Adjusting Covariate Misclassification in Electronic Health Records-Based Machine Learning Prediction Models

Presentation Time: 08:00 AM - 08:12 AM

Abstract Keywords: Fairness and Elimination of Bias, Quantitative Methods, Artificial Intelligence
Primary Track: Applications

This study developed and evaluated methods to adjust misclassification errors in electronic health record (EHR)-derived covariates using group-wise and individualized weights based on observed sensitivity and specificity to reduce bias in predictive modeling. Logistic regression, XGBoost, and neural networks predicted follow-up adherence in lung cancer screening. The Lung-RADS category, extracted via natural language processing (NLP), was adjusted using group-wise weights and individualized weights from kernel and multinomial regression. Models with adjusted covariates were compared to naïve (unadjusted) and oracle (true value) models. Performance assessed by the area under the receiver operating characteristic (AUROC) curve across 10%, 20%, and 30% validation sets, showed that adjusted models outperformed naïve models, improving AUROC by 0.3%–10.4%. Compared to oracle models, adjusted models reduced the AUROC gap to 2.0%–7.5%. Individualized weights provided more precise corrections than group-wise weights. This scalable framework mitigates misclassification bias in EHR-derived covariates, enhancing predictive accuracy without resource-intensive manual review.

Speaker:
Shuang Yang, MS
University of Florida

Authors:
Shuang Yang, MS - University of Florida; Yonghui Wu, PhD - University of Florida; Mei Liu, PhD - University of Florida; Jiang Bian, PhD - University of Florida; Yi Guo, PhD - University of Florida; Muxuan Liang, PhD - University of Florida;

---

Cryptogenic Stroke and Migraine: Using Probabilistic Independence and Machine Learning to Uncover Latent Sources of Disease from the Electronic Health Record

Presentation Time: 08:12 AM - 08:24 AM

Abstract Keywords: Machine Learning, Causal Inference, Bioinformatics, Quantitative Methods, Artificial Intelligence
Primary Track: Applications
Programmatic Theme: Clinical Research Informatics

Migraine is a common but complex neurological disorder that doubles the lifetime risk of cryptogenic stroke (CS). However, this relationship remains poorly characterized, and few clinical guidelines exist to reduce this associated risk. We therefore propose a data-driven approach to extract probabilistically-independent sources from electronic health record (EHR) data and create a 10-year risk-predictive model for CS in migraine patients. These sources represent external latent variables acting on the causal graph constructed from the EHR data and approximate root causes of CS in our population. A random forest model trained on patient expressions of these sources demonstrated good accuracy (ROC 0.771) and identified the top 10 most predictive sources of CS in migraine patients. These sources revealed that pharmacologic interventions were the most important factor in minimizing CS risk in our population and identified a factor related to allergic rhinitis as a potential causative source of CS in migraine patients.

Speaker:
Joshua Betts, M.D. Candidate
Vanderbilt University School of Medicine

Authors:
Joshua Betts, M.D. Candidate - Vanderbilt University School of Medicine; John Still, Application Developer - Vanderbilt University Medical Center; Thomas Lasko, MD, PhD - Vanderbilt University Medical Center;

---

Machine Learning for Predicting Drug Release Behavior of PLGA Microspheres

Presentation Time: 08:24 AM - 08:36 AM

Abstract Keywords: Machine Learning, Drug Discoveries, Repurposing, and Side-effect, Data Mining
Primary Track: Applications
Programmatic Theme: Clinical Research Informatics

PLGA microspheres are widely used in long-acting drug formulations due to their ability to provide sustained release, improving patient adherence and reducing dosing frequency. However, drug release behavior is influenced by complex formulation and processing factors, making traditional trial-and-error development inefficient. This study leverages machine learning to predict drug release profiles from PLGA (poly(lactic-co-glycolic acid)) microsphere formulations. A dataset of 113 PLGA formulations containing small-molecule drugs and large-molecule peptides was collected from published literature. Multiple machine learning models were developed and compared. The best-performing model achieved an R² value of 0. 9415, a RMSE of 6.99% and a MAE of 4.35%, demonstrating strong predictive accuracy for in vitro drug release. Additionally, feature importance analysis was conducted, offering insights into key factors influencing release behavior and guiding the rational design of PLGA-based microspheres.

