Utilizing artificial intelligence as an arbitrary tool in managing difficult COVID-19 cases in critical care medicine

Abstract

This opinion review paper explores the application of artificial intelligence (AI) as a decisive tool in managing complex coronavirus disease 2019 (COVID-19) cases within critical care medicine. Available data have shown that very severe cases required intensive care, most of which required endotracheal intubation and mechanical ventilation to avoid a lethal outcome if possible. The unprecedented challenges posed by the COVID-19 pandemic necessitate innovative approaches to patient care. AI offers significant potential in enhancing diagnostic accuracy, predicting patient outcomes, and optimizing treatment strategies. By analyzing vast amounts of clinical data, AI can support healthcare professionals in making informed decisions, thus improving patient outcomes. We also focus on current technologies, their implementation in critical care settings, and their impact on patient management during the COVID-19 crisis. Future directions for AI integration in critical care are also discussed.

Key Words: Artificial Intelligence; COVID-19; High-resolution computed tomography; Computed tomography; Diagnosis; Treatment; Critical care medicine arbitrary

Core Tip: Coronavirus disease 2019 (COVID-19) presents a range of characteristic patterns and findings on computed tomography (CT) scans that reflect disease progression and severity. Accurate interpretation is crucial for patient management, yet this task is complicated by the variability in radiologists' experience and training. Standardizing CT reporting by grouping findings into distinct categories based on disease stage could improve consistency. However, variability and potential subjectivity persist, highlighting the need for artificial intelligence (AI) support in imaging diagnostics. AI can aid radiologists in achieving more accurate, objective interpretations by identifying, classifying, and quantifying changes, ultimately contributing to a more reliable and standardized approach to diagnosing and managing COVID-19 in critical care settings.

INTRODUCTION

The coronavirus disease 2019 (COVID–19) pandemic, caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2), challenged healthcare worldwide. Firstly, identified in Wuhan, China, the virus rapidly affected the whole world within months, affecting the lives of every citizen and causing closed schools, universities, businesses, etc. As of April 13, 2024, 704753890 cases were confirmed and 7010681 deaths[1]. In addition, according to the World Health Organization, 13.64 billion COVID-19 vaccine doses were administered[2]. The disease had different fatality ratios in the different age groups, from 0.004 in the age 0–34 up to 28.3% in the age 85+ years[3].

The pandemic started in Bulgaria on March 8, 2020, with the first patients being a 27-year-old man and a 75-year-old woman. Initially, there was a lack of PCR tests. In some cities, samples were taken and sent to other towns for evaluation, leading to errors. It was necessary to immediately diagnose the suspected COVID-19 patient, and the preferred method of choice was high-resolution computed tomography (HRCT) due to its high sensitivity, specificity, and immediate results. Based on the results of the HRCT examination, it was decided whether to hospitalize the patient or not[4,5]. The role of HRCT changed once countries had sufficient PCR tests. The method is no longer used for initial diagnosis, according to the Canadian Society of Thoracic Radiology, The Royal College of Radiologists, the American College of Radiology, etc.[6,7]. CT is now used in patients with negative reverse transcription-PCR (RT-PCR) tests[8] and in patients whose symptoms are worsening[9]. CT plays a vital role in distinguishing COVID–19 pneumonia from other types of respiratory diseases because X-ray has lower sensitivity and specificity. COVID-19, caused by SARS-CoV-2, was initially thought to be a primary respiratory illness that sometimes led to viral pneumonia. After collecting and analyzing data during the pandemic, it has been concluded that it is a complex multisystem disease requiring intensive care treatment[10,11]. HRCT has an invariable role in assessing the severity of changes, which makes the method vital in determining the therapy of patients. The need for quantitative and qualitative determination requires the use of modern post-processing methods, such as artificial intelligence (AI). There is a need to link the CT findings with laboratory tests. Standardization in the approach to the diagnosis and follow-up of patients is required. Nowadays, when the pandemic has been under potential control, COVID-19 and other respiratory diseases still require rigorous diagnostic approaches and precise arbitrary, especially in critical care settings[9].

This opinion review aims to examine the potential of AI as a supplementary tool in managing complex COVID-19 cases in critical care settings. This review highlights the value of AI-driven systems in providing enhanced accuracy and objectivity in diagnostic imaging, especially in the context of CT scans, which are critical for assessing the extent and severity of lung inflammation. By exploring the capabilities of AI in standardizing CT report interpretations and supporting clinical decision-making, this review discusses how AI can help overcome the challenges of diagnostic subjectivity and variability, ultimately contributing to improved patient outcomes and streamlined care in critical COVID-19 cases.

