Title: Support Vector Machine for Dynamic Survival Prediction with Time-Dependent Covariates

Abstract: Predicting time-to-event outcomes using time-dependent covariates is a challenging problem. Many machine learning approaches, such as tree-based methods and support vector regression, predominantly utilize only baseline covariates. Only a few methods can incorporate time-dependent covariates, but they often lack theoretical justification. In this paper, we present a new framework for event time prediction, leveraging the support vector machines to forecast the associated counting processes. Utilizing the kernel trick, we accommodate nonlinear functions in both time and covariate spaces. Subsequently, we use a chain algorithm to predict future events. Theoretical analysis proves that our method is equivalent to comparing time-varying hazard rates among at-risk subjects, and we obtain the convergence rate of the resulting prediction loss. Through simulation studies and a case study on Huntington’s disease, we demonstrate the superior performance of our approach compared to alternative methods based on machine learning, deep learning, and statistical models.

Bio: Donglin Zeng obtained PhD from the Department of Statistics at the University of Michigan in 2001. He was a faculty in the department of Biostatistics at the University of North Carolina from 2001 to 2023 and has recently joined the department of Biostatistics at the University of Michigan. He is an elected fellow of  the American Statistical Association and the Institute of Mathematical Statistics. His research interests include machine learning, precision medicine, high-dimensional data, semiparametric models and causal inference.

Title: Advancing Clinical Decision-Making with Representation Learning: Tackling Confounding and Censoring Biases in Counterfactual Survival Analysis

Abstract: Observational time-to-event data presents unique challenges for counterfactual survival analysis in clinical decision-making, primarily due to confounding and censoring biases. In this talk, we will explore these biases and their impact on treatment effect estimation with observational data. We will then introduce a novel approach that harnesses the power of representation learning to address these challenges. Our presentation will demonstrate the potential of representation learning in enhancing clinical decision-making through improved counterfactual survival analysis. By highlighting the benefits of this approach, we aim to encourage its adoption in healthcare and inspire further research in this promising and rapidly evolving field.

Bio: Paidamoyo Chapfuwa is a Senior Researcher at Microsoft Research, Health Futures, where she focuses on advancing representation learning approaches to decode the adaptive immune system. Her work involves leveraging statistical machine learning techniques, including causal inference, generative modeling, and Bayesian nonparametrics, to better understand complex biological data. She is particularly interested in modeling approaches for survival analysis, time-series, and high-dimensional sparse protein data. Paidamoyo received her B.S.E., M.S., and Ph.D. degrees in electrical and computer engineering from Duke University, where she was advised by Drs. Lawrence Carin and Ricardo Henao. She also completed a Postdoctoral fellowship in health policy at Stanford University, where she worked with Dr. Sherri Rose. Prior to her work in machine learning and computational biology, she had experience in radar signal processing, medical equipment repair and maintenance, software engineering, and hospital & health systems. For more information about her work, please visit her website at https://paidamoyo.github.io.

Title: Deep Learning for Censored Survival Data

Abstract: Unlike standard tasks, survival analysis requires modeling incomplete data, such as right-censored data, which must be treated with care. While deep neural networks excel in traditional supervised learning, it remains unclear how to best utilize these models in survival analysis. A key question asks which data-generating assumptions of traditional survival models should be retained and which should be made more flexible via the function-approximating capabilities of neural networks.  In addition, most of these methods are difficult to interpret and mathematical understanding of them is lacking. In this talk, we explore these issues from two directions. First, we study the partially linear Cox model, where the nonlinear component of the model is implemented using a deep neural network. The proposed approach is flexible and able to circumvent the curse of dimensionality, yet it facilitates interpretability of the effects of treatment covariates on survival.  Next, we introduce a Deep Extended Hazard (DeepEH) model to provide a flexible and general framework for deep survival analysis. The extended hazard model includes the conventional Cox proportional hazards  and accelerated failure time models as special cases, so DeepEH subsumes the popular Deep Cox proportional hazard (DeepSurv) and Deep Accelerated Failure Time (DeepAFT) models.  We provide theoretical support for the proposed models, which underscores the attractive feature that deep learning is able to detect low-dimensional structure of data in high-dimensional space.  Numerical experiments further provide evidence that the proposed methods outperform existing statistical and deep learning approaches to survival analysis.  Time permitting, we will explore how to perform hypothesis testing for survival data. 

Bio: Jane-Ling Wang is Distinguished Professor at the University of California, Davis. Her research interests include survival analysis, functional data analysis and machine learning.  She also enjoys collaborations with domain scientists. She is a member of the American Association for the Advancement of Science and was recently elected Academician of Academia Sinica.  Currently, she is serving her last year term as co-editor of the Journal of American Statistical Association. 

Title: Federated Survival Analysis: Challenges, Opportunities, and Solutions

Abstract: Federated Survival Analysis (FSA) aims to conduct survival analysis using decentralized health data sourced from various medical institutions. FSA empowers individual medical institutions, referred to as clients, to enhance their survival predictions while preserving privacy. In this talk, we will delve into the challenges encountered in FSA due to the presence of non-linear and non-i.i.d. data distributions among clients, as well as the biases associated with censoring. We will present potential solutions to address these challenges by introducing a novel deep learning framework for FSA, known as FedPseudo. FedPseudo harnesses deep learning models to learn robust representations from non-linear survival data, utilizes newly designed federated pseudo-values to address non-uniform censoring, and employs Federated Learning algorithms like FedAvg to facilitate the learning of model parameters. Additionally, we will showcase how FedPseudo can be extended to handle Competing Risks Analysis in the Federated Learning settings. We will present empirical results obtained from experiments conducted on both synthetic and real-world survival datasets and showcase FedPseudo's performance across different censoring settings.

Bio: Sanjay Purushotham is an Assistant Professor in the Department of Information Systems at the University of Maryland, Baltimore County (UMBC). His research interests are in artificial intelligence and machine learning, and their applications to health and climate sciences. He received his Ph.D. in Electrical Engineering from the Viterbi School of Engineering at the University of Southern California (USC). Before joining UMBC, he worked as a postdoctoral research scholar in the Department of Computer Science at USC. He has published more than 75 peer-reviewed publications at leading AI/ML venues such as AAAI, ICML, KDD, and NeurIPS, and has received best paper and poster awards for his contributions. He is currently serving as an editor for the special issue "AI for Healthcare" for the Asia Pacific Signal and Information Processing Association (APSIPA) journal. His research activities have been supported by NSF, NASA, NIH, and ARL, and in 2023, he was awarded the prestigious NSF CAREER award. His current research work focuses on advancing the trustworthy federated learning ecosystem for computational healthcare.