Kaoutar Azijli Abdellaoui

In this thesis we wanted to find answers to the following questions: Is it possible to better predict the outcomes of patients with suspected infection at the ED? How can we differentiate between bacterial and viral infection? How can we diagnose bacterial super infection? While we were busy with our journey to answer these questions the world was struck by the COVID-19 pandemic and we had to halt some of our projects temporarily and started working projects to improve the care for COVID-19 patients.

At the beginning of this pandemic no reliable rapid antigen tests were available and there was a need for a reliable diagnostic test. Therefore, we decided to add additional research questions such as: Does the chest CT and the lung ultrasound have a place in diagnosing COVID-19? And how can we diagnose a bacterial infection in COVID-19 patients? How can we measure disease severity in COVID-19 patients attending the ED?

Early warning scores (EWS) were designed to predict chance of deterioration in hospitalised patients. In 2016 qSOFA was introduced as part of the new sepsis criteria. In chapter 3 we started a prospective multicentre study to compare the performance of qSOFA, NEWS, MEWS and SIRS in the ED in patients with suspected infection. All patients 18 years and older and with suspected infection were included. We compared the scores of NEWS ≥5, MEWS ≥3 , qSOFA ≥ 2 and SIRS ≥ 2 to predict 30 days mortality and ICU admission. NEWS ≥5 showed to have the best balance between sensitivity (75.8% (95% CI 63.3-85.8), specificity (65.9% (95% CI 63.2-68.5) and NPV (98.2% (95% CI 97.3-98.9) for early identification of high risk patients with suspected infection in the ED.

In chapter 4 we conducted an observational, retrospective multicenter cohort study. All patients aged 18 and over with suspicion of COVID-19 who visited the ED were enrolled. Multiple variables were collected and analysed as potential risk factors for poor outcome. Ten predictors of poor outcome were identified, namely current malignancy (OR 2.774; 95% CI 1.682-4.576), systolic blood pressure (OR 0.981; 95% CI 0.964-0.998), heart rate (OR 1.001; 95% CI 0.97-1.028), respiratory rate (OR 1.078; 95% CI 1.046-1.111), oxygen saturation (OR 0.899; 95% CI 0.850-0.952), body temperature (OR 0.505; 95% CI 0.359-0.710), serum urea (OR 1.404; 95% CI 1.198-1.645), C-reactive protein (OR 1.013; 95% CI 1.001-1.024), lactate dehydrogenase (OR 1.007; 95% CI 1.002-1.013) and SARS-CoV-2 PCR result (OR 2.456; 95% CI 1.526-3.953). The AUC was 0.86 (95% CI 0.83-0.89), with a Brier score of 0.32 and, and R2 of 0.41. The AUC in the external validation in 500 patients was 0.70 (95% CI 0.65-0.75). The COVERED risk score showed an excellent discriminatory ability also in an external validation.

In chapter 5 we aimed to prospectively validate the COVID-19 Reporting and Data System (CO-RADS) at the ED and we analyzed whether the CT-severity Score (CTSS) was associated with hospital admission, ICU and mortality. Diagnostic accuracy measures were calculated for CO-RADS using PCR as reference. The area under the curve (AUC) of 0.91 (95% CI 0.89-0.94) was found for CO-RADS using PCR as reference. The optimal CO-RADS cutoff was 4 with a sensitivity of 89.4% (95% CI 84.7-93.0) and specificity of 87.2% (95% CI 83.9-89.9). A significant association between CTSS and admission, ICU and 30 day mortality was found, adjusted ORs per point increase in CTSS were 1.19 (95% CI 1.09-1.28), 1.23 (95% CI 1.15-1.32), 1.14 (95% CI 1.07-1.22) respectively. These findings support the use of CO-RADS and CTSS in triage and diagnosis for patients presenting with possible COVID-19 at the ED.

