Statistical Machine Learning on Time Series with Applications to Manufacturing and Healthcare
Year: 2023
Type: dissertation
Abstract: Predictive analysis in time series is critical to making informed decision-making and has been extensively studied in a wide range of fields including manufacturing and healthcare. The temporal patterns contained in time series hold the promise of making predictions on future behaviors as those patterns often reflect the health state evolution. However, developing such predictive models is a challenging problem because (1) the actual health state is typically not directly observable and can only be assessed from the health condition monitoring data (e.g., sensor measurements of the engineering system, physiological signals of the patient); (2) the health state transition may depend heavily on historical states due to complicated failure mechanisms or complex disease progression; (3) the health evolution trends generally exist significant variability due to some critical exogenous factors (e.g., operational and environmental conditions of the engineering systems, gender and age of the patients) and heterogeneity; and (4) when making predictions at an early stage, the temporal patterns may be too subtle to be detected by traditional methods (e.g., abnormality in early ECGs associated with future cardiac diseases). This dissertation addresses these challenges by developing novel and flexible statistical models and leveraging machine learning algorithms, and investigating the performance of the proposed methods in simulation studies as well as real-world applications. We first develop a flexible prognostics framework based on a higher-order hidden semi-Markov model (HOHSMM) for systems or components with unobservable health states and complex transition dynamics. The HOHSMM extends the basic hidden Markov model (HMM) by allowing the hidden state to depend on its more distant history and assuming generally distributed state duration. An effective Gibbs sampling algorithm is designed for model inference and is evaluated by a simulation study. We design a decoding algorithm to estimate the hidden health states and use a simulation approach to predict the remaining useful life for a functioning system in operation. A case study on the NASA turbofan engines demonstrates the practical utility of the proposed prognostics framework. To further capture the variability among time series due to some critical factors and patient-level heterogeneity, we propose a novel and flexible Bayesian mixed-effect higher-order HMM (MHOHMM) to predict patients' future events based on their electronic health records. A range of MHOHMMs are designed to capture different data structures and the optimal one is identified by using the $k$-fold cross-validation. An effective two-stage MCMC sampling algorithm is designed for MHOHMM inference. A simulation study is conducted to evaluate the performance of the proposed sampling algorithm and the MHOHMM-based classifier. The results show that considering the impacts of mixed effects and higher-order transition dynamics significantly improves prediction performance when such impacts present in the data. Furthermore, we conduct a case study on acute hypotensive episode prediction using real-world clinical data to demonstrate the effectiveness and advantage of the proposed prediction framework. For problems of making predictions at an early stage, we adapt state-of-the-art deep learning techniques to detect subtle patterns. To investigate whether mutant mice present subtle abnormalities in their early ECGs before severe arrhythmias occur, we propose a deep transfer learning (TL) model, DeepMiceTL, to predict cardiac conduction diseases using early mouse ECGs. We further apply the Bayesian optimization and $k$-fold cross-validation methods to tune the hyperparameters of the DeepMiceTL. Our results show that DeepMiceTL achieves promising performance to identify ECG patterns during the early course of cardiac conduction disease in mice. Our study also suggests the feasibility for early clinical diagnosis of human cardiac conduction diseases and other types of cardiac arrhythmias using deep learning in the future.
Author: Liao, Ying
Language: English
Subfield: Safety, Risk, Reliability and Quality
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7134143582