Sunday, November 1, 2026

11:00 - 12:30

Lecture Theatre 3

---

Details

In response to the growing global interest in machine learning (ML) within pharmacoepidemiology, this symposium aims to explore ML's role beyond standard predictive modeling, focusing on how ML can enhance traditional pharmacoepidemiologic methods and causal inference. Specifically, we want to highlight ML approaches for data collection, feature extraction, confounding adjustment, and causal effect estimation, raise awareness of the pros and cons of those approaches, discuss perspectives, and overall promote rigorous and standardized ML applications in pharmacoepidemiology in the Asia-Pacific region. Observational studies remain vulnerable to confounding, treatment selection bias, and patient heterogeneity. While ML is widely recognized for its utility in predictive modeling and clinical decision support systems, its potential to address these core epidemiologic challenges remains underexplored. We want to propose a cross-country symposium highlighting initiatives across the Asia-Pacific region where ML supported pharmacoepidemiology designs, from data collection and processing to analysis. This session describes novel, hybrid approaches that integrate ML abilities into rigorous study design starting from scientific context investigation, to data processing and data analysis to improve validity, transparency, or interpretability. We propose the following agenda: 1. Opening and introduction Dr Nathan Vidal (National Cheng Kung University, Tainan, Taiwan) will open the symposium by defining key terminology and the objectives of the session. An interactive poll will be conducted to assess the audience’s use of and familiarity with ML applications to support traditional analyses, providing an estimate of the prevalence of these hybrid approaches. He will then outline the session plan. Each subsequent presentation will focus on a specific step of the research process, moving from scientific context investigation, through data processing, to data analysis. Speakers will describe the scientific context of their ML project(s) and describe the hybrid methodology, outlining the application criteria, the ML parameters determination, the pros and cons, and benchmarking the ML approach against more traditional methods. 2. Scientific context investigation: machine learning for extracting key scientific information from text - Professor Shirley Xue Li (The University of Hong Kong, Hong Kong) will describe the application of Large Language Models to extract medical data from scientific literature. She will discuss how this approach can facilitate scientific background investigations and study quality assessments. Key notions will include prompt engineering, metrics for evaluating the model’s accuracy, risk of bias, and future perspectives for systematic reviews and quality assessment for individual study. - Short Q&A 3. Data processing: machine learning to extract relevant data features - Dr Richeek Pradhan (Monash University, Parkville, Australia) will share his experience using data-driven definition of relevant features collected in a claims database to build a probabilistic phenotyping algorithm aimed at automating drug surveillance. The presentation will cover the advantages and future perspectives of combining a classifier like regularized regression with human-defined clinical criteria to build an algorithm and automate feature extraction. - Short Q&A 4. Data analysis Machine Learning for confounding adjustment and modeling of complex clinical data - Professor Aimin Yang (The Chinese University of Hong Kong, Hong Kong) will explore how ML can enhance confounding adjustment, improve modeling of complex clinical data, and help identify clinically meaningful heterogeneity in treatment effects in the context of diabetes research. The presentation will emphasize the concept of "causal insight" in ML-enhanced diabetes pharmacoepidemiology, stressing the importance of interpretability for informing therapeutic decision-making in real-world practice. - Short Q&A Machine learning for causal effect estimation (15 min) - Mr Junhyuk Chang (Ajou University, Suwon, Korea) will present causal machine learning methodologies designed for robust causal inference and treatment effect estimation. Using practical case studies from his own research, he will describe causal forests, heterogeneous treatment effects, and the differences between these models and traditional effect modifier investigation. - Short Q&A 5. Global discussion A moderated Q&A will allow attendees to engage directly with the panel. If time permits, the moderator will invite audience members to share brief insights from their own experience using ML to support their pharmacoepidemiologic research. To spark discussion and raise awareness, a specific slide will highlight alternative ML applications not covered by the main panel, including, for instance: - ML for de-identifying clinical data. - ML for proofreading and correcting datasets. - More general use of ML for refining search strategies, screening titles, and summarizing complex findings (e.g., PRISMA-AI guidelines, Consensus.app, PaperDigest). 6. Closing: The moderators will highlight the best practices for each method raised during the symposium and share relevant resources. The session will conclude with a final quick poll to determine how many researchers in the audience plan to incorporate these hybrid methods into their future projects, assessing the immediate impact of the symposium. This initiative aims to: 1. Reviewing current evidence on how machine learning applications can assist researchers in building robust pharmacoepidemiological study designs, ultimately popularizing the most relevant methods. 2. Because ML tools are rapidly evolving and globally accessible, standardizing ML research practices is crucial. By providing an up-to-date snapshot of the ML landscape, we can guide its proper application and openly discuss its limitations, which is especially important for educational purposes. 3. Raising awareness of alternative ML methodologies. These are particularly valuable for researchers who face restricted access to health data, lack the resources or time for extensive data cleaning, or need alternative approaches to causal inference. Ultimately, the materials and presenters will share their singular experience, enabling us as a group to build a central take-home message on rigorous methodology: identifying and mitigating the pitfalls of those ML methods, promoting established best practices, and emphasizing the critical methodological details that must be transparently reported.

---

Speaker