In chapter 6 we conducted a multicenter prospective observational study where we compared the diagnostic accuracy of lung ultrasound (LUS) with chest CT in suspected COVID-19 patients at the ED. Area under the receiver operating characteristic (AUROC) was 0.81 (95% CI 0.75-0.88) for LUS and 0.89 (95% CI 0.84-0.94) for CT. Sensitivity and specificity for LUS were 91.9% (95% CI 84.0-96.7) and 71.0% (95% CI 61.1-79.6) respectively versus 88.4% (95% CI 79.7-94.3) and 82.0% (95% CI 73.1-89.0) for CT. Agreement between LUS and CT was 0.65 and inter-observer agreement for LUS was good 0.89 (95% CI 0.83-0.93). LUS can safely exclude clinically relevant COVID-19 pneumonia and may aid in COVID-19 diagnosis in high prevalence situations.

In chapter 7 we investigated whether the optimal use of procalcitonin (PCT) is different in patients with and without proven viral infections for the purpose of excluding bacteremia. The study was a prospective observational study where all patients 18 year and older for whom a blood culture and viral test were ordered in the ED were included. PCT had an area under the curve of 0.85 (95% CI 0.80-0.91) for prediction of bacteremia. In patients with proven viral infection PCT < 0.5ug/L had a sensitivity of 100% (95% CI 63.1-100) and specificity of 81.2% (95% CI 75.1-86.3) to exclude bacteremia. This study suggests that a PCT concentration of < 0.5ug/L makes bacteremia unlikely in patients with a viral infection. General Discussion and Future perspectives General Discussion The first part of this thesis focuses on answering the following question: Is it possible to better predict the outcomes of patients with suspected infection? The prospective study in Chapter 3 shows that the NEWS score, with a cut-off of ≥5, is the most accurate at predicting death or admission to the intensive care unit (ICU) in patients who are suspected of having an infection compared to the qSOFA, SIRS, and MEWS scores. This finding is in line with the findings of several other (retrospective) studies. 1-5 We didn't include patients with septic shock in this study because they are hemodynamically unstable and easy to identify. Our main goal was to identify the patient with infection in the early stages and see if these prognostic scores could detect a chance of deterioration at the ED. The first set of vitals parameters was used to calculate the risk scores. This is because decisions regarding treatment and admission often have to be made early and within a short time frame. This is in contrast with other studies, which took multiple samples over time. 2-6 The NEWS is a general predictive score already being used in many EDs, and we recommend using it for early risk stratification in patients with suspected or proven infections at the ED. It would give the emergency physicians an early warning to identify which of the patients are at risk of deterioration in the next 24 hours. During the making of this thesis, the world was struck by the COVID-19 pandemic. The pandemic quickly overwhelmed the health care systems all over the world. The medical world went looking for other methods to try to predict the bad outcomes in these patients. That is why, in Chapter 4, we tried to fill this knowledge gap by developing a prediction model that would predict poor outcomes (30-day mortality or ICU admission). We externally validated our final COVERED risk score, which contained 10 parameters, in an Italian cohort. It showed excellent discriminatory capacity with an AUC of 0.81, and it may aid in clinical decision-making. During the first wave of the COVID-19 pandemic, many prediction models were developed. 7-8 A systematic review by Wynants et al. recommended against the use of any of these models in clinical practice. 9 The previous models were poorly tested in a real-life setting, reported a high risk of bias, and were overly optimistic. Our model was developed in a multi-center setting with a large heterogeneous sample size (N = 1193) compared with previous models (sample size of 26–577). Furthermore, we externally validated the COVERED score in an independent international cohort of 500 patients. 7-9 We did not categorize continuous variables because it caused significant data loss, nor did we pool risk into prognostic risk groups because it introduces major problems in model analysis and interpretation. 10-11 However, our prediction model is not used currently due to the fact that by the time the study was finished, the pandemic was in decline. It may be possible to predict the outcomes of patients with suspected infections by using simple early warning scores rather than prediction models. This is because a prediction model is difficult to implement and generalize. Early warning scores, like NEWS, are already widely used in daily practice. In practice, we can use early warning scores to alert us earlier. However, given their moderate specificity, it is necessary to combine them with the passage of time and clinical judgment. The second part of this thesis focuses on answering the following questions: How can we differentiate between bacterial and viral infections at the ED? How can we diagnose a bacterial super infection? Do chest CT and lung ultrasound have a role in diagnosing COVID-19? At the beginning of the pandemic, no reliable rapid antigen tests were available. There was a need for a rapid and reliable diagnostic test at the ED. Chest radiology (CXR) was used to show whether there were any signs of pulmonary involvement. 