Application of Artificial Intelligence Methods Depending on the Tasks Solved during COVID-19 Pandemic

Abstract

Health systems challenges that emerged during the COVID-19 pandemic, such as a lack of resources and medical staff, are forcing solutions which optimize healthcare performance. One of the solutions is the development of clinical decision support systems (CDSS) based on artificial intelligence (AI). We classified AI-based clinical decision-supporting systems used during the pandemic and evaluated the mathematical algorithms present in these systems. Materials and methods: we searched for articles relevant to the aim of the study in the Scopus publication database. Results: depending on the purpose of the development a clinical decision support system based on artificial intelligence during pandemic, we identified three groups of tasks: organizational, scientific and diagnostic. Tasks such as predicting of pandemic parameters, searching of analogies in pandemic progression, prioritization of patients, use of telemedicine are solved for the purposes of healthcare organization. Artificial intelligence in drugs and vaccine development, alongside personalized treatment programs, apply to new scientific knowledge acquisition. Diagnostic tasks include the development of mathematical models for assessing COVID-19 outcomes, prediction of disease severity, analysis of factors influencing COVID-19 complications. Conclusion: artificial intelligence methods can be effectively implemented for decision support systems in solving tasks that face healthcare during pandemic.

1. Introduction

In the context of the spread of COVID-19, national health systems around the world are under serious pressure. Most countries face the same challenges: lack of medical staff, overcrowding of hospitals, shortage of equipment and facilities for diagnosis and treatment, forced suspension of routine services, etc. According to the World Health Organization, about 40% of patients have a mild disease and do not require hospitalization, 40% of patients have moderate symptoms that may require hospitalization, 15% of patients have a severe disease requiring oxygen therapy and other medical treatment events in the hospital, and 5% of patients develop extremely severe pathologies requiring mechanical ventilation [1].

Another important aspect in the development of the pandemic is the high risk of infection for healthcare workers in primary healthcare centers and hospitals. Up to 10% of reported disease cases in China and up to 9% of all disease cases in Italy are among healthcare workers. In some EU countries, this rate reaches 26% [1,2].

With a sharp increase in the number of patients with coronavirus, a high incidence among medical workers has revealed the unpreparedness (unreadiness) of healthcare systems in many countries to work in a pandemic. In this situation, the introduction of artificial intelligence (AI) methods in the organization of medical care will significantly reduce the burden on medical personnel and help optimize resources.

Currently, there is a technological potential for the introduction of modern information technologies in the work of medical facilities in order to improve the quality of medical care. The aim of this study was to identify the possibilities of using artificial intelligence for clinical decision supporting systems (CDSS) during a pandemic.

2. Materials and Methods

The relevance of CDSS is evidenced by the results of search queries to the Scopus publication indexing system. Thus, the clinical decision supporting system query procured a total of 56,738 articles. Further, within the framework of the obtained search results, an additional search was carried out for information related to the application of machine learning and artificial intelligence methods within the framework of the CDSS. Based on the additional phrase “machine learning”, 7825 articles were filtered. Further, a pair of synonyms was added to the search query, describing belonging to research within the COVID-19 and SARS-CoV-2 pandemics with additional filtering by 2020 and 2022 publication years. After submitting a request to the indexing systems, we received 724 articles that were analyzed to determine the goal relevance for the review. Finally, only 92 original research articles were included finally to the review paper.

The artificial intelligence methods found during the search were classified depending on the direction of the practical problem being solved within the construction and application of the COVID-19-related CDSS.

The citation indexes of the journals were not taken into account when searching for the original research papers. This could be the possible limitation of the study.

3. Results and Discussion

Automation of processes aimed at solving the problems of medicine and healthcare during a pandemic has been investigated in hundreds of works. At the same time, a variety of publications can be summarized for solving organizational, scientific and diagnostic problems (Figure 1).

3.1. Organizational Task

AI-based CDSS, solving the problems of medical care organizing.

The first group included tasks related to the organization of medical care in a pandemic, namely patient routing, resource administration (drugs, bed resources, medical devices), predicting the number of cases, and the prediction of recovery. It is for this group of tasks that the use of AI methods is the most promising. An analysis of the publications made it possible to identify the following areas of AI application:

3.1.1. Predicting the Global Parameters of the Pandemic

The aim of such system is to make decisions in the field of planning the load on the medical infrastructure and to make decisions on its adjustment, taking into account changing predicted values. It is not a classical AI-based CDSS for the management of individual patients, however, it serves to plan and administer global health parameters in a pandemic. These predictions make it possible to reserve the need to create medical resources to counter a pandemic and prevent situations in which the existing healthcare system is unable to accept patients in the quantities that arise during the development of a pandemic.

Actually, almost all of these methods are aimed at analyzing the time series of morbidity/recovery/mortality parameters to construct predictive models. Another specific task is to discover the country regions with similar parameters of pandemic development via clusterization methods [3]; DNN and decision trees [4] describe such areas by some common features and build analogies for the pandemic development prognosis.

Moreover, we include in this group a method of monitoring the local environment to detect violations of the regime and the presence of symptoms, which also allow us to monitor the overall situation, deal with violations of the regime and predict their extent and consequences. The local environment (humidity, temperature, wind) influence on the number of COVID-19 cases also can be analyzed by AI methods [5].

Methods of regression analysis including support vector regression are used to make a continuous prediction [6,7,8,9,10,11], while SVM [7,8], fuzzy inference system [12], varieties of recurrent neural networks [13,14,15,16,17] are used to build predictive models with discreet output. ANN model [18], Genetic Programming (GP) algorithm [19], Susceptible-Exposed-Infectious-Removed model (SEIR) combined with the AI model [20], statistical Auto-Regressive Integrated Moving Average (ARIMA) method [21] and Naïve Bayes [22] are also suitable in predicting the number of confirmed, deceased, and recovered cases and the estimation of epidemiology curve.

3.1.2. Research Identifies Similar Variants of the Disease Development When Comparing the Dynamics of Patients’ Treatment

Based on these, data comparison is made between countries in terms of the course of the disease in infected individuals.

