Predicting the Disease Burden of Organ Failure Using Generative Transformers: A Standardized Tool for National Disease Surveillance

Presentation Type: Podium Abstract
Presentation Time: 02:00 PM - 02:12 PM

Primary Track: Clinical Research Informatics


We developed a transformer-based model trained on China’s medical record front page data to predict kidney failure risk among patients with diabetes. Using longitudinal ICD-coded trajectories from 12,194 patients, the model achieved AUROC up to 0.903 across prediction intervals of 3–60 months and remained robust after excluding near-event diagnoses. These findings demonstrate that diagnostic sequences contain meaningful long-range progression signals and offer a scalable foundation for organ failure surveillance and noncommunicable disease burden assessment.

Speaker(s):
Chenxin Song, Master
The First Affiliated Hospital of Sun Yat-Sen University

Author(s):
Chenxin Song, Master - The First Affiliated Hospital of Sun Yat-Sen University; Wei Lixia, Engineering - SYSU-FAH; Haibo Wang - First Affiliated Hospital, Sun Yat-Sen University; Scott McGrath, PhD FAMIA - CITRIS Health UC Berkeley; Katherine Kim, PhD, MPH, MBA, FAMIA - University of California Davis; Nicholas Anderson, PhD - University of California, Davis; Haipeng Xiao, MD, PhD - The First Affiliated Hospital of Sun Yat-sen University;

---

Machine Learning Based County Level Phenotypes Related to Diabetes Prevalence

Presentation Type: Paper - Student
Presentation Time: 02:12 PM - 02:24 PM

Primary Track: Data Science/Artificial Intelligence


In the US, diabetes prevalence rates continue to rise. However little focus is given on the association of diabetes with the social determinants of health (SDoH). This study focuses on developing phenotypes based on county-level SDoH which are related to diabetes prevalence. Machine learning algorithms such as classification and regression tree (CART) model were used to define phenotypes based on county-level SDoH. Random forest was also used to identify additional risk factors. Five different phenotypes identified by the CART model divided the counties into five groups. Counties with high food insecurity rates (more than 16%) and high poverty rates (more than 24%) were found to have higher mean prevalence rate of diabetes (17.64%, SD 2.42). These can help individuals involved in healthcare and policy makers to make tailored region-based interventions to reduce diabetes prevalence and improve living conditions for people with diabetes.

Speaker(s):
MD FITRAT HOSSAIN, Ph.D.
University of Maryland, Baltimore

Author(s):
Fadia Shaya, PhD, MPH - University of Maryland Schools of Pharmacy & Medicine; MD FITRAT HOSSAIN, Ph.D. - University of Maryland, Baltimore;

---

Spatio-Temporal Modeling for Multi-County Opioid Overdose Surveillance: A Unified Graph Convolutional Framework

Presentation Type: Paper - Regular
Presentation Time: 02:24 PM - 02:36 PM

Primary Track: Data Science/Artificial Intelligence


This study introduces a unified spatio-temporal predictive framework for estimating opioid-involved mortality across six U.S. regions. A spatio-temporal graph convolutional network was used to generate monthly grid level predictions while incorporating local spatial adjacency and temporal progression derived from observed mortality trends. The standardized representation enables comparison of spatial clustering, temporal variability, and prediction behavior across jurisdictions that differ in geographic layout, population distribution, and mortality burden. The framework provides a basis for examining how regional characteristics relate to predictive patterns and offers a way to assess whether structures learned in one region also appear in others with distinct environments. This approach may support analysis of regional variation in mortality dynamics and help identify consistent features of opioid involvement that emerge across heterogeneous public health settings.

Speaker(s):
Dohyo Jeong, PhD
University of Kentucky

Author(s):
Dohyo Jeong, PhD - University of Kentucky; Daniel Harris, PhD - University of Kentucky;

---

Opioid Overdose Death Prediction with Graph Neural Networks

Presentation Type: Paper - Student
Presentation Time: 02:36 PM - 02:48 PM

Primary Track: Data Science/Artificial Intelligence


The opioid crisis has severely impacted Ohio, with overdose death rates surpassing national averages and disproportionately affecting rural and Appalachian regions. Accurately predicting county-level opioid overdose deaths (OD) is critical for timely intervention but remains challenging due to the wide differences in opioid OD between large and small counties. We propose a Spatial-Temporal Graph Neural Network (ST-GNN) framework that integrates graph neural networks (GNNs) to capture spatial relationships between counties and Long Short-Term Memory (LSTM) networks to model temporal dynamics. Using quarterly OD data from Q1 2017 to Q2 2023 for 88 Ohio counties, we incorporate a nine-dimensional dynamic feature set, including naloxone administration events and high-risk opioid prescribing, along with a static Social Determinants of Health (SDoH) index. Compared to baseline models, our ST-GNN demonstrates superior performance, particularly in larger counties, while specialized strategies improve predictions for small counties, leading to more stable and reliable results. Our findings emphasize the need for spatial-temporal modeling and customized training to enhance public health decision-making in addressing the opioid crisis.

