Early Predictors of Clinical Deterioration in a Cohort of 239 Patients Hospitalized for Covid-19 Infection in Lombardy, Italy

We described features of hospitalized Covid-19 patients and identified predictors of clinical deterioration. We included patients consecutively admitted at Humanitas Research Hospital (Rozzano, Milan, Italy); retrospectively extracted demographic; clinical; laboratory and imaging findings at admission; used survival methods to identify factors associated with clinical deterioration (defined as intensive care unit (ICU) transfer or death), and developed a prognostic index. Overall; we analyzed 239 patients (29.3% females) with a mean age of 63.9 (standard deviation [SD]; 14.0) years. Clinical deterioration occurred in 70 patients (29.3%), including 41 (17.2%) ICU transfers and 36 (15.1%) deaths. The most common symptoms and signs at admission were cough (77.8%) and elevated respiratory rate (34.1%), while 66.5% of patients had at least one coexisting medical condition. Imaging frequently revealed ground-glass opacity (68.9%) and consolidation (23.8%). Age; increased respiratory rate; abnormal blood gas parameters and imaging findings; coexisting coronary heart disease; leukocytosis; lymphocytopenia; and several laboratory parameters (elevated procalcitonin; interleukin-6; serum ferritin; C-reactive protein; aspartate aminotransferase; lactate dehydrogenase; creatinine; fibrinogen; troponin-I; and D-dimer) were significant predictors of clinical deterioration. We suggested a prognostic index to assist risk-stratification (C-statistic; 0.845; 95% CI; 0.802–0.887). These results could aid early identification and management of patients at risk, who should therefore receive additional monitoring and aggressive supportive care.


Introduction
In late December 2019, clusters of patients with pneumonia of an unknown cause were reported in Wuhan, Hubei Province in China. Subsequently, a novel coronavirus (SARS-CoV-2), causing a severe acute respiratory syndrome, had been isolated from those patients [1][2][3]. In recognition of its global transmission, the World Health Organization declared Covid-19 as a pandemic on 11 March 2020 [4]. On 21 February, the first case of Covid-19 was detected in Italy. As of 23 April 2020, the Italian National Healthcare Service has declared about 190,000 cases of Covid-19 with 25,549 deaths.
Covid-19 has a spectrum of manifestations ranging from asymptomatic infection, to mild upper respiratory symptoms, to bilateral pneumonia with respiratory failure requiring advanced respiratory support [5][6][7][8]. The rapid diffusion rate of the disease in the population, due to asymptomatic carriers, associated with the possible sudden deterioration of clinical conditions requiring critical care admission has contributed to the exceeding of hospital and intensive care units (ICUs) capacity in several Italian hospitals [9][10][11][12]. To increase capacity, elective surgeries were cancelled, semi-elective procedures postponed, and operating rooms turned into makeshift ICUs [9,10,12,13].
Since the current Covid-19 pandemic poses a major strain especially on critical care facilities [9][10][11][12], being able to stratify patient risk at admission would help focus treatment efforts to potentially prevent deterioration [14]. Although recent studies investigating predictors of poor prognosis at an early stage identified potential risk factors including older age, high SOFA score, d-dimer, lymphopenia, and the presence of secondary infection or cardiac disease [8,15], such evidence has not been reported from studies conducted outside of China yet.
In this paper, we described the clinical characteristics and outcomes of a cohort of 239 SARS-CoV-2 positive patients admitted to Humanitas Research Hospital (Milan, Italy) and assessed predictors of clinical deterioration, defined as ICU transfer or death. This study could have significant implications in the clinical management of Covid-19 patients.

Study Design and Participants
This retrospective cohort study included all males and non-pregnant females, 18 years of age or older, consecutively admitted to Humanitas Research Hospital between 22 February and 22 March, 2020, with a laboratory-confirmed diagnosis of Covid-19.
Hospital admission criteria were based on a positive assay for SARS-CoV-2 associated with respiratory failure requiring oxygen therapy, or radiological evidence of significant pulmonary infiltrates on a chest computed tomography (CT) scan, or reduction in respiratory/cardiopulmonary reserve as assessed by a 6 min walking test, or due to frailty related with patient comorbidity.
We assessed a composite outcome of ICU transfer or death. Patients transferred to ICU were those requiring invasive ventilation or non-invasive mechanical ventilation with an oxygen fraction over 60%. Patients with continuous positive airway pressure therapy (CPAP) were followed up by an ICU outreach team and ward physicians in Covid-19 wards. Acute respiratory syndrome (ARDS) was defined according to the Berlin definition [16]. Acute kidney injury (AKI) was diagnosed according to the Kidney Disease Improving Global Outcomes (KDIGO) clinical practice guideline [17].

