Computational Phenotyping and Unsupervised Subgroup Discovery for Hypermobile Ehlers-Danlos Syndrome in a National EHR Dataset

Presentation Type: Oral Presentation - Student

Click to View Presentation

Presentation Time: 01:30 PM - 01:42 PM

Abstract Keywords: Clinical Decision Support and Care Pathways, Diagnostics, Health Data Science, Precision Medicine, Multi-Omics, and Pharmacology Integration
Primary Track: Big Data for Health


Hypermobile Ehlers-Danlos syndrome (hEDS) is a heterogeneous connective tissue disorder lacking a definitive biomarker, leading to delayed diagnosis. Using the N3C EHR dataset, we first developed computational phenotyping models to identify potential hEDS patients . A scoring-based rule model was built from enriched conditions, while logistic regression and tree-based machine learning (ML) models were trained on the same features, with SHAP values for interpretability. The rule-based model identified 6,819 diagnosed patients and predicted 376,955 undiagnosed (~3% of the cohort), showing broad sensitivity but low precision; the ML model identified 13,357 diagnosed and 12,397 predicted undiagnosed (~0.1%), with higher selectivity and recall. Second, to explore heterogeneity, we clustered hEDS cases using phecode-frequency vectors. Dimensionality reduction and multiple algorithms yielded seven stable subgroups with distinct organ system involvement, all including musculoskeletal, gastrointestinal, and neurological domains, ranging from mild musculoskeletal-dominant to severe multisystem manifestations. These results show that EHR-based computational phenotyping can accurately detect hEDS and that clustering identifies clinically meaningful subtypes, providing a scalable framework for early identification, phenotypic stratification, and future biomarker discovery in precision medicine.

Speaker(s):
Megan Pearson, BS
University of Colorado Anschutz Medical Campus

Author(s):
Bryan Laraway, MS; Melissa Haendel, PhD - University of North Carolina at Chapel Hill; Megan Pearson, BS - University of Colorado Anschutz Medical Campus;

---

Enhancing Cancer Risk Identification Through a Digital Genetic Risk Assessment Tool

Presentation Type: Oral Presentation - Regular

Click to View Presentation

Presentation Time: 01:42 PM - 01:54 PM

Abstract Keywords: Population Health, Digital Therapeutics, Remote Patient Monitoring (RPM), and Digital Engagement, Innovation Partnerships, Implementation Science, and Learning Health Systems, Clinical Decision Support and Care Pathways
Primary Track: Implementing Real-World Change, Digital Engagement, and Connected Health


Many individuals at high risk for hereditary cancer lack awareness and access to expert guidance. MSK's Genetic Risk Assessment tool helps users evaluate hereditary cancer risk and access appropriate care. This study analyzes 15,871 assessments from 2021-2025, before and after Epic integration. Direct-to-test recommendations increased from 23% to 33% post-Epic, eliminating pre-test visits and saving one appointment slot per patient. Findings demonstrate how digital tools create clinical workflow efficiencies while supporting cancer risk identification.

Speaker(s):
Fernanda Polubriaginof, MD PhD
Memorial Sloan Kettering Cancer Center

Author(s):
Margaret Sheehan, MS - Memorial Sloan Kettering; Brian Dolan, MA - Memorial Sloan Kettering; Grisselle DeFrank, MS - Memorial Sloan Kettering; Erin Salo-Mullen, MS MPH - Memorial Sloan Kettering; David Wylie, MS - Memorial Sloan Kettering; Gilad Kuperman, MD, PhD - Columbia University; Kenneth Offit, MD MPH - Memorial Sloan Kettering; Zsofia Stadler, MD - Memorial Sloan Kettering;

---

Unmasking Inhibitor Development in Hemophilia A and B Using Interpretable Machine Learning: A Logistic Regression Approach on CHAMP and CHBMP Mutation Data

Presentation Type: Oral Presentation - Student

Click to View Presentation

Presentation Time: 01:54 PM - 02:06 PM

Abstract Keywords: Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Health Data Science, Precision Medicine, Multi-Omics, and Pharmacology Integration
Primary Track: Big Data for Health


Predicting inhibitor development in hemophilia remains one of the most challenging problems in genetic medicine. Inhibitors disrupt standard factor therapy, increase clinical burden, and disproportionately affect patients with severe genetic variants. Yet predictive modeling in this domain often struggles with data imbalance and the need for clinical interpretability.
In this study, we apply an interpretable machine learning framework—logistic regression—to mutation-level datasets from the CDC Hemophilia Mutation Project (CHAMP for Hemophilia A and CHBMP for Hemophilia B). Our goal is to determine whether pre-treatment genetic information can help identify patients with a history of inhibitor development. We conducted comprehensive exploratory analysis, characterizing variant type, mutation mechanism, exon distribution, and severity across both datasets.
Models were trained separately on CHAMP and CHBMP using one-hot encoded features and SMOTE to address class imbalance. Performance was evaluated using AUC, precision-recall curves, F1-scores, and confusion matrices. The CHAMP model achieved an AUC of 0.72, demonstrating meaningful discrimination despite limited inhibitor-positive cases. High-impact mutations—such as nonsense, frameshift, and large structural changes-emerged as consistently important predictors in both disorders.
Our findings highlight interpretable ML as a valuable tool in hemophilia genomics, revealing cross-dataset patterns that align with biological expectations. By focusing on transparent methods, this work supports the development of clinically trustworthy prediction tools that can enhance genetic counseling, treatment planning, and patient stratification in Hemophilia A and B.

