The thesis solves a scientific problem in the field of instrumental support of bioinformatics - the development of theoretical foundations, improvement of methodological, algorithmic, software, and technical basis of the computer systems for processing of biosignals and data based on the extensive use of artificial neural networks and deep learning technologies.
Current state and future perspectives of machine learning usage in the computer bioinformatics systems are analyzed in the thesis. It is shown that the heterogeneousness of data and a wide range of bioinformatics tasks influenced the development of specialized solutions for each separate domain or application. This complicates the possibility to compare the effectiveness of certain methods as well as the usage of the best system design variants for the new tasks.
A novel framework related to the development of principles for the design of the biosignals computer processing systems involving a combination of machine learning techniques and digital signal processing is presented in the thesis.
The expediency of separation of the system levels within the process of biosignals processing is reasoned, and their functions are outlined. Innovativeness of the suggested approach lies in the separation of functions of the lower, middle, and upper levels from methods with the help of which they are realized, as well as from the implementation variants for these methods based on the hardware and software components. The middle system level is significantly invariable both in regards to the task to be solved and to the biosignal type. At the same time, the upper level is specific as to the final application, and the lower level is specific as to the type of biosignal. Distinct outlining of functions for each system level and the inter-level interfaces opens perspectives for information structuring during the analysis of the known decisions, which simplifies the analysis and comparison of the effectiveness of these solutions. The design process of the computer system for the specific tasks gets simplified and potentially quickens due to the possibility of transferring the best results among the related tasks. On the basis of the developed three system levels concept the range of tasks related to machine learning application and biosignals processing on all the system levels was studied and analyzed.
A novel method of optimal hyperparameters selection for a multilayer neural network classifier based on the Monte Carlo method and predictive modeling was developed and introduced in the thesis, new algorithms for detection and correction of anomalies in ECG-signals were presented as well. This resulted in reduction of hyperparameters optimization time by 4 orders compared to a grid search approach in the entire hyperspace of possible values.
A new approach to the detection and correction of residual anomalies in biosignals was developed and tested. This approach is based on the use of neural network autoencoders for nonlinear filtering of distortions that are located in the same spectral band as the useful signal and its application reduces identification error by 5-7 times.
A method of software conditioning of biosignals was developed, which enabled parameters unification of digital records of biosignals from open databases by resampling, rescaling, time normalization, etc. and its application improves the issue of low data volume while deep neural networks training.
A comprehensive study of the influence of the variability of ECG signal acquisition systems (different data sets, sampling frequency, recording duration, data volume) on the accuracy of biometric identification was performed. The obtained results proved sufficient stability and reliability of the electrocardiogram as a biometric marker and confirmed the possibility of its real-world application in biometric identification systems.
Neural network equalizer was designed for dynamical error correction of bioimpedance sensors, which resulted in expanding the operating frequency band up to 100 times.
Functionality was expanded, metrological characteristics were improved and the speed of the digital rheograph was doubled by improving the Howland current pump by compensating the basic bioimpedance with a coded potentiometer (trimmer) and using a direct digital synthesis of orthogonal probing signals.
Intelligent processing of vibroarthrography signals based on a combination of machine learning algorithms and wavelet decomposition was developed, which allowed to achieve better accuracy in terms of knee joint disorders analysis.
The results of the work can be used in various applications, such as cybersecurity and access systems (biometric identification), robotics (bio-machine control interfaces), and affective informatics (psycho-emotional state analysis), as well as a medical domain (diagnostics, clinical trials).
