Please use this identifier to cite or link to this item:
|Title:||Analysis of irregularly sampled time series health care data sets||Authors:||Wang, Anni||Keywords:||Engineering::Electrical and electronic engineering||Issue Date:||2021||Publisher:||Nanyang Technological University||Source:||Wang, A. (2021). Analysis of irregularly sampled time series health care data sets. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/154677||Project:||D-255-20211-02961||Abstract:||The real-world healthcare system generates abundant time-series data. In most cases, these data have a high prevalence of missing values and are often irregularly sampled across both time and patient. Moreover, due to the complex level of a different dataset, the preprocessing is more significant and challenging. This dissertation focuses on imputation and prediction tasks to address the challenges of irregularly sampled time series data sets. First, we trained the Recurrent Imputation for Time Series (RITS) model and Bayesian Long Short Term Memory (BLSTM) model on a publicly available PhysioNet dataset for prediction task only. Next, we, train a Bayesian LSTM for imputation of missing values (Considering the irregular sampling as missing values too) and prediction of outcomes, on a proprietary heart failure risk prediction data set. The proposed model represents the distribution of irregularly sampled time series data, imputes both categorical and continuous missing data in the time series, and makes prediction of the outcome of interest. Furthermore, the Bayesian model allows for a reliable estimate of the outcome of interest. While the missing continuous variables are imputed through the MAE error minimization, the categorical variables are imputed using softmax and/or argmax operations. The outcome prediction task in the presence of imbalanced data set is addressed through the weighted loss function. Performance results indicate that the proposed approach is effective in imputing both categorical and continuous variables, with the superior prediction of outcome of interest.||URI:||https://hdl.handle.net/10356/154677||Fulltext Permission:||restricted||Fulltext Availability:||With Fulltext|
|Appears in Collections:||EEE Theses|
Updated on May 15, 2022
Updated on May 15, 2022
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.