For solving this task, clustering methods are applied, such as K-means, or Fuzzy PROMETHEE (Preference Ranking Organization METHod for Enrichment of Evaluations) and VIKOR (from Serbian: VIseKriterijumska Optimizacija I Kompromisno Resenje, that means: Multicriteria Optimization and Compromise Solution) methods. The input parameters are tied to the problem being solved (depending on the type of searched analogy and the parameters of the region or patients) [3,8].

3.1.3. General Review Papers on the Use of Telemedicine and AI within It Contain a Minimum of Specific Information on Solutions within the AI-Based CDSS

These papers are the most significant ones in terms of the number of works, therefore they are mentioned as a part of this review. This section included a number of works that, with some specifics (algorithms, results), could be attributed to other methods [23,24,25] as well as specific software solutions without describing their content and implementation [26]. These works confirm the applicability of AI methods for data analysis, but they are not allowed to be compared with similar tasks. There is the task of diet correction or disease monitoring using telemedicine when the in-person treatment is impossible under the restrictions associated with COVID-19 [24,26].

3.1.4. The Task to Prioritize Patients for Selected Treatments Based on Their Individual Characteristics (e.g., Blood Types) and the Severity of the Current Condition [27,28,29]

The input data mainly includes diagnostic parameters (laboratory, X-ray data). Due to high workload of clinical laboratories an automated methods are needed even for such a trivial task as blood type analysis and machine learning methods have a great potential in this field. In paper [28], a deep convolutional neural network was trained to classify images of a blood test samples after applying specific antigens to determine the blood types [28]. An estimation of risk factors of COVID-19 severity allows us to prioritize patients providing them a better treatment. Machine learning methods such as regression, random forests (RF), support vector machines (SVM), and artificial neural networks (ANN), which can solve feature importance problem along with predictive model construction are used [27,29].

We include in this group methods that allow us to predict disease severity thus aiding us in better resource planning and allocation. Fusion modeling using medical history and current treatment data can forecast the need for hospitalization for patients infected with COVID-19 at the time of the RT-PCR test [30]. Accurate prediction of COVID-19 severity is possible after analyzing clinical data with 5-layer deep neural network (DNN) [31]. Deep learning models with an autoencoder-based approach estimate the spread of the disease and predict the survival possibilities of the quarantined patients in isolation [32].

3.2. Research Tasks

The second group includes research tasks aimed at obtaining new knowledge through the analysis of complexly structured data, a large amount of heterogeneous data and intended to predictive assessment of health status, improving the quality of medical care. We included the following subtasks in this group:

Application of Machine Learning Methods in DNA Analysis for Mutation Detection, Personalized Diagnosis and Treatment, Treatment Efficacy Prediction [33,34], Vaccine Development [35] and Drug Development [36]

The input data are either a sequence of genes [37,38] or the structural chemical composition of the studied substances in the form of a sequence [36]. For example, in [39] 3C-like protease inhibitors were created by a novel generative adversarial network. For gene analysis, statistical methods such as Markov models or Monte Carlo methods is used to build sequences from available samples [34,39].

The output is possible for the next elements of the gene sequence, genomic biomarkers that affect the course of diseases, or options for potentially effective drugs.

3.3. Diagnostics Tasks

This group includes approaches aimed at assisting the physician in making a diagnosis, prescribing drugs, and providing hardware assistance. Despite the objective importance of diagnostic tasks, we left them only the third place in our classification, since the decision about the patient’s condition, the conclusion about treatment administration, in any case, should be made by the physician. There are following subsections within the framework of AI participation in diseases diagnostics.

3.3.1. Diagnostic Models Aimed at Assessing the Likelihood of Any Manifestation of the Disease, Its Development and Outcome

The best performance was observed in models based on clinical, laboratory and radiological variables [40]. Within the pandemic, these approaches have been used to predict the outcome of COVID-19 and have shown similar results. However, most of these studies do not include continuous physiological signals (e.g., ventilator parameters, vital signals) for prediction and do not work in real-time. Such information would significantly improve the characteristics of predicting in terms of expanding the classes of problems being solved (not only a single prediction, but also monitoring in dynamics) [41]. A subspecies of diagnostic application is an approach for predicting the development of a disease and developing personalized patient management and treatment plans based on the success of previous patients with a similar prognosis [42].

An analysis of the solutions found showed that the created systems are mainly aimed at predicting survival/death (the metric of belonging to a class can be used to assess the risk of death). In the framework of solving these problems, statistical regression methods and their modifications are most often encountered [27,41,42,43,44,45,46,47,48]. There are also decision trees [10,27,49], K nearest neighbors (KNN) [50], and SVM [46,48]. Neural network algorithms are found, but relatively infrequently due to small amount of annotated data [27]. Methods for constructing decision-making systems based on several classifiers and combining their decisions (ensemble methods [46]) have a mean F1 score of more than 0.8. At the same time, within the framework of the study, both for this one and for other classes, the data sets differ greatly. Standard datasets used simultaneously in many significant works have not been identified. This leads to the fact that quality metrics can be used for comparison within the same work but are poorly applicable for comparison within different works on different data. This makes it possible to assess the general applicability of solutions, but it does not allow us to clearly identify the best ones.

In addition, the analysis showed the predominance of systems for constructing a predictive model for a patient without real-time monitoring of his condition. This can be seen from the prevailing use of input data such as age, gender, chronic diseases (clearly static parameters within the disease), blood test results (not monitored in real time, although it is possible to conduct repeated studies and track dynamics).

The latest works show that AI-based analysis of data from wearable devices (with LTSM-based autoencoder) [51], biosensors (with Fuzzy-based decision tree) [52] and audio data (with Deep learning–enabled tracking tool using gated recurrent units) [53] can accurately differentiate between healthy and COVID-19 infected people and predict recovery mode. Furthermore, custom gradient boosting ensemble learning method makes it possible to predict infection in real time after analyzing the data from wearable devices [54].

3.3.2. One of the Diagnostic Use of CDSS Is Lung Monitoring Based on Classical X-ray and CT Image Analysis Both for Direct Diagnosis of COVID-19 [55] and for Separating Pneumonia into Bacterial and Disease-Induced COVID-19 [56]

This category of tasks uses image analysis methods almost always associated with the use of deep learning methods, namely convolutional neural networks [55,57,58,59,60,61,62,63,64,65,66,67,68,69,70] and its combination with other methods: SVM [71], U-net [72], Feed-Forward and Neural Network [73], random survival forest [74], belief function-based CNN [75]. Extreme learning machine (ELM) method [76] and SVM [57] are also encountered. The use of SVM to extract parameters according to the encoder principle with subsequent application of other methods (KNN, RF, SVM for decision making [56]) seems promising.

