Researchers worldwide have had to face a myriad of challenges at a near-relentless pace, driven by the complex nature of the disease caused by the Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2). Although mass vaccination campaigns have significantly alleviated the effects of the pandemic, it is becoming increasingly clear that Coronavirus Disease 2019 (COVID-19) is likely to accompany us for many years to come. The vaccines have proven to be particularly effective in reducing the risk of severe and devastating forms of the disease caused by the original strain of the coronavirus. However, the emergence of new and more infectious variants of the coronavirus casts further doubts on our ability to completely eradicate the disease (1). Since the initial variant, known as Alpha (B.1.1.7), several other variants have been identified, including the Delta (B.1.617.2) and the Omicron (B.1.1.529) variants. The Delta variant has been estimated to spread about 225% faster than the Alpha variant (2). Similarly, the Omicron variant has shown increased transmissibility compared to its predecessors. As recently reported by WHO (World Health Organization), the Omicron XBB.1.16 variant has shown an increase in the prevalence, and it may become dominant in some countries and cause a rise in case incidence due to its growth advantage and immune escape characteristics. However, despite its high transmissibility, early data suggested these Omicron variants are associated with milder symptoms and a lower risk of hospitalization compared to other variants. Given that, and also the important fact that vaccines provided a significant level of protection against severe disease and other complications, the proposed study focused only on Alpha and Delta variants.

Unfortunately, up to 50% of people of all age groups end up having Post-Acute Sequelae of COVID-19 (PASC). These symptoms may linger for months or even years and seem to depend on a variety of factors including vaccination status, the virus variant and geographic region (3). It is strongly suspected that these phenomena are closely linked to the innate and adaptive immune response mechanisms of individuals (4–6).

The immune system activates different pathways of innate and adaptive immune response mechanisms to effectively combat viral infections and their associated disease progression. In previous epidemics such as SARS and MERS (Middle East Respiratory Syndrome), the immune response mechanisms of the host, including both innate and adaptive components, have been closely associated with the progression of the disease and clinical outcomes in affected patients (7). Also, in Zan et al. (8) Killer-cell Lectin-like Receptor D1 (KLRD1) has been identified as a potential biomarker for flu susceptibility. During the current ongoing pandemic, additional factors such as advanced age and male gender have been identified as risk factors for the progression of COVID-19 toward severe and often fatal disease (9).

The outbreak of COVID-19 has undoubtedly prompted a large number of scientists to employ Artificial Intelligence (AI) to help combat the pandemic crisis. In fact, Machine Learning and Deep Learning, particular approaches of AI, are a very useful tool to understand and fight this disease.

In literature there are several thousands of publications that show a various range of ML applications that have been developed to address different issues related to the virus and to the associated disease, including forecasting the number of future infections and the trajectory of the outbreak (10, 11); diagnosing the disease, e.g., using imaging tools in addition to the standard Polymerase Chain Reaction (PCR) test (12); predicting the mortality and severity risk (13); improving drug development and vaccination (14); enhancing contact tracing methods from individual interviews to more efficient digital contact tracing (15). What becomes apparent is that, despite the remarkable progress in the field of ML applied to COVID-19, the current approaches suffer from some important drawbacks (16–21). First and foremost, the lack of data is a crucial element in the development of models. In fact, it is well-known that patients’ data are difficult to gather, and considering that ML models rely on large datasets, this problem plays an important role on models’ generalization and performance. Secondly, many laboratory parameters included in the studies are collected after the patient has contracted the virus and the disease has already evolved, thereby excluding the option of early and tailored treatment in high-risk individuals (22, 23). Another drawback is that the proposed algorithms have not yet reached the level of a human expert. This particular flaw is closely related to the need for ML models with high interpretability and transparent output. These qualities are crucial so that human readers, e.g., medical professionals, can have trust in the conclusion made by the model. Decision Tree-based models are particularly helpful in this case as they can provide comprehensible rules behind every decision taken.

What emerges from the published studies is that ML is a popular paradigm to create new solutions that can help in the fight against the virus and the related disease. However, as far as the authors’ knowledge, no studies that can identify “a priori” subjects at risk of developing the severe clinical manifestation of COVID-19, based on immunogenetics, demographic and clinical characteristics, were published.

