An Intelligent System for Medical Decision Support in Differential Diagnosis and Treatment of COVID-19

Abstract

The development of electronic services to help doctors in the diagnosis and treatment of COVID-19 is discussed. The existing systems for such purposes are analyzed, and requirements for them are formulated. The architecture of an intelligent medical decision support system and the basic principles of its development using an ontology-oriented approach are given. The unique capabilities of the system are shown, and its information and software components are described.

The problem of early differential diagnosis and prescription of personalized treatment for COVID-19 has not lost its relevance even two years after the onset of the pandemic; moreover, it is becoming increasingly important [1]. Compared to many other groups of diseases that have been studied for a long time and have a large evidence base, therapies for COVID-19 are in the process of formation [2]. New strains are emerging, knowledge about which is still insufficient, and treatment regimens are often changed due to their imperfection (from March 2020 to December 2021, 13 versions of clinical recommendations were issued in Russia).

New strains of SARS-Cov-2 are found in the population unexpectedly and unpredictably; their detailed study takes time; and it is in short supply, because the emergence of each new virus variant is associated with the need to take urgent diagnostic and treatment measures. It should be noted that the differential diagnosis of viral diseases is quite complicated and accompanied by many medical errors (according to statistics, up to 30%), since many of these pathologies have similar symptoms. In the case of COVID-19, the situation is aggravated by the fact that, due to the scale of the epidemic, doctors of various profiles are forced to be involved in the process of diagnosis and treatment, including those who do not have special knowledge in viral infections and who make decisions based on incomplete information. To help in the diagnosis and treatment of this class of diseases and to reduce significantly the percentage of doctors’ errors [3, 4], as noted in many literature sources, intelligent systems for supporting medical decision making exist. This article is devoted to description of the basic principles, architecture, information, and software components of the system aimed at the differential diagnosis and treatment of COVID-19.

LIMITATIONS ON TECHNOLOGIES FOR THE DEVELOPMENT OF MEDICAL DECISION SUPPORT SYSTEMS

The emergence of COVID-19 has led developers to create new and update existing software products in this area. As a result, the medical community has gained access to many services, including the Russian coronavirus calculator [5], klinica.com [6], mayoclinic [7], health.mail.ru [8], the electronic assistant MeDiCase [9], WML.Symptom Checker [10], Infermedica [11], Isabel [12], Botkin.AI [13], the cough sound detection neural network for COVID-19 [14], WebMD [15], and others. We would like to note particularly the service of differential diagnosis and treatment on the Med-IACPaaS platform, which was the first to appear in March 2020 [16].

Among the important criteria for evaluating or classifying software tools aimed at early diagnosis of COVID-19 and its treatment, we highlight the following:

• target audience (patients or doctors), as well as supported stages of the medical process;

• level and detail of the patient description;

• number of pathologies analyzed;

• explainability of the proposed hypotheses;

• sources of knowledge and opportunities for their accumulation;

• integration with an electronic health record.

Let us take a closer look at each of these six criteria.

The target audience: patients and doctors. Most of the available services are designed for patients; their main task is to use a set of fairly simple signs to help determine the likelihood of a certain condition and generate useful recommendations. No more than 30% of available Internet services are useful not only for patients but also for doctors. Only a few are intended exclusively for physicians; however, support for decision-making of the entire range of tasks facing the doctor (diagnosis, treatment methods, risk assessment, etc.) is not fully implemented in any of them.

Level and detail of patient description: only complaints, considering objective studies and taking into account the dynamics of symptoms. Patient services use a simplified set of incoming information, usually static, almost always based on patient complaints, and do not offer an objective investigation. Often, a preprogrammed interview is offered, starting with the first main symptom (complaint). Entering the results of any measurements and laboratory tests is rare. The ability to add signs of the disease in the process of interacting with the service is also a rare option. The results of instrumental studies are used in neural network algorithms, but they allow one to analyze only images (sometimes the sound of a cough). Single services provide a means of describing the history of a developing disease.

Number of pathologies analyzed. Some services support the analysis of one pathology [5, 14, 17], and most of them, from tens to hundreds [11, 12, 15].

Explainability of the proposed hypotheses: absence, enumeration of typical symptoms, and detailed explanation. Where a decision is supported not for preventive purposes but for the purpose of choosing treatment tactics, especially personalized (including surgical intervention), the doctor needs an explanation of the expected outcomes. Services implemented on the basis of machine learning methods are characterized by the absence of justification for decisions (WML.SymptomChecker, WebMD, Isabel). For services using knowledge, an explanation can be generated. For example, the Infermedica service provides an explanation for each diagnosis from the list generated as concise lexical constructions, for example, this one: “Influenza. Moderate evidence is the presence of symptoms: rapid heartbeat, headache, dry cough.” A useful detailed explanation can be provided mainly by medical decision support systems based on ontological knowledge.