Speaker(s):
Swarna Bharathi Kathi, Masters of Science in Health and Bioinformatics, Bachelors in Dental Surgery
Grand Valley State University

Author(s):
Swarna Bharathi Kathi, Masters of Science in Health and Bioinformatics, Bachelors in Dental Surgery - Grand Valley State University; Pravallika Manchu, Masters - Grand Valley State University; Md Kamrul Hasan, Assistant Professor of Practice; Kamrul Hasan, Ph.D - Grand Valley State University; Suhila Sawesi, PhD - GVSU;

---

Graph Transformer–Driven Multimodal Representation Learning for Drug Discovery

Presentation Type: Oral Presentation - Regular

Click to View Presentation

Presentation Time: 02:06 PM - 02:18 PM

Abstract Keywords: Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Precision Medicine, Multi-Omics, and Pharmacology Integration, Health Data Science
Working Group: Education Working Group
Primary Track: Big Data for Health


Predicting protein–ligand binding affinity is a cornerstone of virtual screening and drug design, yet many models struggle to generalize across across diverse chemical and structural spaces. We developed a multimodal graph transformer framework that unifies sequence-based protein embeddings, AlphaFold3-derived structural descriptors, and 3D ligand conformers. Rather than relying on a single representation, our approach encodes these features into unified protein–ligand graphs, utilizing a hybrid architecture of SchNet, Graphormer, DimeNet++, and GATv2Conv to capture both local atomic interactions and global structural context. Testing on the PDBbind v2020 benchmark demonstrates that this integration significantly reduces error compared to single modality baselines, achieving a Pearson correlation of 0.803 and an RMSE of 1.308 pK. These results highlight the necessity of cross-model feature fusion for robust, structure-based drug discovery.

Speaker(s):
BILAL RATHER, PH.D
State University of New York

Author(s):
BILAL RATHER, PH.D - State University of New York; Ram Samudrala, PhD - University at Buffalo; Zackary Falls, PhD - University at Buffalo, Jacobs School of Medicine and Biomedical Sciences;

---

The Invisible Data- Mapping The Perimenopausal Data Gap

Presentation Type: Oral Presentation - Student

Click to View Presentation

Presentation Time: 02:18 PM - 02:30 PM

Abstract Keywords: Population Health, Digital Therapeutics, Remote Patient Monitoring (RPM), and Digital Engagement, Social Determinants of Health (SDoH), Health Data Science
Primary Track: Implementing Real-World Change, Digital Engagement, and Connected Health


Perimenopause impacts millions of women worldwide, yet it remains underdiagnosed and poorly represented in clinical research. This study highlights a broad spectrum of symptoms, patterns in symptom reporting, information exchange, and coping strategies discussed within digital communities. The findings aim to reduce existing data gaps by translating lived perimenopausal experiences into insights that can inform more empathetic digital health communication tools, enhance quality of life for individuals experiencing perimenopause, and offer deeper understanding for symptom management.

Speaker(s):
Nidhi Mittal, Student
University of California,Davis

Author(s):
Katherine Kim, PhD, MPH, MBA, FAMIA - University of California Davis; Nidhi Mittal, Student - University of California,Davis; Akshita Sivakumar, Ph.D. Communication and Science Studies ,Master of Science in Architecture Studies (SMArchS) Massachusetts Institute of Technology, (MArch.), Bachelor of Science. Engineering Core - University of California, Davis; Seth Frey, PhD, B.A - University of California, Davis;

---

Qualitative Methods to Build Quantitative Machines: A Mixed-Methods Framework for Context-Aware Transfusion Decision Support

Presentation Type: Oral Presentation - Student

Click to View Presentation

Presentation Time: 02:30 PM - 02:42 PM

Abstract Keywords: Clinical Decision Support and Care Pathways, Human Factors and Usability, Analytical Artificial Intelligence: ML, Digital Pathology, Imaging AI, Predictive Analytics, Governance, Diagnostics, Clinician Well-Being, Ethics, Health Data Science, Outcomes Improvement and Equity
Primary Track: Advancing Wellness for Providers and Community with Consideration of Human Factors


We developed a mixed-methods framework that integrates trauma surgeons’ qualitative reasoning into quantitative transfusion decision support. Using questionnaires and interviews with attending surgeons, we identified key decision heuristics—mechanism of injury, hypotension, mental status, and hemodynamic trends—and tested them against ten years of trauma data (n=33,824). These variables were strongly associated with transfusion volume, informing a Parse → Structure → Validate approach for building context-aware, interpretable clinical decision support models.

Speaker(s):
Anais Kolesnikov, Masters of Science
OHSU

Author(s):
Michael Kolesnikov - OHSU; Anais Kolesnikov, Masters of Science - OHSU; Vishnu Mohan, MD, MBI, FACP, FAMIA - Oregon Health & Science University;

---