The aim of the proposed study was to develop an algorithm capable of predicting the risk of developing a more aggressive and severe disease course in individuals before they become infected with SARS-CoV-2. An AI method, combining common demographic risk factors (age and gender) (24, 25) with key immunogenetic risk markers, including HLA class I molecule alleles (HLA-A, -B, -C) and the 14-base pair polymorphism in the HLA-G gene, is described. In fact, classical and non-classical HLA antigens play a crucial role in activating and regulating innate and adaptive immune responses (26). Additionally, KIR genes regulating NK lymphocyte activity, which have a significant impact on the severity of SARS CoV-2 infection (27), have been included.

The proposed predictive model involves the use of Decision Trees (DT) (28, 29), which belong to the supervised machine learning classification methods. DTs are able to provide high classification accuracies together with a simple and intuitive representation of gathered knowledge. This peculiarity made them quite popular in different areas of medical decision making (30). DTs are compared with Naïve Bayes (NB) classifiers, as they are a common benchmark in classification problems.

Considering the above, the main contributing parts of this work are described as follows:

The remaining sections of the paper are structured as follows: section 2 provides a description of the Materials and Methods. This section is further divided into subsections. Section 2.1 discusses the ethical statement, section 2.2 describes the database containing the parameters of the COVID-19 infected subjects, sections 2.3 and 2.4 present a theoretical explanation of the NB and DT models, respectively, and section 2.5 introduces the indices used to assess the model’s performance. The experimental results are presented in section 3. In section 4, an organized discussion and possible limitations of the method are presented. Finally, the paper ends with a conclusion in section 5.

Figure 2 shows the odds ratio for COVID 19 protective and risk factors and their related confidence intervals (39). The plots highlight the factors more closely associated with protection or an increased risk of a severe clinical disease course. The age, which maximized the separation between asy/pauci and severe groups, was 55.

The immunogenetic factors that seemed to offer increased protection against the more severe clinical forms of COVID 19 were: age less than or equal to 55 years, female gender, presence of the KIR2DS2 gene, particularly when combined with HLA-C alleles of the C1 group, presence of KIR2DS3 and presence of the HLA-G 14-bp Del/Del or Ins/Ins genotype. It is interesting to note that the presence of the protective haplotype (HLA-A*02:05, B*58:01, C*07:01, DRB1*03:01), not shown in Figure 2, yielded an OR of 0, owing to the fact that this haplotype was completely absent in the severe group of patients. This factor has previously been shown to have a strong protective effect and, despite its relatively high frequency (2.6%) in the Sardinian population, was only observed in 4 of the 80 (3%) individuals pertaining to the asy/pauci group.

The strongest risk factors for a severe clinical disease course were: age above 55 years, male gender, absence of the KIR2DS2 gene alone or in combination with HLA-C alleles of the C1 group (KIR2DS2/HLA-C C1+), presence of the 14-bp polymorphism (Ins/Del) of the HLA-G gene, presence of the KIR2DS5 gene and presence of the KIR telomeric region A/A (Tel A/A).

The associations found for COVID 19 protection/risk factors using univariate analysis deserve further investigation in larger cohorts of patients.

Table 2A shows the average performance of the 10 NB models for training, leave-one-out, and test data sets. The range of the different indices among the 10 models is also provided for the whole database. The NB models, trained on the full set of features, showed relatively high accuracy on the training data with a score of 86.02%; despite this, the average accuracy for leave-one-out and test were 61.87% and 60.33%, respectively. This shows that NB models fit the training data, but perform worse on test data, which means that they experience a high overfitting. Similar results would be obtained by means of DT without any action to limit overfitting.

To avoid overfitting, when training the DT models, the tree-building process was stopped before it produced leaves with very small samples. This heuristic is known as early stopping, although occasionally it is referred to as pre-pruning. The minimum sample size in terminal nodes was fixed at 6, since this helps to maximize the average accuracy on the 10 models for the leave-one-out set (see Figure 3). An example of a pre-pruned tree (one of the 10 trained models) is shown in Figure 1.

Table 2B shows the average performance of the 10 DT models for training, leave-one-out and test data sets. The range of the different indices among the 10 models is also provided for the whole database. With respect to the results obtained on NB models, the DT models showed similar accuracy on the training data with a score of 83.41%, while higher average accuracy, equal to 80.98%, for leave-one-out. Accuracy on test data decreased on average to 74.67%showing also for decision trees a little overfitting. When considering performance on the whole dataset, the averaged accuracy remained high (81.70%). The accuracy in the worst model was still high (79.74%). However, larger deviations were observed for TPR (72.60–84.93) and TNR (75.00–88.75). In fact, models with low TPR generally report high TNR and vice versa, making accuracy values quite stable.