The input parameters are the results of radiographs or computed tomography. The output results within the solutions found are the classification results for the absence of pneumonia/non-COVID-19 pneumonia/COVID-19 associated pneumonia. A solution has been found, which allows for a full cycle of diagnostics and monitoring of patient treatment [64]. However, the analysis of CT results is described without a detailed description of other aspects. For this reason, this work is assigned to this class, and not allocated to an independent class.

Adding more information to CT scans such as acoustic data or results of RT-PCR tests gives high diagnostic accuracy after analyzing with Deep Learning-assisted Multi-modal Data Analysis [77] and EfficientNet-B0 initialized on ImageNet pretrained weights [78], respectively.

3.3.3. Analysis of Factors in the Diagnosis of COVID-19 and Complications [79]

CT results are used as input data. Data are manually processed to extract parameters for subsequent analysis and predictive model construction. The outputs of the system are not a mortality prediction, but a metric describing the severity of the disease and the possibility of complications.

The widest data analysis methods are used, such as neural networks [80,81,82], KNN [80,83], regression methods [79,84,85,86], Bayes classifier and its modifications [80,81,87,88], fuzzy systems [89,90], RF [80], SVM [80,82]. There are also approaches based on the ensemble of different methods [91,92,93].

In addition, a system for predicting complications during treatment (the occurrence of acute respiratory distress syndrome (ARDS)) [92] is singled out separately. Initially, the work was included in the review as an example of a system for monitoring the patient’s condition and controlling the ventilation system, however, after a detailed study, it was found that it was aimed rather at identifying a dangerous complication and signaling about its occurrence. For this reason, the work is moved to the existing class. Works directly related to the management of systems for the treatment and monitoring of patients with COVID-19 are not found in the leading indexed publications.

We make a comparative table on the basis of the identified classes of application of AI methods in the construction of CDSS (

Table 1. Applicability of AI methods in the framework of tasks relatedto the COVID-19 pandemic, input parameters used in their solutions and achievable indicators of the resulting models.

