Seminar by Professor Lei LIU, Division of Biostatistics, Washington University in St. Louis, U.S.A.

We present a deep learning framework for predicting time-to-primary open-angle glaucoma (POAG) using longitudinal visual field (VF) data from the Ocular Hypertension Treatment Study (OHTS). POAG is a leading cause of irreversible blindness, and VF tests are essential for tracking disease progression. Unlike prior studies that rely on single time points or treat prediction as a binary classification (glaucoma yes/no)—which misclassifies censored cases and reduces power—our approach models the full temporal trajectory of VF data. We integrate landmark analysis with advanced deep learning architectures—LSTM, CNN-LSTM, Transformer, and CNN-Transformer—to simultaneously capture the spatial structure of VF maps and their temporal evolution over repeated visits. Models are optimized using a negative partial likelihood loss tailored for survival analysis, enabling direct handling of censored time-to-event outcomes. Evaluation via 10-fold cross-validation at multiple landmark times shows that CNN-LSTM consistently achieves the highest concordance indices. In contrast, Transformers underperform in this relatively small and irregular dataset due to their on large training samples and lack of inherent temporal structure, while CNN-Transformers are hampered by feature compression that reduces the richness of attention inputs.

Dr. Lei Liu is a Professor in the Division of Biostatistics at Washington University in St. Louis. He has diverse research interests in biostatistical and data science methods, including survival analysis, longitudinal data analysis, spline regression, personalized medicine, and machine learning. His research is focused on the analysis of high dimensional omics (epigenetics and microbiome) data, medical cost data, and joint models of multi- outcome data. He collaborates with clinicians in various medical fields, e.g., cancer, cardiovascular, addiction, ophthalmology, nephrology, infectious disease, asthma, and diabetes. Dr. Liu is a Fellow of the American Statistical Association. He is an associated editor of Biometrics and Statistics in Medicine, and an editorial board member of the Journal of the National Cancer Institute and Frontiers in Psychiatry. He is a standing member of NIH Biostatistical Methods and Research Design Study Section (2016-22), the only study section focusing on biostatistical methodology development. He also reviews grants frequently for other NIH study sections and other funding agencies around the world.