The reduced overfitting obtained on DT models with respect to NB models could be due, as well as to the pre-pruning stage, also to the capability of DTs to perform a feature selection. Figure 4 shows the relative importance of the most significant features (in order of importance) used by the DT models. In order to easily compare the importance obtained in the different tests, the impurity gain for each feature was normalized by the sum of the impurity gains of all features. The red circle indicates the average normalized importance, while the blue bars indicate the interval obtained among the 10 models.

As can be seen, the most important features were age group and gender. This finding is furthermore supported by the results of univariate analysis shown in Figure 2.

The genetic feature with the highest importance was the 14-bp insertion/deletion polymorphism of the HLA-G gene, already identified as a risk or protection factor in univariate analysis. For the sake of comparison, the results on the NB models by using only the 11 features selected by DTs shown in Figure 4 are reported in Table 2C. As it can be noticed, the performances on the leave-one-out after the feature selection are better than those obtained with the full features dataset and closer to the results obtained by DT models.

In line with current literature, age group and gender were leading predictors for all the considered models. The age group threshold identified by the DTs was 55. Table 3 reports the number of asy/pauci and severe classified cases according to the different age and gender groups. In the database, the asy/pauci and severe cases were equinumerous (19 asy/pauci and 18 severe) in the group of women over 55 years of age and, therefore, did not provide any information for this group. On the other hand, 40 out of the 49 men older than 55 developed severe symptoms. Among younger patients (age less than 56 years), only 2 of the 35 women developed severe symptoms. Thus, young age is in itself a protection factor for women. Among younger men, only 13 out of 32 developed severe symptoms.

After extrapolating the rules leading to classification scores greater than 2/3 on the training set, present in at least three models, the retained rules involved four genetic parameters: presence or absence of the HLA-G 14-bp Del/Ins genotype, presence or absence of KIR2DS5, presence or absence of 2DS2/HLA-C1 and presence of the KIR telomeric region (Tel A/A). Among older women, addition of the parameter KIR2DS5 allowed for correct classification in 27 out of 37 cases: 14 out of 19 asy/pauci cases and 13 out of 18 severe cases. Among older men, after adding presence or absence of the KIR2DS2/HLA-C C1+ functional unit and the KIR telomeric region (Tel A/A), it became possible to correctly classify 42 out of 49 cases: 5 among the 9 severe cases and 37 among the 40 asy/pauci cases. In the group of younger men, presence or absence of the HLA-G 14-bp Del/Ins genotype allowed for the correct classification of 25 out of 32 cases: 9 among the 13 severe cases and 16 among the 19 asy/pauci cases.

Figure 5 illustrates the diagnosis classification scheme. The red and green rectangles represent severe and asy/pauci classification, respectively. The two numbers in these rectangles indicate the number of asy/pauci and severe patients falling within each class. Overall, 126 out of 153 cases were correctly classified: 63 of the 80 asy/pauci cases and 63 of the 73 severe cases. The classification failed for 10 older women (5 asy/pauci and 5 severe), 7 older men (4 asy/pauci and 3 severe), 2 severe young women and 7 young men (3 asy/pauci and 4 severe). As shown, for older women, older men and younger men, the additional genetic parameters improved the classification based on age group and gender by about 21.62%, 4.08%, and 20.00%, respectively.

Performance of the two models—one based on age group and gender and the other expanded to include genetic parameters—is shown in Table 4A. Although the model based on age group and gender has a quite good performance in recognizing actual asy/pauci patients, it exhibits a very low ability in recognizing actual severe patients. The inclusion of genetic parameters helps to increase the accuracy on the overall dataset by more than 10 % points, thereby consistently increasing the TPR from 54.79% to 80.82%. This property is confirmed by the PPV and NPV. In fact, while the classification based on age group and gender has a low NPV, demonstrating a low ability in identifying asy/pauci cases, the proposed expanded model yielded both PPV and NPV above 80%.

The test on the aging of the prediction model in Figure 5 by means of the 42 patients infected by the Delta variant, lead to 24 (57.14%) patients with severe disease assigned to the severe class, and 18 (42.86%) patients, who were either asymptomatic or pauci-symptomatic, assigned to the asy/pauci class. Out of the 18 female patients, most of whom were under the age of 55, 77% did not develop severe disease. Out of the remaining 24 male patients, 58.3% developed the severe form of the disease.