Sources of knowledge and opportunities for their accumulation: popular science literature, professional community, and available data from practice. Most of the available services offer one of the most common manifestations of the disease, which is more like information from popular science literature than expert assessments (health.mail.ru, WebMD); some report that they used the knowledge of professional doctors (Infermedica, MeDiCase). In addition, algorithms for statistical analysis and calculation of event probability, and mathematical models for representing patients for diagnostics are added (klinica.com, Botkin.AI). Most of the available services are rarely updated, only a new disease is added to the database and not variants of its manifestations. However, some services promise continuous development of the resource due to the mechanisms for accumulating new information and its further use (klinica.com).

Integration with an electronic health record (EHR). Most of the services are implemented as separate programs, and the user is required to fill out an online form: enter the results of analyses, metrics, etc. Some developers claim that they can integrate these results into medical information systems, and others, that they have a text analysis component from EHR (Webiomed—WML.SymptomChecker, klinica.com). One of the problems of integrating services with an electronic health record is that each of them has its own limited set of names (and meanings) of symptoms and other signs, often with ambiguous meaning. A unified medical classifier or terminological base is not used in this case.

As was mentioned, at present, research teams in Russia and abroad are actively conducting research on the creation of decision support systems for practicing physicians. Most of this work is related to the analysis of images (usually computer tomograms). Image analysis is an important component of the diagnostic process, but for it to be accurate and, as a result, accompanied by the correct treatment, it is necessary to take into account clinical data such as complaints, laboratory tests, life history, etc. The available decision support systems based on the analysis of clinical data use simplified knowledge models (without considering fuzzy information, the development of the disease over time), and also do not conduct differential diagnostics (by form, severity), prescribe personalized treatment in accordance with clinical recommendations, and do not provide a detailed explanation of the proposed interventions. An innovative solution based on an ontological approach to support early diagnosis, differential diagnosis of diseases, and treatment prescription is described below.

BASIC PRINCIPLES OF THE CREATION AND ARCHITECTURE OF AN INTELLIGENT SYSTEM

The treatment and diagnostic process consists of several main stages [18]: preliminary diagnosis; treatment based on the preliminary diagnosis, symptoms, and severity of the patient’s condition; differential diagnosis, taking into account the form and stage of the disease; and correction of treatment based on an accurate diagnosis and its monitoring.

At the stage of preliminary diagnosis, a search for diagnostic hypotheses (possible diseases) is carried out: collection and analysis of the patient’s history, analysis and clarification of complaints, and objective examination. At this stage, an intelligent medical decision support system (IMDSS) should help the doctor make a preliminary diagnosis by analyzing all possible options, as well as issue recommendations for laboratory and instrumental studies, considering the preliminary diagnosis.

Before obtaining the results of laboratory and instrumental studies, the task of the IMDSS is to prescribe the initial treatment based on the patient’s complaints and the results of his/her objective examination, taking into account concomitant diseases, history of allergies, age, etc. Then an accurate diagnosis is carried out, if necessary, with additional studies and correction of the initial appointments. The IMDSS can assist the clinician in making a diagnosis, choosing the right drugs and their combinations, choosing the dosing regimens, determining the duration of taking each drug, and developing a plan for monitoring the patient’s condition. These tasks are often solved with errors.

Considering the described functionality that IMDSS should support and the characteristics of the disease, we single out the following requirements for the system [19]:

• rapid deployment (knowledge of diagnosis and treatment is constantly updated);

• maintenance of all the stages of the treatment and diagnostic process mentioned above based on the data of the electronic health record;

• compliance of diagnostics and prescribed treatment with clinical recommendations of the Russian Ministry of Health;

• generation of a detailed explanation that is understandable to doctors, that is, using generally accepted terminology;

• implementation of the system as a cloud service (any update is available immediately to all users);

• medical specialists should modify and expand the body of knowledge.

To implement these requirements and ensure functionality, we propose an architecture that is schematically illustrated in Fig. 1. The system is implemented on the basis of ontological knowledge bases [20], formed by experts with the participation of editors. Among the main ones are “Knowledge Ontology of Disease Diagnostics,” “Knowledge Ontology of Disease Treatment,” and “Pharmacological Guide Ontology.” To describe knowledge, the “Database of Medical Terminology and Observations” and “Pharmacological Guide Database” are used. An ontology-driven smart solver generates a detailed explanation based on the data of the electronic health record, also based on the ontology.

