 2025 per PRISMA guidelines. Studies reporting associations between MRI-derived ECV or native T1 and clinical outcomes in patients with HFpEF were included. Random-effects meta-analyses were used to pool unadjusted and adjusted hazard ratios (HRs) with 95% confidence intervals (CIs) for a 1-unit increase in ECV and for comparisons of high vs. low T1 and ECV values based on cut-offs defined in each of the studies in relation to adverse outcomes (all-cause mortality, HF hospitalization, and cardiac mortality). Results: Sixteen studies with a total of 1276 patients with HFpEF were included. Mean age was 65.8±11.0 years, and 50% were male. After excluding overlapping cohorts, higher ECV was significantly linked with a higher risk of adverse events in unadjusted models (HR 1.14, 95% CI 1.09–1.19, p < 0.001, I2 = 46%). ECV was also linked to higher all-cause mortality and HF hospitalization (HR 1.19, 95% CI 1.06–1.33, p < 0.001). Associations remained significant after adjustment for clinical covariates in the studies (aHR 1.09, 95% CI 1.05–1.13, p < 0.001, I2 = 41%). In the comparison of high vs. low ECV, patients with higher ECV tend to have significantly higher rates of adverse events (aHR, 1.71, 95% CI 1.25-2.34, p < 0.001, I2 = 0%). Furthermore, patients with high native T1 values had an increased risk of adverse events, compared to low T1 group (aHR 1.74, 95% CI 1.21–2.50, p < 0.001, I2 = 0%). Conclusion: MRI-derived ECV and native T1 are clinically relevant predictors of adverse outcomes in HFpEF. By further large-scale studies conducted in future, these imaging markers may help guide risk stratification and personalize treatment decisions in this challenging patient population.

  PublisherDOI

• Highly Accelerated Real-time Cine MRI Pulse Sequence for Cardiac Implantable Electronic Devices and Arrhythmias

  Radiology Cardiothoracic Imaging · 2025-11-13

  articleOpen access

  A 32-fold accelerated real-time cine MRI sequence (radial sampling) demonstrated better temporal resolution and image quality than a 16-fold sequence (Cartesian sampling) in patients with cardiac implantable electronic devices and arrhythmias.

  PublisherOA PDFDOI

• AI-Driven Scribble-Based Foundation Model for Left Ventricular Scar Quantification on cardiac MRI

  Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025-09-16

  article

  Motivation: Late Gadolinium Enhancement imaging is the gold standard for assessing myocardial fibrosis, with LV scar volume as a key predictor of major adverse cardiac events. However, manual segmentation is labor-intensive and variable, limiting practical use. Goal(s): Develop an automated deep learning method for accurate LV scar quantification, tackling complex scar appearances and incorporating a scribble GUI for refining scar segmentation. Approach: A foundation model combining MedSAM's representation capabilities with U-Net's localization, enhanced by scribble-based annotations. Results: The deep learning model achieved an average scar Dice score of 0.917 for connected scars and 0.719 for disconnected scars, surpassing existing methods. Impact: Our foundation model offers a significant advancement in automated LV scar assessment, improving reliability, reducing manual workload, and enhancing consistency in clinical cardiac imaging, which can lead to better patient outcomes through timely and accurate diagnosis.

  PublisherDOI

• Benchmarking cell type and gene set annotation by large language models with AnnDictionary

  Nature Communications · 2025-10-28 · 4 citations

  articleOpen access

  We develop an open-source package called AnnDictionary to facilitate the parallel, independent analysis of multiple anndata. AnnDictionary is built on top of LangChain and AnnData and supports all common large language model (LLM) providers. AnnDictionary only requires 1 line of code to configure or switch the LLM backend and it contains numerous multithreading optimizations to support the analysis of many anndata and large anndata. We use AnnDictionary to perform the first benchmarking study of all major LLMs at de novo cell-type annotation. LLMs vary greatly in absolute agreement with manual annotation based on model size. Inter-LLM agreement also varies with model size. We find that LLM annotation of most major cell types to be more than 80-90% accurate, and will maintain a leaderboard of LLM cell type annotation. Furthermore, we benchmark these LLMs at functional annotation of gene sets, and find that Claude 3.5 Sonnet recovers close matches of functional gene set annotations in over 80% of test sets. Cell type labelling in single-cell datasets remains a major bottleneck. Here, the authors present AnnDictionary, an open-source toolkit that enables atlas-scale analysis and provides the first benchmark of LLMs for de novo cell type annotation from marker genes, showing high accuracy at low cost.

  PublisherOA PDFDOI

Recent grants

• NIH Grant R21NS067278

  NIH · $477k · 2014

• Rapid Real-Time Cardiovascular MRI for Detecting Coronary Artery Disease

  NIH · $1.1M · 2017–2022

• Quantitative Detection of Coronary Microvascular Dysfunction in Long COVID Patients using a Comprehensive, Rapid, Free-Breathing Cardiovascular MRI

  NIH · $1.6M · 2023–2027

• Real-time Wideband Cardiac MRI for Patients with a Cardiac Implantable Electronic Device

  NIH · $3.3M · 2020–2025

• Quantitative Detection of Coronary Microvascular Dysfunction in Long COVID Patients using a Comprehensive, Rapid, Free-Breathing Cardiovascular MRI

  NIH · $777k · 2023–2027

Frequent coauthors

• Michael Markl

  74 shared

• James Carr

  Environmental Molecular Sciences Laboratory

  66 shared

• Alan H. Kadish

  University Health System

  61 shared

• Daniel Kim

  59 shared

• Jacqueline Fontugne

  Institut Curie

  58 shared

• Juan Miguel Mosquera

  Weill Cornell Medicine

  57 shared

• Edwin Wu

  Facultad Latinoamericana de Ciencias Sociales

  56 shared

• Mark A. Rubin

  University of Bern

  56 shared

Education

• Ph.D., Homotopy Theory

  Massachusetts Institute of Technology (MIT)