Speaker(s):
Xianhui Chen, Master
The Ohio State University

Author(s):
Zishan Gu, Master - OSU; John Myers, Master - The Ohio State University; Joanne Kim, PhD - The Ohio State University; Changchang Yin, MS - The Ohio State University; Naleef Fareed, PhD MBA - The Ohio State University Dept Biomedical Informatics; Neena Thomas, MS - The Ohio State University Dept of Biomedical Informatics; Soledad Fernandez - The Ohio State University Dept of Biomedical Informatics; Ping Zhang, PhD, FAMIA - The Ohio State University;

---

Association Between Substance Use and Stigmatizing Language Documented in Hospital Birth Admission Notes

Presentation Type: Podium Abstract
Presentation Time: 02:48 PM - 03:00 PM

Primary Track: Data Science/Artificial Intelligence


Background: Patients with substance use disorders (SUD) experience bias from clinicians when receiving pregnancy care. Stigmatizing language documented in electronic health records is a marker of bias and can reinforce inequitable care.
Objective: To examine the associations between substance use and stigmatizing language documented in clinical notes from the birth hospitalization.
Methods: This was a cross-sectional study of electronic health record data from patients ≥20 weeks gestation admitted to labor and birth at two urban hospitals between 2017-2019 (N=19,094). We identified patients with SUD through ICD-10 codes and natural language processing (NLP)-based keyword searches. We employed an NLP algorithm to identify three categories of stigmatizing language in clinical notes. We conducted multivariable logistic regression to examine odds of stigmatizing language by SUD history.
Results: Patients with SUD had higher odds of any stigmatizing language compared to patients without SUD (adjusted odds ratio [aOR]=2.16, 95% confidence interval [CI]=1.63, 2.85). Patients with SUD also had higher odds of language representing marginalized language/identities (aOR=3.98, 95% CI=2.98, 5.31) and 'difficult patient categories' (aOR=1.76, 95% CI=1.36, 2.28). There were no significant differences in the unilateral/authoritarian language category between patients with and without SUD (aOR=1.21, 95% CI=0.92, 1.61).
Discussion: Patients with SUD were more likely to have stigmatizing language documented during labor and birth. Findings indicate a need to implement scalable, automated methods to improve documentation practices in perinatal care, including real-time monitoring of EHR notes to flag stigmatizing language to promote perinatal health equity.

Speaker(s):
Sarah Harkins, PhD, RN
Columbia University Irving Medical Center

Author(s):
Max Topaz, PhD, RN, MA, FAAN, FIAHSI, FACMI - Columbia University; Veronica Barcelona, PhD, MSN, RN, PHNA-BC, FAAN - Columbia University School of Nursing; Sarah Harkins, PhD, RN - Columbia University Irving Medical Center;

---

Electronic health record-based healthcare continuity tracking after natural disasters: a case study of the 2025 St. Louis tornado

Presentation Type: Podium Abstract
Presentation Time: 03:00 PM - 03:12 PM

Primary Track: Data Science/Artificial Intelligence


Long-term healthcare access impacts of the May 2025 St. Louis EF3 tornado were evaluated using electronic health records (EHRs) from major regional health systems. 16,000 patients were identified in the response zone and initial summary statistics demonstrated the feasibility of EHR–based healthcare monitoring. This work establishes a foundation for multi-system data integration to assess chronic disease management, care disruptions, and disparities, providing a scalable EHR model for tracking healthcare continuity after natural disasters.

Speaker(s):
Abigail Lewis
Institute for Informatics at Washington University in St. Louis, School of Medicine

Author(s):
Alexander Garza, MD, MPH - Saint Louis University School of Medicine; Matthew Haslam, MPH - City of St. Louis Department of Health; Victoria Anwuri, MPH - City of St. Louis Department of Health; Adam Wilcox, PhD - Washington University in St. Louis; Albert Lai, PhD, FACMI, FAMIA - Washington University; Philip Payne, PhD, FACMI, FAMIA - WashU Medicine and BJC Healthcare;

---