АRTIFICIAL INTELLIGENCEIN RADIOLOGY

Radiologists play a crucial role in interpreting medical images to diagnose COVID-19. The last two years showed that AI technologies demonstrated performance that matches the accuracy of all radiologists in several specific tasks. However, a step in this area is the development of AI models that can compete with healthcare personnel in particular tasks. The last data reveal the success of using AI to detect breast cancer in screening mammography[12-15]. This success was achieved after 10 years of hard work and the creation of the OPTIMAM image database (OMI-DB), a database with a total group of > 150000 customers[16]. The database contains over 2.5 million images from over 170000 women collected from three United Kingdom breast screening centers. The image collection is ongoing, and the database is constantly updated. OMI-DB has been shared with over 30 research groups and companies since 2014[17].

Similarly, from 2019 till now, significant progress has been made in diagnosing lung conditions using chest X-ray (CXR) and CT. Many datasets have been published to help radiologists determine different pulmonary conditions from CXR[18,19] and malignancy lung conditions with CT[20,21]. AI may, therefore, support the differentiation of pneumonia caused by COVID-19 and other forms of pneumonia (e.g., other respiratory and infectious diseases)when used in high-prevalence and symptomatic populations[22,23].

The Joint Research Centre published a review comparing the added value of AI vs humans to detect COVID-19 by medical imaging[24]. Concerning diagnostic performance, in testing datasets, reported sensitivity was 42%–100% (human readers), 60%–95% (AI systems), and 81%–98% (AI-supported readers). Reported specificity was 26%–100% (human readers), 61%–96% (AI systems) and 78%–99% (AI-supported readers). The potential for AI is very good for assessing changes in the severity of lung lesions, while several studies have shown potential time saving for COVID-19 detection with AI. All these 1270 publications indicate that COVID-19 has accelerated the development of AI in medical imaging[24].

In areas affected by the pandemic, negative RT-PCR tests but positive CXT and CT are significant signs of COVID-19 and can emphasize the importance of rapid detection of SARS-CoV-2 infection. This gives the community and clinicians a much better chance of reducing the spread of the virus[25,26]. CT of the lungs of pneumonia caused by COVID-19 depicts bilateral, subpleural, groundless opacities with air bronchograms, undefined edges, and a slight predominance in the right lower lobe[27]. The image model facilitates the differentiation of different lung infections regarding their origin and structure. In the presence of large sets of images and specialized reading of the image reports, these data can be used for diagnostics and made available as structured machine-learning labels[28-30]. AI provides consistent and detailed image analysis to reduce diagnostic errors. With AI, the anomalies that are often invisible to the human eye are detected, thereby significantly enhancing the accuracy of diagnosis. In addition, AI is crucial in reducing human errors, e.g., effectively counteracting human fatigue and overseeing and providing reliable interpretations without interference from external factors[31-33]. AI contributes efficiency and speed, dramatically accelerating the process of interpreting medical images. This acceleration is sometimes critical because it is associated with life-saving decisions[34,35].

AI algorithms quantify areas of lung involvement, highlight abnormal patterns and suggest possible classification categories (e.g., early, progressive, severe) in structure report format. The next step is the clinical validation of the AI-generated findings by radiologists specializing in image interpretation. They produce an objective and standardized report for clinical review and help the clinical team make timely treatment strategy decisions[36,37]. The most important elements of COVID-19 difficult cases are presented in Figure 1.

Figure 1 Artificial intelligence-enhanced accuracy in computer tomography for coronavirus disease 2019 cases. AI: Artificial intelligence; CT: Computer tomography; COVID-19: Coronavirus disease 2019.

STANDARDIZATION OF THE REPORT OF THE CT CHANGES

The COVID-19 pneumonia causes a broad variety of changes, such as ground–glass opacities–usually in the periphery of the lungs, bilaterally, mainly in the basis of the lungs, crazy paving appearance, air space consolidation, bronchovascular thickening, traction bronchiectasis[27,38-40]. Peripheral distribution of the findings, ground glass opacities and bronchovascular thickening in the lesion are reported to have the highest discriminatory value(P < 0.001)[41]. The following findings are unusual in COVID–19 patients and suspect added bacterial pneumonia – lymphadenopathy, tree-in-bud, effusions, pulmonary nodules, cavitations. Pneumothorax and pneumomediastinum are seen as complications of intubated patients.

The Radiological Society of North America (RSNA), the Society of ThoracicRadiology, and the American College of Radiology classified the CT changes into 4 groups, which terms should be used for standardized reports[42].