12 The problem with CXR is that it has poor diagnostic accuracy for COVID-19 pneumonia. That is why the clinicians used chest computed tomography (CT) to diagnose patients with COVID-19. 13-14 The Radiological Society of the Netherlands (NVvR) introduced the COVID-19 Reporting and Data System (CO-RADS) and the corresponding CT severity score (CTSS) to analyze and report the CTs in a systematic and reproducible manner. However, there was no external validation of this system. In chapter 5, we prospectively validate the CO-RADS as a COVID-19 diagnostic tool at the ED and determine if the CTSS is associated with prognosis. The study found that CO-RADS could clearly tell the difference between a positive and negative PCR result (AUC of 0.91 (95% CI: 0.89–0.94)) and that the CTSS at the ED presentation was linked to hospital admission, ICU admission, and death within 30 days. The CO-RADS is an easy-to-use tool to assist radiologists and clinicians in diagnosing COVID-19 in patients with moderate to severe symptoms presenting to the ED. This study was conducted in a high-prevalence setting, and it is unclear how it will function when the pandemic subsides. Chest CT was used to diagnose patients with COVID-19, but its role was suboptimal due to its higher radiation dose and elevated cost, and not all hospitals had access to a CT 24x7 at the ED. This was why the bedside lung ultrasound attracted attention. The advantages of ultrasound are that it is radiation-free, can be performed at the bedside in a few minutes, and the interpretation is real-time for multiple diagnoses. And it could be used in low-income countries that did not have easy access to a CT. Despite these advantages, the literature shows different values of sensitivity and specificity. 15-17 In chapter 6, we show that lung ultrasound (LUS) and chest CT have equal diagnostic accuracy for COVID-19 pneumonia, with AUROC between 0.8 and 0.9. LUS can help during triage by excluding clinically relevant COVID-19 pneumonia at the ED and may aid in the diagnosis of COVID-19 in a high-prevalence setting. It may prove especially useful in situations where CT or PCR results are not readily available. The results of this study are in line with the literature on the diagnostic accuracy of LUS in acute respiratory failure caused by different etiologies, ranging from pneumonia to ARDS. 17-20 However, LUS had a lower specificity than CT (71.0% (95% CI 61.1–79.6)). This could be due to the fact that the sonographic features of COVID-19 are not exclusive to COVID-19. They are also observed in other diseases such as (viral or atypical) pneumonia, acute respiratory distress syndrome (ARDS), heart failure, and atelectasis, but this is also the case in chest CT, but to a lesser extent. 15–20 LUS was useful in the COVID-19 pandemic to exclude severe COVID-19. But outside the pandemic, it would be difficult to make a differentiation because the sonographic features of COVID-19 are not exclusive to COVID-19. The use of imaging in diagnosing infections is not useful in all kinds of infections, and it definitely does not tell the physician whether a bacteria or a virus is responsible for the infection. One of the hardest parts for an emergency physician at the ED, especially during the winter season in the Netherlands, is to distinguish between a viral and bacterial cause of respiratory complaints. The rate of bacterial (co) infection in viral infections is highly variable. In influenza, this varies between 2-65%. That is why antibiotics are often started empirically in patients with suspected respiratory infections. This was also the case during the COVID-19 pandemic, especially during the first wave. A meta-analysis by Langford et al. reported a rate of bacterial (co)infection in patients with severe COVID-19 at presentation of 3.5%, which might increase to 16% during the course of the disease. 5 Estimated rates of unnecessary antibiotic use vary between 30 and 60%; this is the most preventable cause of antibiotic resistance in general. 21-24 In the literature, it has been proposed to use PCT for identifying a bacterial (co)infection; however, the studies show conflicting results. One of the reasons that is given for the conflicting results is the use of different cutoffs (<0.10 ug/l, 0.25 ug/l, etc.). In a study by Schuetz et al., it was advised to take the pre-test likelihood of bacterial infection into account to choose the appropriate cutoff. 25 The study described in Chapter 7 investigated if the PCT cutoff to exclude bacteremia should be different in patients with and without confirmed viral infections presenting during a viral epidemic or pandemic. The study found that PCT performed better in patients with a viral infection with a higher cutoff than in patients without a viral infection. If we know that the patient has a very low chance of bacteremia, we might be able to make better decisions on antibiotic use. In fact, if we had used the PCT cutoff point of < 0.25 µg/L as guidance for antimicrobial therapy, antibiotic use would have decreased from 50.9% to 21.9% in patients with a viral infection. However, in the group without viral infection, a cutoff of < 0.25 ug/L would have resulted in a much smaller reduction of 54%–42%. A pitfall of PCT is that it can sometimes take up to 8–24 hours before reaching high values. 