Task Class	Input Data	Output Data	Decision Method	Data Sets	Accuracy	Reference
AI-based CDSS for organizational problem solving
1.1 Predicting the Global Parameters of the Epidemic	Time series of sick/recovered patients	The number of sick and recovered patients in 6 countries.	Simple recurrent neural network (RNN), long-term short-term memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Repeating Units (GRU), and Variational Autoencoder (VAE).	From COVID-19 outbreak for corresponding countries (from 22 January 2020 to 17 June 2020)	Root mean square error (RMSE), mean absolute percentage error (MAPE), mean absolute error (MAE), explained variance (EV), and log mean squared error (RMSLE). Separately for 6 countries.	[14]
	Interest in search terms “Corona”, “COVID-19”, “Coronavirus”, “Antiseptic sale”, “Antiseptic purchase”, “Hand washing”, “Hand sanitizer”, “Ethanol”, “Antiseptic” in Persian for the previous day The incidence of COVID-19 the day before (Iran).	Similar parameters for the current day, of particular importance, the incidence of COVID-19 (in Iran).	Linear regression and long-term short-term memory (LSTM).	From 10 February 2020 to 18 March 2020	The mean square value for the linear regression model was 7.562. (SD 6.492). The root mean square LSTM was 27.187 (SD 20.705).	[6]
	Time series of cases/recoveries/deaths (India) Number of new cases, deaths and consequences; vaccinations (α).	Number of sick/recovered/dead (India) SIRVD (Susceptible, Infected, Recovered, Vaccinated and Dead) Model Parameters, on their basis—predictive values.	Artificial Neural Network based Adaptive Incremental Learning (ANNAIL).	From 30 January to 13 June, India.	Not separately identified	[15]
	Time series of sick/recovered/dead	Number of sick/recovered/dead	MLP	Information about infected, recovered and deceased patients in 406 locations for 51 days (from 22 January 2020 to 12 March 2020)	Coefficient of determination (R2) 0.94 for confirmed cases, 0.781 for recovered patients and 0.986 for models of deceased patients	[16]
	Population density, average temperature, maximum temperature, minimum temperature, precipitation, wind speed, humidity. their minimum, maximum, mean, standard deviation, skewness, kurtosis	Minimum, maximum, mean, standard deviation, skewness, kurtosis of the infection rate	Combination of Virus Optimization Algorithm (VOA) and Adaptive Network Fuzzy Inference System (ANFIS)	A total of 1657 records from various administrative divisions of the USA	R2 = 0.8338	[12]
	The number of COVID-19 cases and deaths around the world and in India by day from 1 January 2020 to 19 June 2020	Number of upcoming COVID-19 cases and forecast of deaths over the next 365 days	Support Vector Regression and Linear Regression	The number of cases and deaths around the world and in India by day from 01 January 2020 to 19 June 2020	R2 = 0.80	[7]
	Country, date, cancellation of public events (due to public awareness), severity index, testing policy (category of test facility available to the general population), total positive cases per million, new cases per million, total deaths per million, number new deaths per million, cardiovascular disease mortality rate, available hospital beds per thousand people, life expectancy, risk awareness (rate) about COVID-19, level of hazard and exposure, people using at least basic sanitation services (ratio), awareness about vulnerability (ratio), awareness about health status (ratio), awareness about vulnerability to epidemics (ratio), mortality rate, prevalence of malnutrition, lack of coping capacity, access to health care, physician number, current health care expenditure per capita, maternal mortality	Similarity clusters	K-means	Owid-covid-data, COVID-19-testing-policy, public-events-covid, covid-containment-and-health-index, inform-covid-indicators	Absence of reference clusters	[3]
	Daily data on new confirmed, deaths and recoveries of COVID-19 cases in Pakistan	Forecast of the number of sick, recovered and dead for 10 days in Pakistan	Models of vector autoregression of time series	Daily data on new confirmed, deaths and recoveries of COVID-19 cases in Pakistan from 8 March to 27 June 2020 has been downloaded from the World Health Organization	Not evaluated	[8]
	Total number of confirmed cases in India for the period 30 January 2020 to 29 April 2020 (showing the usefulness of 8 values for predicting the next day)	Forecast of the number of cases for the next day	A Logistic Growth Model to determine the stability of a pandemic and a Prophet Model to predict the total number of infections in India	Total number of confirmed cases in India for the period 30 January 2020 to 29 April 2020	Not evaluated	[9]
	Average age, PM 2.5 (air pollution level), population density, GDP per capita, average temperature and average humidity	R0 (infection index, more than 1—growth, less—decline)	Linear Regression, Linear Kernel Support Vector Machine (SVM), Radial Kernel SVM, Polynomial Kernel SVM, and Decision Tree	US state statistics	R2 = 0.473	[10]
	Image of the situation from the cameras	Type of violation of the regime (no mask, gloves, violation of social distance, contact, coughing, sneezing, spitting, hugging)	CNN	A total of 11,175 photos with/without described situations	F1 = 0.82	[13]
	Statistical data about COVID-19 cases in Indian states	Total cases/Discharges/Total deaths	Polynomial Regression, Decision Tree Regression, and Random Forest Regression	Number of Total cases, Discharges, Total deaths is various Indian states	RMSE value of 0.08	[11]
	Several factors including age, gender, province, epidemical data	Prediction of COVID-19 severeness	Deep learning models with autoencoder-based approach, SVM	Data Science for COVID-19 in South Korea dataset. Training includes 5165 coronavirus instances, validation—1533 quarantined patients	Average accuracy of 99%	[32]
	Daily number of confirmed, deceased, and recovered COVID-19 cases in 401 locations, over 78 days	Estimation of confirmed, deceased, and recovered cases and the estimation of epidemiology curve for specific countries	Genetic algorithms	Repository by the Johns Hopkins University Center for Systems Science and Engineering	R2 scores of 0.999	[19]
	Demographics, medication, past medical procedures, comorbidities, and laboratory results	Prediction of COVID-19 by multimodal fusion	Multimodal fusion AI model. XGBoost and multi-branched deep dense network	A total of 3194 COVID-19 patients	84% overall F1-score [CI 82.1–86.1]	[30]
	Risk factors were lymphocyte level, platelet count, and shortness of breath or dyspnea from medical records	Low and high severity of COVID-19	Deep neural network with 5 layers	A total of 5601 COVID-19 patients	balanced accuracy 90.3%, and AUC 0.96	[31]
	Statistical information from World Health Organization	Prediction of confirmed cases and deaths of COVID-19	ANN	Cases and deaths data are collected from 20 January to 11 November 2020 by the World Health Organization	Correlation coefficient R is 0.9948	[18]
	Temperature, humidity and wind as the independent variables	The relationship between COVID-19 spread and environmental factors	Adaptive Neuro-Fuzzy Inference System, CNN, Multiple Linear Regression	Data taken from another work [Pirouz, Behrouz, et al. “Investigating a serious challenge in the sustainable development process: analysis of confirmed cases of COVID-19 (new type of coronavirus) through a binary classification using artificial intelligence and regression analysis.” Sustainability 12.6 (2020): 2427.]	Average R-values of 0.90	[5]