Table 4B compares the performance of the proposed predictive model to that of the model exclusively based on age group and gender. Although the model based on age group and gender achieved quite good performance in identifying asy/pauci patients, it had a very low ability in recognizing severely ill patients. After adding the genetic parameters, accuracy increased by more than 7 % points with a remarkable increase in TPR from 44.44% to 77.78%. Moreover, TNR% yielded a value greater than 79%. It should be noted that the proposed model also yielded quite good values for PPV and NPV. The latter reached a value of about 83% thereby providing us with a high degree of confidence that, in case of negative predicted value, the tested patient will not develop the severe form of the disease even if he/she has been infected by the Delta variant virus. It is worth noting that PPV and NPV are closely related to the prevalence, and thus should be carefully evaluated during screening stages. The comparison with the numbers in Table 4A confirms the conclusions made when referring to the datasets of patients infected by the Alpha variant, thereby showing a limited aging effect; all the performance indices decreased but still remained well above the 70% threshold.

Medicine is one of the main sectors in which the use of AI has been successfully applied, not only achieving improvement in diagnostic strategies but also enabling more specific and targeted therapies that in future will allow for increasingly “personalized” medicine. The application of AI embraces a variety of medical fields that range from the digitalization of diagnostic imaging and the replacement of paper files with digital reporting to the development of biotechnologies associated to “omics” sciences, all of which have led to a true explosion of the so-called Internet of Medical Things (IoMT). IoMT is focused on healthcare and is an extended and more specific version of the Internet of Things (IoT) (40–42). Combined to point-of-care diagnosis, it could be used within the current crisis to develop a digital healthcare platform capable of administering proper care to COVID-19 patients at home as well as to provide government and healthcare organizations with a comprehensive disease management database (43).

On the topic of prediction of mortality risk and severity, several works have been published. The aim of the first line of research was to assess the mortality risk, i.e., to determine patients with the highest risk of dying from COVID 19 using clinical, demographic and radiological data on admission to the hospital from the electronic medical records (44–48). Each study selected various biomarkers that can predict the mortality, e.g., age, male gender, white blood cell, comorbidity profile of the patient such as diabetes mellitus, respiratory distress, coronary artery disease, and others.

The second line of research aforementioned, focused on the assessment of the severity risk, i.e., the development of a ML model based on clinical, radiological and laboratory characteristic to predict, on the basis of the patients outcome, the severity of the disease (49–53). Male gender, obesity, smoking, cerebrovascular disease, chronic liver disease, etc., were clinical determinants of Covid-19 severity. The study in Toraih et al. (54) explored pre-existing cardiovascular diseases as a risk factor of a severe form of COVID 19. This study suggested that, beyond the well-known pulmonary complications, high severity and mortality rates are closely related to viral myocarditis, myocardial damage, microangiopathy, and acute coronary syndromes. The conclusion of this study was that elevated cardiac injury biomarkers may improve the identification of patients with high risk of mortality and severity. In de Freitas et al. (55)’, the main symptoms, signs and demographic data that were most often associated to the hospitalization in patients with respiratory problems, have been defined by means of a ML model. In Sun et al. (56), researchers employed a Support Vector Machine (SVM) model where input data are associated with the clinical course of COVID-19. The results showed that a four-feature combination of GSH, CD3 ratio, total protein and age resulted in an Area Under Receiving Operating Curve (AUROC) of 0.9996 and 0.9757 in the training and testing datasets, respectively. Kaplan–Meier survival and cox-multivariate regression analyses confirmed the ability of the model to individuate severe cases using the selected clinical parameters. Moreover, after evaluating 253 clinical blood samples from the city of Wuhan in China, a group of researchers found 11 key relevant indices (total bilirubin, creatine kinase isoenzyme, glucose, creatinine, kalium, lactate dehydrogenase, platelet distribution width, calcium, basophil, total protein, and magnesium) which can detect, with an overall specificity and sensitivity of 95.95 and 96.97%, respectively, COVID 19 infected patients (57).

In general, most of these algorithms couple demographic parameters such as the sex and age of the patients to simple clinical and laboratory parameters such as blood group, transaminase, bilirubin, creatinine, number of platelets, etc. Some other studies have included more sophisticated laboratory markers: the dosage of cytokines, chemokines, the structure of lymphocyte subpopulations and various genetic factors including mutations in coagulation genes (58, 59). In Yasar et al. (60), three COVID-19 positive patient groups (mild, severe, and critical) and a control group have been classified based on the blood protein profiling.