Speaker:
Ling Zheng, PhD in Computer Science
Monmouth University

Authors:
Andrew Catapano, BS - Monmouth University; Ling Zheng, PhD in Computer Science - Monmouth University; Xudong Yuan, PhD - ACON Pharmaceuticals; Kelly Yuan, NA - ACON Pharmaceuticals;

---

A Multi-Phase Analysis of Blood Culture Stewardship: Machine Learning Prediction, Expert Recommendation Assessment, and LLM Automation

Presentation Time: 08:36 AM - 08:48 AM

Abstract Keywords: Infectious Diseases and Epidemiology, Large Language Models (LLMs), Machine Learning, Artificial Intelligence, Bioinformatics
Primary Track: Applications
Programmatic Theme: Clinical Research Informatics

Blood cultures are often overordered without clear justification, straining healthcare resources and contributing to inappropriate antibiotic use—pressures worsened by the global shortage. In study of 135,483 emergency department (ED) blood culture orders, we developed machine learning (ML) models to predict the risk of bacteremia using structured electronic health record (EHR) data and provider notes via a large language model (LLM). The structured model’s AUC improved from 0.76 to 0.79 with note embeddings and reached 0.81 with added diagnosis codes. Compared to an expert recommendation framework applied by human reviewers and an LLM-based pipeline, our ML approach offered higher specificity without compromising sensitivity. The recommendation framework achieved sensitivity 86%, specificity 57%, while the LLM maintained high sensitivity (96%) but overclassified negatives, reducing specificity (16%). These findings demonstrate that ML models integrating structured and unstructured data can outperform consensus recommendations, enhancing diagnostic stewardship beyond existing standards of care.

Speaker:
Fatemeh Amrollahi, PhD
Stanford University

Authors:
Fatemeh Amrollahi, PhD - Stanford University; Nicholas Marshall, MD - Stanford; Fateme Nateghi Haredasht, PhD - Stanford University; Kameron Black, DO, MPH - Stanford University; Aydin Zahedivash, MD, MBA - Stanford University; Manoj Maddali, MD - Stanford; Stephen Ma, MD, PhD - Stanford University School of Medicine; Amy Chang, MD, PharmD - Stanford University; Stanley Deresinski, MD - Stanford; Mary Goldstein, MD, MS in HSR - Stanford University; Steven Asch, MD - Stanford; Niaz Banaei, MD - Stanford; Jonathan Chen, MD, PhD - Stanford University Hospital;

---

Transformer-based Artificial Intelligence for Predicting Prescribed Medications from Unstructured Clinical Data

Presentation Time: 08:48 AM - 09:00 AM

Abstract Keywords: Clinical Decision Support, Artificial Intelligence, Deep Learning, Natural Language Processing, Workflow, Health Equity, Healthcare Economics/Cost of Care
Primary Track: Foundations
Programmatic Theme: Clinical Informatics

This study evaluated transformer-based artificial intelligence for predicting medications prescribed during primary care visits using unstructured clinical data. The fine-tuned model achieved a clinically useful top-12 accuracy of 80.1%, significantly outperforming prior LSTM-based approaches. The model requires minimal preprocessing, facilitating its adoption and portability across healthcare settings. These findings demonstrate the potential of transformer-based clinical NLP models to support proactive, cost-informed prescribing decisions through the anticipatory provision of patient-specific co-pay information.

Speaker:
Christian Balbin, PhD
University of Utah

Author:
Kensaku Kawamoto, MD, PhD, MHS - University of Utah;

---

Towards Interpretable, Sequential Multiple Instance Learning

Presentation Time: 09:00 AM - 09:12 AM

Abstract Keywords: Artificial Intelligence, Bioinformatics, Machine Learning, Clinical Decision Support
Primary Track: Foundations
Programmatic Theme: Clinical Informatics

This work introduces the Sequential Multiple Instance Learning (SMIL) framework, addressing the challenge of interpreting sequential, variable-length sequences of medical images with a single diagnostic label. Diverging from traditional MIL approaches that treat image sequences as unordered sets, SMIL systematically integrates the sequential nature of clinical imaging. We develop a bidirectional Transformer architecture, BiSMIL, that optimizes for both early and final prediction accuracies through a novel training procedure to balance diagnostic accuracy with operational efficiency. We evaluate BiSMIL on three medical image datasets to demonstrate that it simultaneously achieves state-of-the-art final accuracy and superior performance in early prediction accuracy, requiring 30-50\% fewer images for a similar level of performance compared to existing models. Additionally, we introduce SMILU, an interpretable uncertainty metric that outperforms traditional metrics in identifying challenging instances.

Speaker:
Xiaolong Luo, Doctoral Student
Harvard University

Authors:
Xiaolong Luo, Doctoral Student - Harvard University; Hsin-Hsiao Wang, Attending Surgeon - Boston Childrens Hospital; Michael Lingzhi Li, PhD - Harvard University;

---