QuBBD: Classification and clustering of medical time series data: the example of syncope

点击翻译按钮获取中文摘要

Syncope or, more colloquially, fainting, is a surprisingly common and little understood condition. The long term goal of the proposed line of work is the identification of the root causes of syncope. Fainting can result from the failure of one or more internal control mechanisms. There are currently no clear causal links between these controls and the observed symptoms. In order to understand the involved mechanisms, this project will start by analyzing clinical data to determine the number and characterization of the different types of syncope. A better understanding hinges on the analysis of clinical data, here time series, and the ability to infer from these, patient classification. Various scenarios will be tested through mathematical modeling to confirm both the soundness of the obtained classification and the nature and source of the pathology for each identified class. The ability to identify subjects as members of a class or group also makes it possible to leverage information about the other members of that group for individual diagnosis purposes. The methodology developed here will contribute to the implementation of this approach, sometimes referred to as "bringing cohort studies to the bedside". This approach will also be applicable to the study of other diseases where similar clinical data are being collected such as epilepsy.Time series data are ubiquitous. In fact, the development of an ever increasing number of applications depends on their analysis, from stock market and economics to weather predictions. Practical issues include signal matching, classification, pattern detection and early prediction. Signals of interest are often high dimensional and noisy and the underlying dynamics are generally unknown. This award supports initiation of a collaborative research project that addresses all the above issues in the context of medical times series data and, more specifically, for syncope. To do so, a novel paradigm combining non-parametric statistics, machine learning, and applied mathematics is proposed. The first objective is to design and compute features in multivariate time series that are well adapted to classification and clustering. The proposed approach relies on recent advances in the evaluation of variable importance for scattered data. The second objective is to facilitate calibration of mathematical models by extending machine-learning concepts to this new realm. The goal is to determine how much data is necessary to "learn" biomathematics models and increasing their predictive power. This award is supported by the National Institutes of Health Big Data to Knowledge (BD2K) Initiative in partnership with the National Science Foundation Division of Mathematical Sciences.