Studies On Hierarchical-Factor-Graph-Based Automated ECG Diagnosis

Electrocardiogram (ECG) is the record of variation of bioelectric potential with respect to time as heart beats. It is one of the most important tools for diagnosis of cardiovascular diseases in clinical applications. Research of automated ECG diagnosis is significant for improving quality of accuracy and real time of ECG diagnosis and reducing labour intensity of physicians. On the basis of analysis of the characteristics of ECG diagnosis problem and the limitations of existing methods, a novel method based on hierarchical factor graph (HFG) is proposed for automated ECG diagnosis in this thesis. And works focus on several critical issues, including basic theory of HFG, HFG construction and inference for automated ECG diagnosis, and ECG signal preprocessing, feature extraction and diagnosis during the course of automated ECG diagnosis.The main contributions of this thesis are summarized as follow:(1) A novel model for knowledge representation and processing, namely, hierarchical factor graph, is proposed. Methodologies for HFG inference and construction are also studied. HFG is an extention of stardard factor graph. It has all merits of existing probabilistic graphical models and important characteristics as follows: First, local functions are introduced to describe relationships among variables in problem domain, so that more independence relationships can be represented in this model and different kinds of knowledge can be brought into an entire framework and dealt with uniformly. Second, hierarchical structure is introduced in this model to represent structural knowledge in problem domain, which provides a convenient and effective approach for modeling and inferring of complex system hierarchically. HFG inference can be performed by transforming HFG to standard factor graph and applying inference algorithms of factor graph. Methods for transforming HFGs to factor graphs are presented in this thesis. On the basis of study on the exiting inference algorithms of factor graph, a approach based on sum-product algorithm is proposed for computing marginal functions of arbitrary multi-variables in HFG inference.(2) The methodologies of HFG construction and inference for automated ECG diagnosis are studied. Firstly, according to the characteristics of ECG diagnosis, a strategy for model construction with combination of domain knowledge and sample data is proposed and hierarchical construction method based on top-down idea is adopted to construct HFG for automated ECG diagnosis. Then, on the basis of flow of automated ECG diagnosis, the basic structure of HFG for automated ECG diagnosis is identified by analyzing the main variables in automated ECG diagnosis and the relationships among them. Methods to specify local functions in HFG for automated ECG diagnosis are also proposed. Finally, main tasks and plans to perform inference in HFG for automated ECG diagnosis are identified by analyzing the process of inference according to the basic structure of HFG for automated ECG diagnosis.(3) Noise HFG based methods are studied for ECG signal preprocessing. With the goal of constructing noise HFG, the variables and their relationships are analyzed, and main tasks are identified in ECG signal preprocessing. Then, studies focus on the application of morphological filter methods to ECG signal preprocessing. A method based on multi morphological filter operations is presented for removing baseline drift in ECG signal. Another method based on morphological operations and adaptive threshold is presented for denoising electromyogram noise in ECG signal. Noise HFG is constructed based on the proposed methods for ECG signal preprocessing and the messages involved in inference are computed.(4) Feature HFG based methods are studied for ECG feature extraction. Main tasks in ECG feature extraction are analyzed and identified according to the goal of constructing feature HFG. Feature extraction methods for QRS complex, ST-T segment, and P wave are studied respectively. For QRS complex, a multi-structuring-element morphological approach with feedback correction is brought forward for QRS detection, and a method based on curve analysis is proposed for characteristic point location and morphological automatic analysis. For ST-T segment, a method for characteristic point location of T wave based on local distance transform and a method for ST segment shape analysis are proposed. For P wave, a novel method based on location estimation and recognition post-processing is proposed for P wave detection and location. Feature HFG is constructed based on the proposed methods for ECG feature extraction and the messages involved in inference are computed.(5) Disease HFG based methods are studied for ECG disease diagnosis. Due to the complexity of diseases, a strategy based on decomposition idea is presented for model construction and problem resolution of disease HFG. Methods are studied respectively for knowledge acquisition, identification of variables and their values, and identification of topological structure and local functions of disease HFG. And disease HFGs for normal ECG diagnosis and myocardial infarction ECG diagnosis are constructed completely. Methods are discussed and determined for computing messages in Feature HFG and performing inference in HFG for automated ECG diagnosis. Finally, real ECG records of PTB diagnostic ECG database are utilized as diagnostic instances to validate the proposed models and methods in this thesis. Experimental results show that the automated ECG diagnosis method based on HFG proposed in this thesis can solve ECG diagnosis problem correctly and efficiently.