Work

Data analytics and mathematical modeling to advance cardiac care

Year: 2021
Type: dissertation
Abstract: Cardiac disease is the leading cause of death in the United States and results in a great amount of medical cost every year. The low cost-effectiveness of cardiac care is in part due to insufficient understanding of disease mechanisms and suboptimal treatment planning and post-operative care. To improve the quality and reduce the cost of cardiac health care, advanced computational techniques, i.e. data analytics and computational modeling, have been widely used in clinical applications. The former can extract valuable information from the medical dataset to assist decision making, while the latter can offer mechanistic insights into disease mechanisms. Although significant research efforts have been made in these two fields, there are still research gaps need to be filled to improve the quality of cardiac care. Therefore, this dissertation aims to address several research gaps in the fields of predictive analytics for post-transplant cardiac care and computational modeling of cardiac electrophysiology. Specifically, the following four tasks are investigated: \n1) There is a lack of risk-stratification model to guide skin cancer screening after heart transplantation. Many research efforts have been conducted to identify risk factors associated with non-melanoma skin cancer (NMSC) after heart transplantation. However, few risk-stratification models have been developed to guide the screening of NMSC. In addition, most of the available risk assessment tools were developed based on small cohorts of recipients for specific types of transplanted organs, which may not be applicable to heart transplant recipients. To address this research gap, the data of heart-transplant recipients are queried from a national organ transplant database, i.e. the United Network for Organ Sharing registry, based on which a risk prediction model was developed for the screening of Basal Cell Carcinoma, one type of the NMSC. \n2) There is a lack of efficient algorithm to impute the missing values in Electronic Health Records (EHR) and exploit the imputed information for predicting Heart Failure (HF) readmission. The prevalence of missing values in EHR system poses great challenges for data analysis and clinical applications. To deal with this issue, there are three commonly used approaches. The first and simplest approach is to delete the instances that contain missing information and conducting analysis with only the complete cases, which might overlook a lot of useful information. The second approach is to impute the missing values with some straightforward methods such as last observation carried forward or mean imputation, which could not provide accurate estimation and might even affect subsequent analysis. The last approach is to impute the missing values using some advanced statistical or machine learning methods. It should be noted that the missing values could not be perfectly estimated but with some uncertainty. However, the available methods could not provide easy quantification of such uncertainty. What’s more, the uncertainty should be included in the subsequent analysis to improve model performance. To address this research gap, Gaussian Process Latent Variable Model is applied to impute the missing values and a constrained support vector machine is further developed to make full use of the imputed information for predicting HF readmission. The proposed method was tested using Intensive Care Unit (ICU) data and showed better performance than existing models. \n3) To enable the personalization of cardiac model, some statistically sound and effective algorithms need to be developed. Although personalization of cardiac electrophysiological models has been studied in the literature, most of the existing methods focus on the estimation of one or some of the model parameters and leave the other unchanged, and the uncertainty of the estimated parameters was rarely quantified. My research intends to develop a Bayesian optimization method to address this research gap. Specifically, a two-stage optimization technique is developed to estimate the cardiac cell membrane potentials, and the posterior distribution of model parameters are calibrated by minimizing the estimation errors. The estimation accuracy and computational efficiency of the proposed method were verified through multiple sets of simulated data. \n4) There is a lack of efficient and accurate uncertainty propagation method to incorporate uncertainty in multi-scale cardiac models. Uncertainty is ubiquitous during the process of model development, which can originate from model assumption, calibrated parameters using noisy data and cell-to-cell variability. Although many research has been conducted, there are still limitations. Some studies adopted the Monte Carlo (MC) method, which is easily to implement but computationally inefficient for cardiac models. Others used the generalized polynomial chaos (gPC) expansion, a non-sampling-based technique, but only applied it to the ionic-channel model. To fill this gap, my research used gPC to propagate parametric uncertainty across different organizational levels of ionic channel, cell, and tissue in a multi-scale cardiac model. The efficiency of the propagation was realized through the identification of significant parameters within each level of the model. The uncertainty propagation results are compared with that from MC to verify the accuracy and efficiency of the proposed method. \nThe methodologies and tools developed in these research tasks can improve the effectiveness of medical decision making using EHR and provide a better understanding of cardiac disease mechanisms.
Author: Zhiyong Hu
Language: English
Open Access status: green
OpenAlex ID: W3202106343
Cites: 145
Cited by:
Related to: 9