Among all ML methods, Logistic Regression, SVM, Random Forest (RF) and Gradient Boosting Decision Trees were the most used. Particularly, the best possible results may be achieved by the gradient boosting models, RF and NNs.

Additionally, machine learning models tuned by metaheuristic algorithms like Genetic Algorithms and Particle Swarm Optimization have gained popularity due to their ability to find near-optimal solutions in complex problem domains.

In the context of COVID-19, metaheuristic frameworks that combine the Generalized Boltzmann distribution and orthogonal Jacobi polynomials have been successfully employed to analyze the spread of the virus (61). These metaheuristic approaches have also found applications in other critical areas, such as the monitoring and management of COVID-19 patients. By leveraging the Internet of Things (IoT), real-time monitoring of COVID-19 patients could be achieved through wearable devices, enabling automatic alerts to mitigate risk factors (62).

Throughout the pandemic, millions of people with or without symptoms have been isolated at home, waiting for the evolution of the infection either toward recovery or worsening of the clinical symptoms. This has often led to delays in hospitalization and the administration of timely and effective treatment (63). Indeed, early treatment of patients at a greater risk of developing the more severe forms of the disease is fundamental to the healing process and reduces the incidence of patients admitted to intensive care and the overall mortality rate (64).

AI algorithms have been widely used to predict the clinical evolution of SARS-CoV-2 infection. These algorithms have a high accuracy and ability to discriminate severe cases. However, an intrinsic limitation is represented by the fact that the laboratory and instrumental parameters on which these tools are based are altered by the interaction of the organism with the virus. This limits their predictive ability in the early stages of infection, when it is more important to be able to predict the evolution of the disease toward severe forms in order to ensure timely and appropriate treatment strategies.

The antiviral Paxlovid (Pfizer Inc.) for COVID-19 appear promising for treating COVID-19 in the early stages and seem to be particularly indicated for adults with mild to moderate disease and at risk of developing a more severe disease course (65). Therefore, it becomes even more important to identify individuals who are at a high risk of experiencing severe illness and hospitalization.

Unlike previous algorithms described in the literature, the proposed algorithm is based on demographic and epidemiologic data as well as the specific immunogenetic characteristics of each individual which are totally unrelated to the humoral and tissue changes generated during the viral infection.

This approach, which has a TPR and TNR of around 80%, allows us to identify “a priori” subjects at high risk of developing the severe clinical manifestations of COVID-19 should they become infected with SARS-CoV-2. Despite the limited number of cases used for AI training, the extremely high predictive ability of the model supports the reliability and effectiveness of the proposed algorithm.

The high ability to discriminate subjects at high risk of developing the severe and potentially lethal forms of COVID-19, both before and in the early stages of SARS-CoV-2 infection, makes this particular algorithm an interesting system for use in clinical practice. Another problem that needs to be addressed is that even fully vaccinated individuals do not always mount an appropriate immune response when challenged with the virus and in some cases may require booster shots to ensure adequate protection. The proposed algorithm might be capable of individuating these cases and could perhaps help researchers upgrade therapies as well as the trained immunity currently offered by COVID-19 vaccines.

The main limitation of this study lies in the relatively small number of subjects examined, as well as the focus on only two variants [B.1.1.7 (Alpha) and B.1.617.1 (Delta)]. Therefore, the high predictive capabilities of the proposed AI algorithm need to be validated on larger and more diverse cohorts. Furthermore, the use of these approaches is currently hindered by the challenge of regularly collecting immunogenetic variables in non-hospitalized patients, thus limiting its application beyond hospital settings. However, it is important to acknowledge the growing trend of integrating genomic information into routine clinical practice. Technological advancements and decreasing costs of genomic sequencing are making complete genomic profiling of individuals at birth increasingly feasible and potentially commonplace. As genomic medicine progresses, the availability and integration of immunogenetic variables in clinical settings are expected to improve, creating new opportunities for the broader application of AI algorithms in healthcare.

However, the algorithm can be significantly simplified by reducing the number of immunogenetic markers to only four genetic parameters strongly associated with a more severe disease course: absence of the KIR2DS2 gene alone or in combination with HLA-C alleles of the C1 group (KIR2DS2/HLA-C C1+), presence of the 14-bp polymorphism (Ins/Del) of the HLA-G gene, absence of the KIR2DS5 gene and absence of the KIR telomeric region A/A (Tel A/A).

