Tianyi Sun dissertation defense – October 31
PhD candidate Tianyi Sun is defending her dissertation on Thursday, October 31, at 9 a.m. Central Time. Her advisor is Dandan Liu. All are invited and encouraged to attend.
The defense will be conducted in Suite 1020, Room 10105 (10th floor conference room) at 2525 West End Avenue. It will also be streamed on Zoom; for virtual access, contact the department at biostatistics[at]vumc[dot]org.
Addressing Statistical Challenges in Implementing Real-World Evidence-Based Risk Prediction Models into EHR Systems
Clinical prediction models have been widely acknowledged as informative tools providing evidence-based support for clinical decision making. However, prediction models are often underused in clinical practice due to various reasons. One major challenge is to handle missing risk factors in real-time risk score calculation. In this dissertation, we proposed a novel submodel approach for prediction models developed using the model approximation approach for model selection, and later we extended this approach for prediction models developed using logistic regression. The proposed submodel approaches have the advantage of borrowing information from the target population. We conducted comprehensive simulations studies to assess the model performance of our proposed approaches and compared them with the existing “one-step-sweep”-based submodel approach as well as the imputation approach. The simulation results show the proposed submodel approaches are robust to various heterogeneity scenarios and are comparable to the imputation-based approach, while the “one-step-sweep” approach is less robust under certain heterogeneity scenarios. The proposed submodel approaches were applied to address missing risk factor issues in the real-time implementation of the STRATIFY prediction model to safely discharge low-risk acute heart failure patients. Another common challenge is to adapt the prediction model into local setting with uncollected risk factors. Here, we assessed and compared multiple approaches to revise a risk prediction model with proxy risk factors of those uncollected risk factors. The proposed approaches were compared under various simulation scenarios and were later applied to revise the STRATIFY prediction model using EHR data from Henry Ford Health.