Laboratory Test, Demographic, and Medical History
Laboratory testing at hospital admission included: complete blood count, renal and liver function (transaminase, total/direct/indirect bilirubin, gamma-glutamyl transferase, alkaline phosphatase), creatinine kinase, lactate dehydrogenase, myocardial enzymes, electrolytes and triglycerides. A panel of acute phase reactant including interleukin-6 (IL-6), serum ferritin, d-dimer, c-reactive protein, fibrinogen procalcitonin was performed routinely. Body temperature, blood pressure, heart rate, peripheral saturation and respiratory rate were measured in all patients. Chest CT scan and arterial blood gas analysis were performed in the emergency department.
Pneumococcal and Legionella urinary antigen test were routinely performed at hospital admission. Nasopharyngeal swab for influenza A, B, H1N1 was routinely performed to exclude co-infection.
Additional microbiological tests were performed (bacterial cultures of sputum, blood and urine) when suggested by clinical conditions. We obtained a comprehensive medical history from patients.
Positivity was assessed with reverse-transcriptase-polymerase-chain-reaction (RT-PCR) assay for SARS-CoV-2 on a respiratory tract sample tested by our laboratory as of 16 March 2020, in accordance with the protocol established by the World Health Organization (WHO; Geneva, Switzerland) [18].
Previously, respiratory specimens were tested at San Matteo Hospital (Pavia, Italy). Due to the high false-negative rate of RT-PCR from a pharyngeal swab, two different swabs were performed in each patient to increase the detection rate [19]. In cases of a negative assay, but suggestive clinical manifestations, presence of contact history or suggestive radiological evidence for Covid-19, the detection was performed on bronchoalveolar lavage fluid or endotracheal aspirate, which has higher diagnostic accuracy.
Demographic, clinical, laboratory, and outcome data were obtained from electronic medical records and were checked by three expert physicians. For a few clinical variables and in the case of a high proportion of missing information, the same physicians manually reviewed the patients' chart notes using a standardized data collection form (Table S1, Supplementary Appendix). The data cut-off was 25 March 2020. The study was approved by the local Ethical Committee, and the requirement for informed consent was waived.