These four genetic markers made it possible to identify with certainty about 80% (TPR) of all patients who developed the severe form of COVID-19 whereas the use of sex and age alone only identified half of those at risk of developing severe SARS-CoV-2 infection.

It is interesting to note that the proposed algorithm resulted to be effective in discriminating against severe forms of COVID19 infection independent of whether the patients had been infected by the Alpha variant and the more common and widespread Delta variant. It can therefore be assumed that its application would also be effective in the case of newly emerging Omicron variants.

In synthesis, all subjects who test positive by a nasopharyngeal swab can quickly be classified in a low or high-risk range. Those at high risk can receive early treatment with the anti-COVID19 drugs such as the oral antiviral drug Molnupiravir (66) and/or passive immunotherapy based on the use of anti-spike monoclonal antibodies which has shown to yield the best results when administered in the initial stages of infection (63, 67).

The proposed algorithm could be made user-friendly by creating a rapid antigen test to analyze the four most relevant genetic markers either at home or in the clinic. The algorithm can be inserted into an application for iOS or Android to facilitate its use even in an out-of-hospital setting. This would allow for early treatment of patients at high risk of developing severe disease with the newly emerging but highly expensive drugs.

However, considering that it is extremely difficult to develop high quality clinical prediction models that are beneficial to patients and providers in the different areas of healthcare (22), the proposed algorithm will need to undergo scrupulous validation on large cohorts before it can be recommended for widespread use in routine clinical practice.

Overall, the algorithm tested in this pilot study is extremely reliable and combines clinical and demographic features with some simple immunogenetic markers associated with the more severe forms of COVID-19. With appropriate modifications, the proposed approach could represent a starting point for the creation of other algorithms applicable to a wide spectrum of viral infections.

In conclusion, the application of artificial intelligence (AI) in the field of medicine, particularly in the context of the COVID-19 pandemic, has shown great potential for improving diagnostic strategies and enabling personalized medicine. The use of AI algorithms has facilitated the prediction of clinical evolution and identification of individuals at high risk of developing severe forms of COVID-19. While existing AI algorithms primarily rely on laboratory and instrumental parameters that are influenced by the interaction between the virus and the organism, the proposed algorithm takes into account demographic, epidemiologic, and specific immunogenetic characteristics. This unique approach has demonstrated a high accuracy in identifying individuals at high risk of severe clinical manifestations of COVID-19, even in the early stages of infection. Furthermore, the algorithm’s ability to discriminate between severe and mild/moderate cases of COVID-19 is independent of the viral variants, including the Alpha and Delta variants, suggesting its potential effectiveness against variants and viruses with the same characteristics. To simplify the implementation of the algorithm in clinical practice, it is proposed to reduce the number of immunogenetic markers to four genetic parameters strongly associated with a more severe disease course. This would make it feasible to determine the risk level of patients using a rapid antigen test in a user-friendly manner, either at home or in a clinic setting. However, before widespread implementation, the algorithm requires rigorous validation on large cohorts to ensure its reliability and effectiveness. Additionally, the integration of the algorithm into a digital healthcare platform, along with the administration of early treatment using antiviral drugs and passive immunotherapy, could significantly improve disease management and reduce the incidence of severe cases and overall mortality rate. In conclusion, the present study demonstrates the potential of AI algorithms, combined with demographic, epidemiologic, and immunogenetic data, in identifying individuals at high risk of severe COVID-19 and facilitating early treatment. This approach could serve as a foundation for the development of similar algorithms for various viral infections, highlighting the broader applications of AI in healthcare.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://www.ncbi.nlm.nih.gov/bioproject/PRJNA904643.

The studies involving human participants were reviewed and approved by Ethics Committee of the Cagliari University Hospital; date of approval: 27 May 2020; protocol number GT/2020/10894. The patients/participants provided their written informed consent to participate in this study.

All authors listed have made a substantial, direct, and intellectual contribution to the work and approved it for publication.

This study was partially funded by the Italian Ministry of University and Research, PNRR, mission 4, component 2, investment 1.3 (Partenariati estesi alle università, ai centri di ricerca, alle aziende per il finanziamento di progetti di ricerca di base), title HEAL ITALIA, project number PE00000019, CUP: F53C22000750006 (AP, University of Cagliari). This research was supported by “Fondazione di Sardegna,” grant no. 2023.0160 to the non-profit organization “Associazione per l’Avanzamento della Ricerca per i Trapianti (AART-ODV)”.

BeC was employed by BithiaTec Technologies.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.