	The epidemic data	Epidemic trend	Susceptible-Exposed-Infected-Removed (SEIR) model, Deep neural network, Recursive Neural Network (RNN)	COVID-19 data including the number of confirmed, cured and deaths from 23 January to 6 March 2020, combined with Baidu population migration data and relevant city data of the National Bureau of Statistics	Good graphical correspondence with real data	[20]
	Time-series data extracted from integrated Command and Control Center, Delhi, India	Forecasting of COVID-19 spread	ARIMA and Deep Convolutional Neural Network	Data from iCCC dashboard (https://covid.dashboard.jantasamvad.org/ accessed on 25 August 2022) was used	Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) are 57.093, 7.859	[21]
	Real time-series dataset of global record of confirmed, recovered, deaths and active cases of COVID-19 outbreak	COVID-19 future trends	Naïve Bayes, Support Vector Machine (SVM) and Linear Regression	Two datasets: the first contains a cumulative count of worldwide recovered, confirmed and death cases of COVID-19 from 22 January 2020 to 19 May 2020 and the second contains the global time-series records of COVID-19 from 22 January 2020 to 19 May 2020	MAE = 488,806.7492 and MSE = 400,919,367,451.7439	[22]
	John Hopkins’s COVID-19 dataset	Prediction of active cases of COVID-19	Recurrent Neural Network and Long short-term memory neural network	The dataset for 28 states and 8 union territories of India extracted from John Hopkins’s dataset	Good graphical correspondence between real data and modeled one	[17]
	Medical COVID-19 data from Virudhunagar district, India	COVID-19 affected regions classified into: danger, moderate, and safe zone	Deep neural network and decision trees	COVID-19 dataset of Virudhunagar district from the period of March to July 2020	Accuracy of 98.06%	[4]
1.2 Prescribing treatment by identifying similar patients and their dynamics/identifying analogies between countries as the pandemic progresses	Country, date, cancellation of public events (due to public awareness), severity index, testing policy, pandemic statistics	Similarity clusters	K-means	Owed-covid-data (number of infected), COVID-19-testing-policy (detection of COVID-19) public-events-covid (reaction to COVID-19 infections), covid-containment-and-health-index, inform-covid-indicators	Absence of reference clusters	[3]
	Adverse effects, route of administration, cost, plasma turnover, fever rate, age, pregnancy, and renal function	Evaluation of current treatment options, including favipiravir (FPV), lopinavir/ritonavir, hydroxychloroquine, interleukin-1 blocker, intravenous immunoglobulin (IVIG), and plasmapheresis	Fuzzy methods PROMETHEE and VIKOR	Not given	Ranked methods: Plasmapheresis was the most preferred alternative, followed by FPV and IVIG, and hydroxychloroquine was the least favorable	[43]
1.3 General review papers on the use of telemedicine and AI within it	Not applicable	General analysis for early detection and forecasting of the global dynamics of the pandemic in South America.	Not applicable	Not applicable	Not applicable	[23]
	Ketonuria, diet disorders and blood glucose values	Correction of diet/insulin adjustments	Algorithm not presented, ready-made SineDie solution	A total of 5108 patients, follow-up for diabetes	During follow-up, the system allowed for dietary adjustments in 20% of patients, 12 patients started insulin treatment, and 41.7% of them subsequently required therapy adjustments. 45.2% of insulin proposals submitted by the system were accepted, 29.0% shelved, and 25.8% rejected	[26]
	Not applicable	Follow-up of patients with Alzheimer’s disease during a pandemic	Not applicable	Not applicable	Not applicable	[24]
	Not applicable	Clinical, technical, financial and cultural barriers to telemedicine, and discusses expected benefits, including those using AI in telemedicine	Not applicable	Not applicable	Not applicable	[25]
1.4 Prioritization of patients for selected treatments based on their individual characteristics (such as blood types) and the severity of their condition	Age, gender, body mass index, medical staff, history of pregnancy, history of smoking, history of exposure, family history, number of cases among family members, people who were in Wuhan, ICU stay, interval between contact date and start date, Interval between start date and visit date, interval between start date and hospitalization date, interval between start date and date of antiviral therapy, blood parameters, etc. Based on the results of the analysis, the following criteria were selected as informative: gender (male vs. female), age > 70 years, temperature > 39 °C, cough, shortness of breath, hypertension, diabetes, secondary bacterial infection, lung injury, leukocyte count (10 × 109), neutrophils/lymphocytes (<3 vs. ≥3), alanine aminotransferase (≤40 vs. >40 U/L), aspartate aminotransferase (≤40 vs. >40 U/L), creatine kinase (≤185 vs. >185 U/L), lactate dehydrogenase (≤250 vs. >250 U/L), C-reactive protein (≤10 vs. >10 mg/L)	Prognosis of the likelihood of ARDS in patients with COVID-19	Logistic regression (LR), random forest (RF), support vector machine (SVM), decision tree (DT) and deep neural networks (DNN)	Clinical data of 659 COVID-19 patients from 11 regions in China	AUC = 0.99, Accuracy = 0.97	[27]
	Photo to determine the blood type. Biomarkers albumin, IgM/IgG, cytokines/chemokines, peroxiredoxin II, C-reactive protein, PaO2/FiO2 for prioritization using subjective and objective decision by opinion evaluation method and use in decision matrix.	Prioritization of plasma transfusion treatment based on donor and patient parameters.	Decision matrix, subjective and objective decision by opinion evaluation method. CNN for blood typing by image interaction with control groups.	Not applicable	Not performed	[28]
	Demographic, clinical and blood test data.	Prediction of which patients may have positive for SARS-CoV-2 test or require hospitalization or intensive care.	Logistic regression, neural networks, SVM, random forests and gradient boosting.	A total of 5644 patients	Patients positive for SARS-CoV-2 a priori with a sensitivity of 75% (95% CI 67%-81%) and a specificity of 49% (95% CI 46–51%), patients with SARS- Positive for CoV-2 requiring hospitalization with an area of 0.92 under the receiver operator characteristic curve (AUC; 95% CI 0.81–0.98), and SARS-CoV-2 positive patients requiring intensive care with an AUC of 0.98 (95% CI 0.95–1.00).	[29]
AI-based CDSS for research problem solving
2.1 Application of machine learning methods in DNA analysis for mutation detection, personalized diagnosis and treatment, treatment efficacy prediction, vaccine and drug development	DNA nucleotide sequence.	Prediction of the next nucleotide in a DNA sequence	Variations of Hidden Markov Models	A total of 1492 sequences	Up to 20%	[34]
	SARS-CoV-2 proteome, 100 most common HLA-A, HLA-B and HLA-DR alleles in the human population	Predictive schemes for the development of universal vaccines against SARS-CoV-2, which contain a sufficiently wide repertoire of T-cell epitopes that can provide coverage and protection to the entire population of the world	Monte Carlo method	About 22,000 people to create a “digital twin”, GISAID database at the stage of genome analysis	Not applicable	[39]
	Genome sequences	Search for genomic biomarkers that may predispose to more severe infection or death. Search for genomic biomarkers of response in patients treated during infection. Prognostic stratification of patients	Not conducted	Sample of 7000 patients, subgroup of 300 patients for sequencing	Not applicable	[38]
