The success and wide-range applications of artificial intelligence (AI) technologies and, in particular, deep learning neural networks methods have led us to a clear understanding of two main problems: the problem of errors (the reliability problem) and the problem of explicitly explaining the decisions made by AI (the explanability problem). These problems are closely related: unexplained AI errors can happen again and again. This is completely unacceptable from the perspective of AI applications in health care because it is critical to the lives and health of patients. If left unresolved, the problems of error and explainability can lead to the rejection or significant restriction of AI systems in medical applications. In this paper, we discuss the problems of explainable artificial intelligence (XAI) for medicine and consider different approaches to solving them.

Auscultation is a classic method of examining patients with respiratory and cardiovascular pathologies. Auscultation is a subjective method, its diagnostic accuracy is highly dependent on the doctor’s experience. Electronic stethoscopes can increase the volume of audio recordings, eliminate noise, and store and transmit sound to a computer or smartphone. Wavelet transform, Butterworth filter, low and high pass filters are used to filter the resulting audio recordings. Machine learning methods, which often surpass to experienced doctors in accuracy, are used to identify various sounds. Methods of mathematical analysis make it possible to differentiate pathological sounds from and innocent heart murmurs, wheezing in the lungs, asthmatic breathing and other pathologies. This review describes various studies on the diagnosis of respiratory and cardiovascular pathologies based on auscultation data.

Background. Long-term outcomes of screening programs are challenging to evaluate in randomized clinical trials. The role of predictive modeling is becoming increasingly popular in oncology. Modeling the interventions consequences in oncology is based, among other things, on the use of toolkits, denoted by the term «mathematical oncology»

Aim. To study approaches to modeling screening scenarios for breast cancer, aimed at developing tools to support medical decision-making in the healthcare system, including the development of clinical guidelines for cancer screening.

Methods. The search for relevant studies was performed through PubMed (Medline) and direct google-search. Key words for the search included breast cancer», «screening», «modeling», «oncology informatics», «cancer care», «big data» etc.

Results. We analyzed several breast cancer screening models. Results of the modeling included broad spectrum of clinically and economically parameters relevant for the screening scenarios characterization. The basic concepts of constructing valid models, including the analysis and simulation of individual histories of the tumor progression course (both natural and in interventional settings), were studied.

Conclusion. Simulation modeling allowed linking new advances in cancer research with the most effective strategies for implementing them into clinical practice in order to maximize patient benefit and reduce economic burden at the population level.

Currently, there are no unified methodological approaches to diagnose certain diseases, syndromes, symptoms, where distant diagnostics and telemedicine-based treatment prescription are legal and feasible. There are limited and empirically generated lists of pathological conditions, accounting for possible risks. The development of regulatory and legal support for telemedicine-based medical care should be based on a scientific approach, providing for safety and quality assurance.

Objective: to develop a methodology for determining the probability of a positive outcome applicable for assessing the possibility of diagnosing and treating patients using distant interaction between healthcare professionals and patients (legal representatives) via telemedicine technologies.

Materials and methods. The study used the principles of a systematic approach. Regulatory and legal acts in the field of organization and provision of medical care were reviewed, including telemedicine-based medical services; duly approved standards of medical care; medical service flowcharts; and duly approved clinical guidelines. Analytical methods (induction, analysis, and synthesis), the method of direct placement for determining weight coefficients, as well as mathematical modeling were applied.

Results. The basic concept has been elaborated as follows: the possibility of making a diagnosis during a direct-to-patient telemedicine consultation should be determined mathematically (by calculating the risks) based on the volume and quality of data on the health status of a given patient. The concept was developed in stages: 1) development of a system of criteria for assessing the volume and quality of medical data; 2) determination of the context and methodology for using the criteria system; 3) design of a mathematical model. The methodology was intended for the provision of primary healthcare outside the healthcare facility or in outpatient settings.

Conclusions. A specific methodology was developed to assess the feasibility of distant diagnosis and effective treatment prescription, with prediction of a positive outcome in a given clinical situation. The methodology includes a system of criteria for assessing the volume and quality of medical data, requirements for the context of clinical application, and an original mathematical model. The methodology can be applied in experimental legal regimens green-lighted for the development of digital healthcare and telemedicine technologies.

Abstract.

Background. Considering the growing interest of researchers and clinical specialists in algorithms for processing medical data, the prospects for the applied application of such approaches have significantly increased, primarily, involving the use of deep neural networks in the tasks of detecting pathological areas. However, the use of such approaches is associated with a low level of localization accuracy, insufficient to translate the developments into the field of assistive systems for making medical decisions.

Aim. This work is aimed at assessing the speed and accuracy of the modern architecture of the convolutional neural network RFCN ResNet-101 V2 for the prospects for automated processing of clinical data from coronary angiography.

Materials and methods. The basis for the chosen neural network architecture training was the clinical graphic data of 50 patients subjected to routine coronary angiography, which is characterized by the presence of single-focal lesions (stenoses) in more than 75% of all cases. The study evaluated the metrics of classification and localization accuracy in determining the position of a single-focal coronary artery lesion.

Results. The utilized architecture of the neural network was capable of detecting single-focal lesions with an accuracy of 94%. However, to a large extent, it didn’t the performance requirements (processing speed).

Conclusion. The results obtained determine the further direction of development of the presented approach, which should be reducing the time of analysis of each frame of coronary angiography due to image preprocessing methods.

The article proposes technique and results for pre-project research and design of automated information systems in relation to business process management systems of medical organizations on the example of monitoring the implementation of regulations at the departments of nephrology and hemodialysis. When developing the system, an original method of system analysis and modeling of business processes was used based on the experience of the divisions of a private network medical company working in the field of nephrology and hemodialysis. Fragments of the ontological model, a scheme of the structure of the system’s goals, diagrams of cause-and-effect relationships, a strategic map of the hemodialysis department are given. A description of the system is given in the form of landscape diagrams and process choreography in BPMN 2.0 software notation. The service-oriented architecture of the system is shown.

The article considers the provisions of the draft convention on improving supranational legal regulation in the Eurasian Economic Union (EEU) in the field of healthcare in the context of the development of innovative digital technologies. Specific proposals on the content of the project’s draft were formed, as well as individual proposals were identified on separate convention sections.