主講人：林宗霖 博士 (陽明交通大學統計學研究所)

題  目：Evaluation of diagnostic test accuracy using latent class models with conditional dependence

日  期：115年5月19日（星期二）

時  間：下午13:00 ~ 14:10

地  點：科學館S433

摘要Abstract:

Latent class models (LCMs) are essential for evaluating diagnostic test accuracy in the absence of a gold standard. A foundational assumption of these models is local independence, which posits that observed tests are independent conditional on the latent disease status. However, this assumption is frequently violated in clinical practice when tests share biological mechanisms, leading to biased estimates of sensitivity, specificity, and disease prevalence. We propose a flexible extension of the LCM by reformulating it as a constrained log-linear model. This approach simultaneously captures marginal response probabilities and complex within-class dependencies, including higher-order interactions. An iterative proportional fitting algorithm is implemented to systematically align the model with specified marginal constraints and association structures, ensuring numerical stability in parameter estimation. The practical utility of the proposed method is demonstrated using a clinical dataset of 328 older patients to evaluate postoperative frailty. Our findings indicate that accounting for conditional dependence significantly refines the assessment of recovery status compared to standard LCMs.

備  註：本業務與聯合國永續發展目標SDG4優質教育連結

ESG+AI=∞  AI+SDGs=∞