	Crystal structure of 2019-nCoV 3C-like protease, ‘Potency’, ‘IC50’, ‘Ki’, ‘EC50’, ‘Kd’ (assay confidence score ≥ 8).	Potential analogues of existing drugs with a different composition	GAN network.	Crystal structure of the 3C-like protease 2019-nCoV. The protease dataset was assembled with molecules active against various proteases in enzymatic assays extracted from the Integrity database, experimental pharmacology module, and ChEMBL. Total—60,293 unique structures	Not done	[36]
AI-based CDSS for diagnostic problem solving
3.1 Diagnostic models aimed at assessing the likelihood of any manifestation of the disease, its development and outcome.	Myo (myoglobin), CD8, age, LDH (lactate dehydrogenase), C-reactive protein, CD45, Th/Ts, dyspnoea, NLR, D-dimer, creatine, etc	Survival/death.	RF	A total of 126 patients, Wuhan.	AUC 0.9905	[49]
	Age, diabetes, coronary heart disease (CHD), lymphocyte percentage (LYM%), procalcitonin (PCT), serum urea, C-reactive protein and D-dimer (DD) (old age, CAD, LYM%, procalcitonin selected as independent)	Survival/death	Least Absolute Shrinkage and Selection Operator (LASSO), multivariate analysis	A total of 2529 patients/452 severe cases	AUC 0.919	[44]
	Neutrophil count, lymphocyte count, lactate dehydrogenase (LDH), highly sensitive C-reactive protein (CRP), and age	Survival/death	KNN	A total of 1766 points from 370 patients	Accuracy 90% 16 days before the outcome	[50]
	Levels of C-reactive protein (CRP), lactate dehydrogenase (LDH), leukocyte, lymphocyte, neutrophil and platelet counts, oxygen saturation level (SpO2), CT results	Survival/death	Statistical methods	A total of 866 patients	AUC 0.927	[41]
	Age group, gender, province, source of information about the disease (hospital admission, physician visit, etc.)	Survival/death	Logistic regression, support vector machine, nearest K neighbors, random forest and gradient boosting	A total of 3524 patients, Korea	Max accuracy 0.96	[42]
	C-reactive protein (CRP), N-terminal pro-B natriuretic peptide (NT-proBNP), myoglobin (MYO)), D-dimer, procalcitonin (PCT), myocardial creatine kinase band (CK-MB), and cardiac troponin I (cTnI)	Survival/death	LASSO	A total of 160 patients, Wuhan	AUC 0.94	[45]
	For age group 80–89 years old: use of ventilation, GFR less than 60 mL/min/1.72 m2 potassium. For age group: 90+: alanine aminotransferase, white blood cells count, pulse rate, procalcitonin, respiratory rate, C-reactive protein, history of congestive heart failure lymphocytes, sodium, hemoglobin. For age group: 70–79 years blood glucose, dementia neutrophils, anticoagulants	Survival/death	Ensembles, Gaussian process, linear models, naive Bayes, nearest neighbors, SVM, decision tree, discriminant analysis	A total of 1478 patients	AUC 0.84	[46]
	Prothrombin activity, urea, leukocytes, interleukin-2 receptor, indirect bilirubin, myoglobin and fibrinogen breakdown products	Pneumonia prognosis	Maximum Relevance and Minimum Redundancy (mRMR) algorithm and logistic regression model with least absolute compression and choice operator	A total of 110 patients	A percentage of 98% sensitivity and 91% specificity	[47]
	Evaluation of sequential organ failure, urea, respiratory rate, blood pressure, age, acute and chronic health state assessment, confusion, etc		Logistic regression	A total of 725 patients	AUC 0.89	[48]
	Gender, severity score at admission, temperature > 39 °C, cough, dyspnea, hemoptysis, hypertension, diabetes, secondary bacterial infection, lung consolidation, blood parameters	Survival/death	Decision Tree, Logistic Regression, Random Forest, Support Vector Machine, Deep Neural Networks	A total of 659 patients	Accuracy 0.97	[27]
	CRP (C-reactive protein), LDH (lactate dehydrogenase) and D-dimer	Survival/death	Deep Learning	Totally, 229 patients	Depends on factors, there is no integral assessment	[43]
	Patient characteristics (age, gender, immunosuppression and kidney disease), clinical parameters (pulse rate, blood pressure, respiratory rate, temperature, presence of shock and confusion), laboratory measurements (urea/blood urea nitrogen, leukocyte count, SpO2, hematocrit, glucose, sodium, and pH), radiological findings (pleural effusion and multilobar pneumonia on chest X-ray), and medical opinion (need for mechanical ventilation)	Survival/death	Models based mainly on regression statistical methods	Various datasets	Max AUC = 0.82	[40]
	X-ray data	Healthy/COVID-19/Pneumonia	COV-ELM (ELM modification)	COVID-19 Image Data Collection (760), COVID-19 Radiography Database (2905), Mendeley Chest X-ray Images (5856)	F1 = 0.94 ± 0.02	[76]
	X-ray data	Healthy/COVID-19/Pneumonia	CNN (the best—ResNet50) + SVM	COVID-19 Image Data Collection (760), COVID-19 Radiography Database (2905)	Max accuracy 98.66%	[57]
	Heart rate and steps interval data from wearable device	Healthy/COVID-19/non-COVID-19 illness	LSTM-based Autoencoder	A total of 25 COVID-positive, 11 non-COVID-19 illness, and 70 healthy patients	Average precision score of 0.91 (SD 0.13, 95% CI 0.854–0.967), a recall 0.36 (0.295, 0.232–0.487)	[51]
	Nanotechnology-based IOT biosensors data	Healthy/COVID-19	Fuzzy-based decision tree	Cord-19 dataset	Accuracy is up to 99%	[52]
	9381 United States Department of Defense personnel wearing Garmin and Oura devices, totaling 599,174 user-days of service and 201 million hours of data	Real-time infection prediction	Custom gradient boosting ensemble learning method	A total of 491 COVID-19 patients	AUC of 0.82	[54]
	Audio data (cough, breathing, and voice)	COVID-19 progression prediction and recovery trend prediction	Deep learning–enabled tracking tool using gated recurrent units	Crowdsourced respiratory audio data, including breathing, cough, and voice samples, from 212 individuals over 5–385 days	AUROC of 0.79, a sensitivity of 0.75, and a specificity of 0.71	[53]
3.2 Lung monitoring based on classical radiograph and CT image analysis for both direct diagnosis of COVID-19 and prediction of disease severity	CT scan	Segmentation (affected segments)	CNN (3D U-Net)	A total of 20 CT results	F1 0.956 for unaffected and 0.761 for affected areas	[58]
	X-ray data	Healthy/COVID-19/Pneumonia	CNN (VGG/MobileNet)	The COVID-19 radiography database consists of 219 COVID-19 image samples, 1345 viral pneumonia image samples, and 1341 normal chest X-rays	Max accuracy 98.75%	[59]
	X-ray data	Healthy/COVID-19/Pneumonia	CNN (VGG)	Cohen Dataset (230 samples) + Wang Dataset (100 + 100 samples)	Max accuracy 92.5%	[60]
	CT scan	Segmentation and assessment of disease severity based on it	CNN (M2UNet)	A total of 666 KT	Max F1 0.785, Max accuracy 0.985.	[61]
	X-ray data	Healthy/COVID-19/Pneumonia	CNN	CXRI + 382 cases with COVID-19	AUC = 0.84–0.88	[62]
	X-ray data	COVID-19/its absence	CNN (ResNet)	Zhao + Kaggle Dataset	Accuracy and specificity 95.09% and 81.89% sensitivity 100%	[55]
	X-ray data	Healthy/COVID-19/Pneumonia	CNN	Open-source COVID-19 + private data (pneumonia and normal cases)	Accuracy 0.96	[63]
	CT scans + 10 lab parameters + 23 observational features	Healthy/COVID-19– Light/COVID-19 –Complicated/Pneumonia	CNN + kNN/RF/SVM	Cumulatively, 689 CT examples (214 patients with non-severe COVID-19, 148 patients with severe COVID-19, 198 uninfected healthy participants, and 129 patients with non-COVID viral pneumonia.).	Accuracy 95.4–97.7%	[56]
	Fever, cough, diarrhea, vomiting, shortness of breath, have a chronic illness, work in health care, travel in the last 14 days. Blood test data CT (at different stages of diagnosis)	Prediction of COVID-19 infection	CNN + kNN/RF/SVM	A total of 689 CT scans (214 patients with non-severe COVID-19, 148 patients with severe COVID-19, 198 uninfected healthy participants, and 129 patients with non-COVID viral pneumonia.)	Accuracy 95.4–97.7%	[56]
	Fever, cough, diarrhea, vomiting, shortness of breath, presence of chronic illness, work in health care, travel in the last 14 days. Blood test data, CT (at different stages of diagnosis)	Prediction of COVID-19 infection	At the presented experimental stage—CNN	A total of 622 observations (122—COVID-19, 500—healthy) in terms of CT analysis	Accuracy 97.78% for CT	[64]