26-27 If patients present shortly after developing symptoms, then PCT could be low. To reduce the risk of missing these bacterial (co)infections, it would be beneficial to repeat the PCT testing. This is in line with previous recommendations. 28 However, given the low rate of (co)infection in our cohort, this finding requires confirmation in a larger population of patients with viral infections. Main conclusion These are the main conclusions of this thesis. • It is possible to predict the outcome of patients with suspected infection using simple early warning scores, particularly NEWS, rather than prediction models. • Lung ultrasound can help in triage by excluding clinically relevant COVID-19 pneumonia at the ED and may aid in the diagnosis of COVID-19 in a high-prevalence setting. • PCT performed better in patients with a viral infection with a higher cutoff than in patients without a viral infection. Future perspectives This thesis demonstrated that it is possible to predict the outcome of patients with suspected infection using simple early warning scores, particularly NEWS, rather than prediction models. We can use it to alert the ED doctors early. However, given their moderate specificity, it is necessary to combine them with the passage of time and clinical judgment. Integrating automated scoring systems with trend analysis and continuous data monitoring can provide valuable insight into patient trajectories. It is critical to collaborate with other countries and healthcare systems when developing a prediction model. This could allow for a better understanding of the generalizability and implementation of prediction models across different settings and populations. This approach could greatly improve the applicability and robustness of prediction models in routine care. 29-33 The COVID-19 pandemic has taught us the importance of having rapid diagnostic tests in the emergency department to quickly differentiate between bacterial and viral infections. It has also underscored the significance of being able to assess investigations systematically and universally. It is important in the future to have a systematic approach for interpreting radiological images to make it possible to address challenges posed by a pandemic like COVID-19 and optimize patient care. Given the need for quick decision-making regarding treatment at the ED, there is a demand for tests with rapid results. Prompt administration of antibiotics is essential for bacterial infections in suspected cases. However, current guidelines rely on cultures, which often take 24 to 48 hours for results, and only 11% of blood cultures taken in the emergency department give positive results. Thirty percent of patients in the emergency department receive unnecessary antibiotics. 9,34 Ideally, a technique providing rapid results to diagnose bacterial infection is preferred. IS-pro appears to be a promising technique for searching for a specific piece of bacterial DNA present in all bacteria, with its length specific to the bacterial species. 35-36 To further differentiate between different bacteria, fluorescent dyes were added, giving each bacterial family its own distinct color. This technique can detect and identify nearly all known bacteria within five hours. Currently, efforts are underway to reduce this time to the point-of-care level. Machine learning is a trendy and hot new topic in the medical world. It is increasingly being used in various ways to improve patient care. Currently, we are conducting the ABC study (appropriate use of blood cultures in the emergency department through machine learning). 37 This is a multicenter trial aimed at reducing blood cultures in the emergency department. After randomization, an algorithm will calculate the likelihood of a positive blood culture in the intervention group. If the likelihood of a positive blood culture is low, the blood culture will be canceled. If the patient is randomized to the control group, the blood culture will be analyzed as usual. The reason for this study is because we take too many unnecessary blood cultures in the ED. Only 11% of the blood cultures are positive, and up to 50% of them contain contamination. 9,34 This has far-reaching consequences, such as longer hospital stays, more unnecessary diagnostics, and antibiotic use. We have developed a tool through machine learning that can predict the likelihood of a positive blood culture. This can assist clinicians in deciding whether or not to take a blood culture, thereby reducing unnecessary blood cultures. In the context of antibiotic use, machine learning can assist in interpreting diagnostic results and, therefore, optimize antibiotic prescribing. Machine learning has immense potential for optimizing care for patients with suspected infections and may shape the future of emergency medicine. With all these developments, we are shifting towards personalized and evidence-based patient care at the emergency department.