|
14:00
15:30
|
-
Hsiu-Hsi ChenTaiwan
Speaker
AI-Driven Digital Twin Medicine for Precision Diabetes CareDiabetes mellitus is one of the most rapidly growing global health challenges, affecting nearly 589 million adults worldwide and projected to reach a prevalence of approximately 13% by 2050. Early detection and timely intervention are therefore critical to preventing long-term microvascular and macrovascular complications. However, type 2 diabetes typically evolves through a prolonged asymptomatic phase, creating a substantial window for population-based screening and preventive strategies.
Using data from the Changhua Community-based Integrated Screening (CHCIS) program in Taiwan, we quantified the natural history of diabetes progression and developed personalized risk prediction models. In this population, prediabetes prevalence reached approximately 18.6% and type 2 diabetes prevalence 10.6%. Continuous-time Markov modeling characterized the dynamic metabolic trajectory from normoglycemia to prediabetes, asymptomatic diabetes, and symptomatic disease, with estimated transition rates of 0.0328, 0.1764, and 0.0988 per person-year, respectively, highlighting a prolonged subclinical phase during which targeted screening and intervention may substantially alter disease progression.
Building on this natural history framework, multistate personalized risk prediction models incorporating demographic, metabolic, and lifestyle factors—including age, body mass index, blood pressure, triglycerides, liver enzymes, smoking status, and family history of diabetes—were developed to estimate individualized progression risks. The machine learning with Random forest was adopted to identify important personalized features for predictive performance across disease stages and provided empirical support for targeted screening and precision prevention strategies.
To further refine individualized prediction, genetic risk markers identified in prospective cohort studies were incorporated into multivariate risk scores. Variants in genes involved in glucose metabolism and β-cell function contributed to pathway-specific risk scores for transitions from normal glucose tolerance to impaired fasting glucose and from prediabetes to diabetes. Integrating clinical and genetic risk factors enabled personalized kinetic trajectories for disease progression and intervention strategies.
Emerging evidence further suggests that integrating multi-omics data—including metabolomics, proteomics, epigenomics, and microbiome profiles—may enhance AI-based personalized risk prediction by capturing dynamic biological states and gene–environment interactions beyond static genetic susceptibility. In parallel, continuous glucose monitoring (CGM) technologies provide high-resolution physiological signals reflecting real-time glycemic dynamics, lifestyle responses, and treatment effects.
Integrating population screening data, multiscale risk prediction, genetic and multi-omics susceptibility profiles, and real-time CGM signals creates the foundation for AI-driven digital twin medicine, in which a dynamic virtual representation of each individual can simulate metabolic trajectories and evaluate alternative interventions before they occur in reality.
Ultimately, the convergence of population screening, multiscale risk prediction, multi-omics profiling, and continuous metabolic monitoring within AI-driven digital twin frameworks may redefine diabetes care—shifting healthcare systems from reactive disease treatment toward proactive precision prevention and lifelong metabolic health management.
Keywords: Diabetes mellitus; Digital twin medicine; Continuous glucose monitoring; Multistate Markov model; Precision medicine; Genetic risk prediction; Population screening; Artificial intelligence
-
I-Wen WuTaiwan
Speaker
AI-Assisted Discovery of Omic Signature in Diabetic Kidney DiseaseDiabetic kidney disease (DKD) is a leading cause of chronic kidney disease and end-stage renal disease, posing significant global health and economic burdens. Traditional management of DKD relies on standardized approaches, which often fail to account for the complexity of individual patient profiles.
Precision medicine leverages individualized patient data—spanning genetic, proteomic, metabolic, and clinical information—to optimize diagnosis, risk assessment, and therapeutic interventions. However, translating this paradigm into clinical practice presents significant challenges and opportunities.
This presentation focuses on the practical aspects of integrating precision medicine into DKD management. Key themes include the role of genetic and epigenetic biomarkers in risk stratification, the integration of multi-omics data with machine learning for predictive modeling and the design of personalized treatment regimens using tools such as pharmacogenomics.
By examining real-world implementation strategies and overcoming barriers, this presentation aims to guide healthcare providers, researchers, and policymakers toward a sustainable and patient-centered precision medicine framework for DKD.
-
Dee PeiTaiwan
Speaker
The roles of Machine Learning in Medical ResearchArtificial intelligence (AI) is fundamentally transforming clinical research by democratizing access to advanced analytical tools and establishing new standards for scientific publication. This presentation outlines a comprehensive framework for integrating machine learning (ML) methodologies into clinical studies—from raw data preparation to manuscript submission—while addressing critical challenges in model development, validation, and interpretation. We emphasize that financial barriers to sophisticated analysis have largely dissolved with the advent of open-source AI platforms, enabling researchers to move beyond legacy statistical software toward reproducible, transparent ML pipelines.
The workflow begins with strategic data preparation, including appropriate imputation techniques (k-NN, MissForest, MICE) and feature standardization. For binary classification tasks—common in clinical prediction—It is advocate a rigorous protocol encompassing stratified cross-validation, hyperparameter tuning via nested CV, and explicit overfitting controls (regularization, feature limitation to ≥10 events per variable). Model selection should prioritize algorithms matching the clinical question: logistic regression for interpretability, random forests for robust baselines, and gradient boosting for maximal performance—while acknowledging trade-offs in complexity and calibration risk.
Evaluation must extend beyond conventional ROC-AUC to include precision-recall curves (especially for imbalanced data), calibration assessment (slope ~1, intercept ~0), Brier score, and decision curve analysis for clinical utility. SHAP values provide essential interpretability for "black-box" models, translating complex predictions into clinically actionable insights. Crucially, it is stressed that accuracy alone is misleading in medical contexts; minimizing false negatives often carries greater clinical consequence than overall accuracy.
Reproducibility demands fixed random seeds, complete pipeline documentation, and packaging preprocessing steps with final models for deployment. As journals increasingly expect ML-enhanced analyses, studies relying solely on traditional statistics face diminished publication prospects. It is concluded that AI integration is no longer optional but essential for contemporary clinical research seeking impact, rigor, and real-world applicability in an era where algorithmic insight complements—not replaces—clinical expertise.
3F Banquet Hall
|