	CT scans	Formation of parameters, Diagnosis of COVID-19	Selection + selection of parameters (GLCM + HFSM) + KNN	A total of 498 + test sample	Accuracy 96%	[83]
	Lung CT images (manual analysis), age, sex, number of days before hospitalization; Impact on the area of the source of the epidemic; hypertension, cardiovascular disease, diabetes mellitus, current smoking, high fever, dry cough, expectoration, blood parameters	Prognosis of the severity of the disease	LASSO, MOTct, POIct and PSI levels	A total of 196 patients	AUC = 0.890	[79]
	COVID-CT dataset (https://github.com/UCSDAI4H/COVID-CT accessed on 24 August 2022)	COVID-19/non-COVID-19	Belief function-based convolutional neural network with semi-supervised training	There were 746 instances from 216 patients	Accuracy of 0.81, an F1 of 0.812 and an AUC of 0.875.	[75]
	X-ray data	Healthy/COVID-19	CNN (ResNet18, ResNet50, ResNet101, VGG16, and VGG19 for feature extraction), SVM	A total of 180 COVID-19 and 200 healthy patients	94.7% accuracy score	[71]
	X-ray data	Healthy/Pneumonia/COVID-19	shuffled residual CNN	558 COVID-19, 1437 Bacterial Pneumonia, 1216 Viral Pneumonia and 10,434 Healthy	F1-score of 97.20% and an accuracy of 99.80%	[65]
	X-ray, CT scans, and clinical indicators data	Healthy/Pneumonia/COVID-19	CNN-based classification framework	Large hybrid dataset available at https://drive.google.com/drive/folders/1-tXCaPtv0vupjXeLvrEX2cGt3tCOqyiv?usp=sharing accessed on 24 August 2022)	F1 scores > 96.72% (0.9307, 0.9890) and specificity > 99.33% (0.9792, 1.0000)	[66]
	X-ray data	Mild/ Moderate/ Severe/ Critical clinical picture.	CNN (ResNet152)	A total of 185 images with data augmentation procedure applied	AUC up to 0.94	[67]
	CT scans	Moderate/Severe COVID-19 groups	Combination of U-net and fully convolutional networks	A total of 465 scans	Spearman’s Correlation coefficient is more than 0.920 between radiologists and DL	[72]
	X-ray data	Healthy/COVID-19	CNN	A total of 253 images with augmentation procedure applied. Total 500 COVID-19 and 500 healthy	Overall accuracy of 99.8%	[68]
	X-ray data	Healthy/Pneumonia/COVID-19	CNN	A total of 2923 healthy, 371 COVID-19, 2778 bacterial, and 2840 viral pneumonia patients	99% accuracy for COVID-19 vs. health, >90% accuracy for other scenarios	[69]
	X-rays and CT scans	Healthy/Pneumonia/COVID-19	Feed-Forward and neural network and CNN, SVM	A total of 255 COVID-19, 255 healthy patients	Accuracy 84–100, AUC 0.85–1 depending on scenario	[73]
	CT scans	Healthy/COVID-19	The CNN was 5-layer deep	A total of 142 healthy, 142 COVID-19 patients	Accuracy of 93.64% ± 1.42%	[70]
	CT scans and clinical data	COVID-19 severity (high-risk and low-risk groups)	Deep convolutional neural network, random survival forest	A total of 1051 patients with RT-PCR confirmed COVID-19	C-index of 0.80	[74]
	CT scans and acoustic data	Symptoms of COVID-19	Deep Learning-assisted Multi-modal Data Analysis	A total of 276 and 502 samples belong to speech and breath audio samples; 1252 COVID-19 and 1229 healthy CT scans	Accuracy of 95.64%	[77]
	X-rays data along with radiology reports and RT-PCR data	COVID-19 severity	EfficientNet-B0 initialized on ImageNet pretrained weights	A total of 6500 non-public chest X-rays comprising diverse COVID-19 severities, along with radiology reports and RT-PCR data	95% diagnostic accuracy	[78]
3.3 Analysis of factors in the COVID-19 diagnosis and its complications	Counting blood cells, neutrophils, eosinophils, monocytes, platelets, lymphocytes, basophils, lactate dehydrogenase, aspartate aminotransferase, alkaline phosphatase, gamma-glutamyltranspeptidase, alanine aminotransferase, C-reactive protein, age	Selection of parameters, Diagnostics COVID-19	FCNB (Bayes)	A total of 207 patients	Max accuracy = 0.99	[87]
	Blood tests	Diagnostics of COVID-19	XGBoost + KNN+ iForest + SMOTE	A total of 5644 data samples with 559 confirmed cases of COVID-19	Accuracy 99.88%	[91]
	(1) Fever (Fe), (2) Headache (H), (3) Myalgia (M), (4) Fatigue (Fa), (5) Nasal congestion (NC), (6) Sneeze (S), (7) Sore throat (ST), (8) Difficult breathing (DB) and (9) Rhinorrhea (R).	Diagnostics of COVID-19	Fuzzy inference system.	A total of 272 patients.	Not performed	[89]
	Blood test (erythrocytes, hemoglobin, platelets, hematocrit, aspartate transaminase, lymphocytes, monocytes, sodium, urea, basophils, creatinine, serum glucose, alanine transaminase, leukocytes, potassium, eosinophils, C-reactive protein and neutrophils)	Diagnostics of COVID-19	NB + RF + SVM	A total of 600 (520 healthy + 80 COVID-19 confirmed) patients.	Accuracy 95%	[93]
	Age, gender, smoking, blood test	Diagnosis of pneumonia ventilator-associated pneumonia (VAP)	KNN, NB, DT, ANN, SVM, RF	A total of 59 patients	Accuracy 0.81 ± 0.04.	[80]
	Thirteen symptoms, estimated local prevalence, image analysis, and molecular diagnostics	Diagnostics of COVID-19	Bayesian inference network (BN) and set-cover models (SC)	A total of 55 patients, including those with fever (78%) or cough (77%), who applied for outpatient (n = 11) or inpatient treatment (n = 44). 51% (n = 28) were women, 49% were <60 years of age. Common comorbidities included diabetes (22%), hypertension (27%), cancer (16%), and cardiovascular disease (13%)	Sensitivity (81.6–84.2%) and specificity (58.8–70.6%)	[81]
	Symptoms found as the results of the examination, data on contacts	Diagnostics of COVID-19	COvid Risk cALculator (CORAL) CDSS, using a modified Delphi method	A total of 2443 patients	0.2% of false negative results	[84]
	Epidemiological history; Wedge-shaped/fan-shaped lesion, bilateral lower lobes; Opacity, chaotic paving pattern	Diagnostics of COVID-19	TRIPOD type 3 or 2b	Undefined	Not shown	[85]
	Diagnostic manuals in text form + specialists able to interpret it	Manuals interpreted by a computer to carry out diagnostics	Data analysis based on the developed rules	Not shown	Not shown	[90]
	Blood tests	Diagnostics of COVID-19	GNB, SVM, ANN	A total of 1186 patients	AUC 0.913, sensitivity 0.801 and specificity 0.890	[82]
	40 features, 12 selected: Lung involvement, cough, fever, dyspnea, oxygen deficiency in the blood, features of digestion, history of ARDS, history of contacts, disability, history of pulmonary infection, respiratory rate, rhinorrhea	Diagnostics of COVID-19	Binary logistic regression (BLR) method and Forward Wald method	A total of 800 patients	Accuracy = 90.25% AUC = 0.835	[86]
	A history of fever or presence of fever. Symptoms and signs of respiratory distress syndrome (cough, cold, sore throat, fatigue). Severe pneumonia or acute respiratory viral infections (ARVI) No other causes based on convincing clinical descriptions. A history of travel or residence abroad that reported local infection transmission	Diagnostics of COVID-19	Naive Bayes Classifier	Not shown	Not shown	[88]
	Patient data, ventilator wave parameters. Patient blood saturation (SpO2), inspiratory volume (TVi), Expiratory tidal volume (TVe), Positive end expiratory pressure (PEEP)	Breath (cycle) recording classification: Normal, Artifact (does not contain Patient-Ventilator Asynchrony (PVA)), Double Trigger Asynchrony (DTA), or Breath Accumulation Asynchrony (BSA) Detection of acute respiratory distress syndrome (ARDS, ARDS)	Ensemble of extremely random trees classifier (ERTC), gradient boosted classifier (GBC), and multilayer perceptron (MLP)	A total of 9715 breath records from 35 ventilated patients	For the first classification accuracy is more than 0.97 for all three classes For the second classification AUC 0.88	[92]