Fig. 1.

Architecture of an intelligent medical decision support system.

Components of the differential diagnosis of COVID-19. To solve diagnostic problems on the medical portal of the IACPaaS cloud platform [21], the already mentioned “Knowledge Ontology of Disease Diagnostics,” “Knowledge Base of Viral Diseases,” as well as software components that provide data processing were placed as information resources: “Intelligent Problem Solver of Diagnostic Problems and Differential Diagnostics” and “Service for Entering an Electronic Medical Record.”

“Knowledge Ontology of Disease Diagnostics” corresponds to modern diagnostic standards and the level of medical knowledge and describes the clinical picture in the temporal dynamics of the pathological process, as well as the impact of therapeutic measures and other events on its manifestations [22]. Each disease is represented by alternative symptom complexes (they include the complex of patient complaints and an objective examination and a complex of laboratory and instrumental studies), the necessary conditions for the onset of pathology, causes, risk factors, and a detailed diagnosis. The latter includes a description of the elements of the clinical classification according to the form, variant, severity, stage, etc., and is a set of signs (or a symptom complex) that helps clarify the main diagnosis, the likelihood of complications, or functional disorders. The ontology used gives medical experts the opportunity to form knowledge bases of diseases with all variants of symptom values in all periods of the course of the disease. A fragment of the ontology is shown in Fig. 2.

Fig. 2.

Screenshot (translated) of “Knowledge Ontology of Disease Diagnostics.”

The “Knowledge Base of Viral Diseases” contains a description of COVID-19, influenza, parainfluenza, numerous infections (acute respiratory, adenovirus, rhinovirus, and others). It is formed in accordance with the ontology that determines the causal relationships of diseases used in medical diagnostics with the dynamics of external manifestations [23] and in accordance with clinical recommendations and methodological guidelines of the Russian Ministry of Health [5, 14, 24]. The description of each disease includes a code according to the International Classification of Diseases of the 10th revision (ICD-10), causes, risk factors, a necessary condition or event that led to the onset of the disease, a number of symptom complexes, and a detailed diagnosis. The need to group symptom complexes is dictated by the clinical features of the course of the disease in different categories of patients: children, adults, the elderly and senile, pregnant women, and athletes, as well as the frequency of atypical forms of the disease. Detailed diagnosis of COVID-19 is described according to clinical classifications: severity, degree of lung damage (based on CT), degree of respiratory failure (by saturation, SpO₂; by gasometric data; by clinical data), course options, complications, etc.

THE PRINCIPLE OF IMDSS OPERATION

The user enters general information about the patient, a list of complaints, and objective examination data recorded in the electronic health record, and starts the process of primary diagnosis. The system searches, generates several hypotheses of the preliminary diagnosis, and offers them to the user. To confirm any of them, additional data must be entered.

The user enters preliminary diagnosis formed by the system into the EHR. It explains the result by generating necessary, characteristic, and possible features. You can then proceed to treatment recommendations or continue with the second step of diagnosis to form a final diagnosis. After entering the data of laboratory and instrumental studies proposed by the system, the process of differential diagnosis takes place, excluding diseases with similar symptoms but inappropriate for any signs, ultimately leading the user to the only likely hypothesis. After the introduction of the main diagnosis in the EHR, the system generates a complete detailed clinical diagnosis with details.

The final result of the system is a structured report. In its format, agreed with specialists, the results are grouped according to their importance. The report provides an indication of the analyzed EHR, confirmed hypotheses for it (one or more in comorbidity), and hypotheses refuted and considered in further analysis.

COVID-19 treatment prescription component. When developing the treatment prescription component, the goal was to comply as closely as possible with the methodological recommendations for the prevention and treatment of a new coronavirus infection, considering personal clinical manifestations and the severity of the disease. To achieve it, taking into account the requirements for the system described above, information components were developed: “Pharmacological Guide Ontology” with the ability to describe the compatibility of drugs, “Knowledge Ontology of Disease Treatment,” which allows describing various models and regimens of drug therapy, taking into account the personal data of the patient, and features of the clinical picture with the possibility of monitoring drug therapy and the patient’s condition. On the basis of these ontologies, relevant information resources have been created to form modern ideas about the treatment of COVID-19.