1 group

In the 1st group, the changes are typical for COVID–19 pneumonia. The changes in this group include ground glass opacities, which are seen in the periphery of both lungs and thickened intralobular septs. Other typical changes are multiple ground glass opacities, which are rounded, with or without consolidation of the intralobular septs. "Reverse halo sign" can also be noted.

2 group

The second group represents a lack of typical changes. The findings in this group are diffuse, multiple ground glass opacities, which lack the specific distribution and can be other than round and not visualized in the periphery.

3 group

In this group, the changes are atypical, representing isolated consolidation, affecting a lobe or segment without ground glass opacity. Other findings are discrete small centrilobular nodules, tree-in-bud, cavitations or smooth interlobular septal thickening with pleural effusion.

4 group

No CT features of pneumonia, absence of consolidation, and ground glass opacities exist. There is a study performed by Simpson et al[42] from the Radiological Society of North America, which compares the CT classification and RT-PCR results, showing 77%-97% typical findings (Group 1), 51%–64% indeterminate findings (Group 2), 3%-5% atypical scans (Group 3) and 20%–25% negative scans (Group 4), confirmed with RT-PCR[42].

CT CLASSIFICATION ACCORDING TO THE STAGE OF THE DISEASE

Assessing the severity of COVID-19 using chest CT is crucial for determining the right therapy. Determining the severity of COVID-19 requires a highly skilled radiologist to provide an assessment through visual analysis. This is time-consuming and sometimes subjective. This is why the application of AI is very suitable for determining the severity of COVID-19 from mild, moderate to severe[43].

The assessment of the changes seen on CT should not only be qualitative but alsosuggest at which stage they are. In recent publications, 4 stages were described: Early stage, which is between 0 and 4th day, showing normal CT or ground glass opacities only; almost 50% of the patients have normal CT scans in the first 2 days. Later, in the progressive stage, days 5–8, there is increased crazy paving and ground glass opacities. In the peak stage, which is between the 9^th and 13^th day, there are parenchymal consolidations. After that, in the final stage of absorption, which is after 2 weeks, we find fibrous changes, which should resolve in one month or more[39,44,45].

THE ROLE OF A IN THE ASSESSMENT OF THE COVID – 19 PATIENTS

The current gold standard diagnostic method is RT-PCR[46]. Alternatively, CXR and CT are commonly used imaging methods. Rapid interpretation of CT scans allows for more accurate and rapid initiation of treatment.

There is a new initiative of a multicenter European project that can implement AI in CT, which will create a deep learning model for automated detection and classification of COVID–19 patients, which can help assess the severity of the changes. The European Society of Medical Imaging Informatics supports the project, and it aims to train an AI algorithm with studies from different countries across Europe. The aim is to automate the diagnosis of COVID-19 on CT and quantify the changes. The model is freely accessible to all participating hospitals; the Netherlands Cancer Institute coordinates the process[47].

But why do we need to use AI to interpret the CT examinations? In the previous peaks of the disease, we saw rapidly growing cases of COVID–19 patients, which overloaded the healthcare systems all over the world. Radiologists play a crucial role in diagnosing these patients, making their jobs harder in emergencies. Typically, the interpretation of CT examination of COVID–19 patients takes up to 10–15 minutes. AI interprets the changes within seconds, which allows the process to be optimized (Table 1).

Table 1 The multiple roles artificial intelligence plays in different computer tomography analysis stages highlight current challenges and provide a clear path forward in utilizing artificial intelligence effectively for critical coronavirus disease 2019 case management.

Aspect	Description	Role of AI	Challenges and limitations	Future recommendations
CT imaging for COVID-19	Provides qualitative and quantitative data on lung inflammation and disease severity	AI assists in enhancing image quality by quantifying areas affected	Variability in scanner technology, inter-radiologist variability	Standardize protocols and AI algorithms for uniform results
Standardized reporting	Ensures that findings are categorized systematically to aid clinician decisions	AI generates standardized templates, suggesting classification	Radiologist training impacts consistency in subjective classifications	Implement AI-guided reports with customizable templates
Stage-based classification	Identifies progression: Early, progressive, or severe stages of disease	AI classifies stages based on set parameters from CT findings	Stage overlap can affect classification accuracy	Use AI to refine and expand classification criteria
Objective diagnosis support	AI reduces subjectivity by assisting radiologists with unbiased imaging analysis	AI improves diagnostic accuracy and reliability in assessments	It requires extensive training data and potential over-reliance on AI	Encourage AI-human collaboration with continuous model updates
Clinical guidance	Facilitates treatment planning by providing objective data on disease severity and progression	AI-generated reports provide objective insights for treatment	Integration with EMR systems and real-time decision-making limitations	Integrate AI findings into EMR for seamless treatment planning

AI: Artificial intelligence; CT: Computer tomography; COVID-19: Coronavirus disease 2019; EMR: Electronic medical record.

