Voici l’annonce du 5e séminaire au DIC pour la session HIVER 2018 par Shengrui WANG
Survival Analysis for Predicting Chronic Obstructive Pulmonary Disease Failures
Résumé - *à noter que la présentation sera en français*
Chronic obstructive pulmonary disease (COPD) yields a high rate of failures such as hospital readmission and death, and predicting such failure is crucial to early intervention and decision-making. In this talk, I will present our recent work on extending survival models for predicting COPD events. Two important issues will be addressed.
The first one is about modeling hazard in longitudinal clinical data. In fact, survival analysis is often confined to specific types of data involving only the present measurements. We consider a more general class of healthcare data found in practice, which includes a wealth of intermittently varying historical measurements in addition to the present measurements. In particular, we propose a new representation of hazard to capture the relationship between survival probability and time-varying risk factors. To optimize model parameters, i.e., regression coefficients, we design and maximize a new joint likelihood that comprises two components used for estimating survival status for failure and censored patients, respectively. A regularized optimization is performed on joint likelihood to prevent overfitting arising from model learning.
The second issue relates to risk factor selection. Current methods address all risk factors in medical records indiscriminately and therefore generally suffer from ineffectiveness in real applications. Numerous studies have been done on selecting factors for survival analysis with limited success in the context of unknown and intricate correlation patterns among risk factors.
These difficulties have prompted us to design a new Cox-based learning machine that embeds the feature weighting technique into failure prediction. In order to improve predictive accuracy, we propose two weighting criteria to maximize the area under the ROC curve (AUC) and the concordance index (C-index), respectively. At the same time, we perform a Dirichlet-based regularization on weights, making differences between factor relevance clearly visible while maintaining the model’s high predictive ability.
Our models have been tested on real-life COPD data collected from patients hospitalized at the Centre Hospitalier Universitaire de Sherbrooke (CHUS) and on a number of public data. Experimental results show they outperform the current state-of-the-art prediction models and reveal their promise in clinical applications for failure prediction.
Shengrui Wang is a Professor at the Department of Computer Science, University of Sherbrooke, Canada. He received his B.S. degree in mathematics from Hebei University, in 1982, and his Ph.D. degree in computer science from the Institute National Polytechnique de Grenoble, France, in 1989. His research interests cover wide areas of data mining, pattern recognition, machine learning with applications in bioinformatics, business intelligence, health informatics, activity recognition and user profiling in smart environments, recommendation systems, aging studies, etc. He is best known for his contributions in high-dimensional data clustering and in sequence analysis. He served in NSERC Computer Science discovery grant committee between 2010 and 2014. He has been serving in NSERC RTI committee in Computer, Mathematical and Statistical Sciences in the past three years and chaired this committee during 2015-16 and 2017-18.