Pharmacological guide ontology. Currently, there are many available pharmacological guides in digital and paper format, which feature more than 4000 drug names. Their structure is generally identical and consists of a set of fields, each of which is presented in text format. Among them are key fields for prescribing treatment such as pharmacological properties, pharmacological action, contraindications, method of administration and doses, interaction with other drugs, etc. The developed ontology of the pharmacological guide is intended to describe in a formal, rather than textual form, all the necessary structural units of the drug [23]. The ontology made it possible to generate a knowledge base that includes 8000 concepts. All of them are structured according to the classical concepts of clinical medicine about a drug and make it possible, first of all, on the basis of clinical recommendations, to form bases on the treatment of diseases—not only COVID-19 but also other socially significant pathologies—gastroenterological, cardiological, and some viral ones.

Knowledge ontology of disease treatment includes a description of a specific pathology, containing the ICD code, model, type, goal, treatment regimen, and recommendations (Fig. 3). The therapy regimen contains a condition for its prescription and a group of medicines used alternatively and in a complex. The term type of therapy combines a whole class of concepts: etiotropic, pathogenetic, symptomatic, empirical, and other types of therapy. The term target of therapy encompasses hemostatic, antiemetic, antipruritic, detoxification, or mucolytic therapy. The term therapy regimen defines the list of active substances, their combinations, the mode of administration, and the dosage of drugs for the optimal treatment of the disease (Fig. 4).

Fig. 3.

A fragment of the ontology of the knowledge base on the treatment of diseases, hosted on the IACPaaS cloud platform (screenshot translated).

Fig. 4.

A fragment of the ontology of the knowledge base on the treatment of diseases “Therapy regimen,” hosted on the IACPaaS cloud platform (screenshot translated).

One of the necessary stages in the formation of the volume and sequence of therapy is the fulfillment of certain conditions. A condition can consist of one block of criteria or several connected by logical links. Each block of criteria can be subjected to an additional constraint—a selection rule. The clinical criterion refers to all kinds of observations of the patient, the results of tests and examinations, and ICD codes. The criterion can be either simple or consist of a certain set of characteristics, on which a separate rule can be imposed.

Based on the ontology, a knowledge base on the treatment of COVID-19 was formed. Interim guidelines for the prevention, diagnosis, and treatment of a new coronavirus infection were used as their source (Fig. 5). Knowledge about the treatment of this disease is presented in strict accordance with the Interim Guidelines of the Russian Ministry of Health (version 13) and is strictly differentiated by type of therapy: “Outpatient treatment” and “Inpatient treatment.” Differentiation is made in the process of analyzing clinical data entered in the EHR.

Fig. 5.

A fragment of the knowledge base on the treatment of the new coronavirus infection, hosted on the IACPaaS cloud platform (screenshot translated).

During the analysis of the patient’s clinical data, the treatment prescription complex generates recommendations for drug therapy in accordance with the criteria for its goals and regimens, providing a list of possible drugs and rules for their use (Fig. 6). The description of the method of application of the recommended medicinal product fully complies with generally accepted standards and includes a decoding of a single dosage, frequency of administration, form of release, and duration of treatment.

Fig. 6.

Generation of recommendations corresponding to outpatient treatment (screenshot translated).

* * *

The intelligent system for supporting medical decisions in the differential diagnosis and treatment of COVID-19, which we offer, has the following main features that fundamentally distinguish it from analogues:

• the system carries out differential diagnosis of the disease and prescribes treatment in accordance with clinical recommendations (based on ontological knowledge bases);

• the model of medical knowledge has a semantic representation, and the presence of a specialized editor provides subject matter experts with the opportunity to introduce knowledge and correct it;

• replenishment of the knowledge base on the diagnosis of new strains of the disease, as well as methods of treatment, does not require programming;

• the system issues recommendations for additional examination if the data entered is not enough to make decisions;

• the model of medical knowledge (the ontology on which the knowledge base is formed) corresponds to modern concepts in medicine; contains tools for describing the fuzziness (modality) of signs, their development over time, and the formation of alternative symptom complexes for various diagnostic methods; prescribes treatment methods in accordance with clinical recommendations; and takes into account the personal characteristics of the patient and the severity of the disease;

• all solutions proposed by the system have detailed explanations;

• the system is a cloud service.

The shell of the proposed IMDSS was used in the implementation of the system for diagnosing and treating COVID-19 using traditional Chinese therapy in the PRC for prompt use by doctors at the height of the epidemic (February–March 2020) [25]. The system was implemented in Chinese (together with Chinese specialists) at the request of the Association of Nongovernmental Medical Institutions of China. Experience has shown that the development meets all the formulated requirements and can be successfully applied in medical practice. Currently, together with specialists, the issue of conducting clinical trials of it in Russian medical institutions is being discussed.

FUNDING

This work was supported in part by the Russian Foundation for Basic Research, project nos. 20-07-00670 А and 19-29-01077.

CONFLICT OF INTEREST

The authors declare that they have no conflicts of interest.