RECOMMENDATIONS AND FUTURE PERSPECTIVES

As AI demonstrates increasing potential in managing critical COVID-19 cases, especially in high-stakes essential care settings, several recommendations can optimize its integration and address associated challenges[48,49]. First, developing standardized protocols and guidelines for using AI tools in clinical decision-making is crucial. These standards should ensure the accuracy, safety, and transparency of AI algorithms, which can support intensive care unit staff with dynamic decision-making, patient monitoring, and predictive analytics for early detection of deterioration. Regulatory bodies should work with healthcare institutions to establish these frameworks, addressing ethical, privacy, and security concerns associated with patient data use.

An essential step is the refinement and expansion of AI models to account for diverse clinical scenarios and patient populations. COVID-19 has revealed variations in disease presentation and progression across demographics, and AI models should be trained on datasets encompassing this diversity to improve predictive accuracy and minimize biases. Collaboration between hospitals, research institutions, and AI developers will foster datasets that include underrepresented groups, thus improving the generalizability and fairness of AI applications[50].

AI in COVID-19 management could also benefit from improved interpretability. AI systems currently in use, such as machine learning models for predicting oxygen needs or ventilator requirements, often function as "black boxes" leaving clinicians with limited understanding of how decisions are made.

Explainable AI (XAI) techniques can enhance trust in these systems by providing clinicians with insights into the decision-making processes of AI tools, enabling more informed and collaborative patient care. According to XAI, the most important markers are C-reactive protein, D-dimer, basophils, lymphocytes, neutrophils and albumin. The models can be implemented in different healthcare tools to predict the severity of COVID-19[51]. So, the XAI makes the models more understandable and could be extensively utilized in the healthcare sector[52].

Finally, a promising future direction lies in integrating AI with other digital health technologies like wearable devices and telemedicine platforms. Wearable health devices have recently been developed to monitor patients' health in real-time. It enables lower medical costs, quick access to patient health data, the ability to transmit data even in harsh environments, and non-invasive deployment[53]. This is a new area in personalized monitoring and treatment, but it willgreatly help in rapid diagnosis. By combining real-time patient data from wearables with AI's analytical power, clinicians could receive continuous updates on vital signs and symptoms, which may facilitate timely interventions outside the hospital setting. Research and development efforts should focus on these hybrid approaches, which may transform COVID-19 management in outpatient and home care settings, along with other socially demanding diseases, such as bladder cancer and colonic polyps management[54,55].

In conclusion, utilizing AI in critical care for COVID-19 presents transformative potential but requires coordinated standardization, interpretability, and inclusivity efforts. With continued research, interdisciplinary collaboration, and ethical oversight, AI could become an invaluable asset in the future of critical care medicine, extending beyond COVID-19 to address broader challenges in patient management and healthcare resource allocation.

CONCLUSION

In conclusion, CT is invaluable in assessing patients with COVID-19, providing both qualitative and quantitative insights into the severity and progression of inflammatory changes within the lungs. The standardized terms introduced for radiologists have improved the uniformity of reporting; however, interpretation often hinges on the capabilities of the specific CT scanner and the radiologist's experience. This can introduce variability and subjectivity into CT report generation, potentially affecting diagnostic accuracy and treatment decisions, such as predicting patient outcomes, optimizing ventilator management, adjusting medication regimens based on real-time data, and supporting early intervention in critical care settings.

Integrating AI into CT scan interpretation could greatly enhance objectivity and consistency in these reports. By applying advanced image analysis algorithms, AI can help standardize interpretations across different machines and experience levels, identifying patterns that may be missed by human observers and providing quantitative metrics that can assist in tracking disease progression or response to treatment. This standardization would be instrumental in reducing diagnostic variability and guiding clinicians more reliably toward optimal, individualized patient management.

Moreover, as AI tools evolve, they may offer predictive capabilities, such as identifying patients at risk for rapid deterioration or forecasting the likely progression of lung inflammation. Such predictions would empower clinicians to make more proactive treatment choices, potentially improving patient outcomes. Thus, adopting AI as an assistive tool in CT interpretation represents a promising step toward enhanced, objective, and data-driven clinical care in managing severe COVID-19 cases and other complex respiratory diseases in the future.