Chernetchenko D. The method and hardware-software implementation of electrocardiac signals processing using artificial multistable neural networks

The object of the study is the process of electrocardiography (ECG) signals processing using artificial multistable neural networks. The purpose of the study is to improve the processing efficiency and accuracy of detection of the common features of the ECG-signal’s P-QRS-T-complex by developing a method and hardware-software implementation based on artificial multistable neural networks. Research methods: discrete information processing and methods of mathematical statistics were used for data processing; development of the structural schematic of system and the hardware-software implementation were performed with the basic principles of systematic approach; mathematical modeling and artificial intelligence approach with help of NEURON and MATLAB software packages were used to develop artificial neuronal models and neural networks; microcontroller embedded software was built using the IAR Embedded Workbench environment, VHDL programming language and Xilinx ISE WebPack software package were applied for design the FPGA-based system; electrical schematic and hardware printed circuit board are designed using Altium Designer software environment; the bibliosemantic method used to study worldwide and national content. Theoretical results: first developed an artificial spiking neural network (SNN), which consist of input encoder, internal layer of recurrent neurons and output neuronal layer, which is a self-learning classification system that automatically adapts to changes in the input raw-signal and provides real-time processing of clinically relevant cases of ECG-signal; first developed an ECG-signal processing method based on the SNN approach, which provides data density reduction by direct encoding of the ECG-signal into a sequence of spikes, also shown that improvement of a whole network stabilization at the certain stable electrical states corresponding to the ECG-signal’s peak moments, pre-processing and filtering of the input raw-data leads to the significant reduction of recognition error; an adapted model of an artificial spiking neuron was investigated and used as a basic component of the neuromorphic module with providing electrical property of multistability, which allows to reproduce all patterns of electrical activity of biological neurons and therefore, increases memory and computing power of neural components; the model and structure of the spiking encoder for precise coding of ECG-signals was improved, which provided the necessary and sufficient background for creating an artificial spiking neural network and the hardware-software implementation for the real-time processing of ECG-signals. The practical results include the development of hardware and software for ECG-signal processing, production on a modern element base, protection of method and circuit design solutions by USA patents, validation and experimental testing of the produced device, accuracy estimation of classification of the common amplitude and temporal ECG signatures on a sample of records. Degree of implementation: the results of the dissertation were implemented in the production process at the company “Scientific and Production Enterprise “SMD” LLC; in the educational process of the Department of Experimental Metal Physics of Oles Honchar Dnipro National University to teach relevant disciplines. Scope (field) of application - medicine.