), illustrating the applicability of individual methods within the designated classes of problems, the input parameters used in the solution, and the achieved accuracy indicators of the obtained models. The list of indicators below contains the exact metrics used in the review, and not a complete list of them:

-

Accuracy—the proportion of correctly defined examples. It is easy to understand and interpret. However, with unbalanced data sets it is not indicative and cannot be used as a quality criterion, because the less represented class, within which the greatest number of errors are made when targeting with a given metric, as a rule, is more important;

-

AUC (area under curve) is the area under the curve of the proportion of correct positive predictions versus the false positive predictions proportion;

-

Sensitivity and Specificity—the proportion of correctly predicted positive labels (true positives) and correctly predicted negative labels (true negative);

-

Coefficient of determination (R2)—proportion of the variance of the dependent variable, explained by the model under consideration;

-

Precision—shows what proportion of objects recognized as objects of a positive class is predicted correctly;

-

Recall—shows what proportion of objects that actually belong to the positive class is predicted correctly;

-

F1 is a metric that combines precision and recall with equal priority.

4. Conclusions

The current state of science makes it possible to adapt existing AI methods for solving new problems quickly, in particular those that have arisen during the COVID-19 pandemic. There are three areas of AI application for making medical decisions of the most diverse focus on work, from an individual patient to obtaining predictions on a global scale.

The most applicable mathematical algorithms for solving the organizational problems during a pandemic include: simple recurrent neural network, long-term short-term memory, gated repeating units, variational autoencoder, support vector regression, linear regression, K-means, logistic regression, neural networks, etc.

Variations of hidden Markov models, Monte Carlo method, GAN network are the best analytical methods for scientific problems.

Clinical decision-supporting systems based on logistic regression, support vector machine, nearest K neighbors, random forest, gradient boosting, ensembles, Gaussian process, linear models, naive Bayes, nearest neighbors, decision tree, discriminant analysis, etc. are used for diagnostic purposes.

Despite the overall positive trend, a number of shortcomings can be identified that can be considered the promising areas in the framework of the creation of AI-compiled CDSS in the context of counteracting COVID-19 and in a more general sense:

The vast majority of the found solutions are not aimed at processing the patient’s current indicators in real time. The papers consider the analysis of the patient’s input indicators, which, as a rule, are static in the context of a pandemic—age, gender, the presence of chronic diseases. On the basis of the static information, a single prognosis/diagnosis is given, on which basis a decision is made on hospitalization/severity of the condition; however, the results of continuous monitoring were not found in the works.

The analyzed works do not present CDSS aimed at managing patient support systems. The work only makes it possible to monitor the operation of the ventilator system, however, data on the management of this system depending on the patient’s condition, adaptation to them are not presented. This shows the prospects of research in the development of such systems that would reduce the burden on physicians, allowing them to monitor the patient’s condition and take part in the validation of decisions made by the system and to launchA appropriate treatment.