Statistical Methods
Descriptive statistics included means with standard deviations (SD) and medians with interquartile ranges (IQR) for continuous variables, and frequency analyses (percentages) for categorical variables.
To identify risk factors associated with clinical deterioration, leading to ICU transfer or death in hospitalized Covid-19 patients, we used time-to-event (survival) methods for censored observations. The composite study endpoint was "ICU transfer or death" within 20 days from hospital admission. Time to event was defined as the time from hospital admission until the date of event or censoring. Patients discharged early and alive from the hospital, and without having experienced ICU transfer, were considered event-free through day 20 [20]. Kaplan-Meier estimates were used to draw the cumulative incidence curves, compared by log-rank tests, as well as by univariable and multivariable Cox proportional hazards (PH) models of relevant prognostic factors. The analyses were based on non-missing data (missing data not imputed). We followed a standard approach for model selection.
In the univariable Cox PH analysis, a criterion of p ≤ 0.10 was used to identify candidate predictors. Additionally, variables were selected according to a review of the literature and consensus opinion by an expert group of physicians and methodologists. Then, we fitted the multivariable model and used backwards selection procedure to eliminate those variables not significant in the multivariable framework. The criterion of p ≤ 0.05 was used for determining which ones to eliminate. After fitting the model, the PH assumption was examined on the basis of Schoenfeld residuals. The hazard ratios (HR) were presented with their 95% confidence intervals (CI) and the respective p-values.
We developed a preliminary prognostic index for clinical deterioration (i.e., ICU transfer or death) of hospitalized Covid-19 patients by converting the beta coefficients from the multivariable model to integer values, while preserving monotonicity and simplicity. We computed the Harrell's C statistic with its 95% CI to measure the predictive power of our prognostic index. To allow for an easy risk stratification, we divided our score into three groups corresponding to low, intermediate, and high risk, and reported the corresponding observed risk of "ICU transfer or death". Calibration was tested by plotting predicted versus observed Kaplan-Meier curves across the three groups of risk [21][22][23][24][25].
Stata 15.0 software was used to analyze the data (Stata Corp., College Station, TX, USA). P-values less than 0.05 were considered statistically significant. All tests were two-sided. No adjustment for multiplicity was applied.
Some patients showed abnormal blood gas analysis findings (decreased partial pressure of oxygen (PaO 2 , 19.0%) and elevated partial pressure of carbon dioxide (PaCO 2 , 10.0%)). About one third of the patients showed signs of moderate (21.2%) or severe (11.1%) acute respiratory distress syndrome.
Several clinical and laboratory characteristics at admission were associated with clinical deterioration ( Table 2). Advanced age, increased respiratory rate, a low PaO₂, a high PaCO₂, an elevated ratio of PaO₂ to FiO₂ (fraction of inspired oxygen), coexisting CHD, abnormal imaging features in the CT scan, leukocytosis (white blood cell count > 10 × 10⁹/L), lymphocytopenia (lymphocyte count < 1 × 10⁹/L), and elevated levels of certain laboratory parameters (i.e., procalcitonin, interleukin-6, serum ferritin, C-reactive protein, aspartate aminotransferase, serum creatinine, lactate dehydrogenase, fibrinogen, troponin-I, and D-dimer) were statistically significant predictors of clinical deterioration leading to ICU transfer or death in hospitalized Covid-19 patients. Gender, body temperature, body mass index (BMI), and clinical symptoms at admission were not associated with the risk of clinical deterioration (Table 2 and Figure S1, Supplementary Appendix).

Prognostic Index
We developed a preliminary tool to predict clinical deterioration. We provided easy-to-apply instructions to derive the score for each patient given his/her respiratory rate, ratio of PaO 2 to FiO 2 , medical history of CHD, C-reactive protein and serum creatinine levels ( Figure S2, Supplementary Appendix).
The prognostic index showed high predictive accuracy (Harrell's C, 0.845; 95% CI, 0.802-0.887). As a sensitivity analysis, we reassessed the predictive power of this tool after exclusion of patients who died or were transferred to the ICU within the first day of hospitalization. The prognostic index was robust (Harrell's C, 0.812; 95% CI, 0.761-0.870).

Discussion
The emerging Covid-19 pandemic has posed serious strain on Italian hospitals, and, in certain areas, has severely challenged the capacity to provide adequate care to all patients [9,10,12]. Besides supportive care, no effective medications have been identified to date. Although close follow-up is sufficient to manage most non-severe cases, aggressive treatments requiring hospital admission and intensive care are needed in more than 20% of cases [5,7,26]. As ICU bed availability ranges from 4 to 20 per 100,000 population in most western countries [27], even just 1% of patients requiring ICU transfer could easily overwhelm hospital surge capacity. In this scenario, rationalizing the available healthcare resources is of paramount importance to guarantee adequate care to the highest number of patients [11,12]. In this study, we reported the clinical characteristics and outcomes of a cohort of 239 patients with confirmed Covid-19 infection at the Humanitas Research Hospital, and identified predictors of clinical deterioration at admission, with the aim to assist risk-stratification of newly admitted patients.
The mean age of our patients was 5-10 years higher than what was reported in previous studies conducted in China [7,8]. This factor alone probably justifies the higher mortality reported in this Italian cohort (15.2%), as age has been associated with mortality [7,8]. Whilst Covid-19 has affected more males than females with a ratio of approximately 2:1, as confirmed in other studies [1,5,7,8], gender was not associated with clinical deterioration. Similar considerations apply to BMI, which was in the range of overweight in most Covid-19 in-hospital patients.
The oxygenation index PaO 2 /FiO 2 was independently associated with risk of clinical deterioration. Patients with PaO 2 /FiO 2 < 100 were 12 times more likely to deteriorate than patients with values > 300. PaO 2 /FiO 2 is the most widely used among oxygenation indexes, and is included in both the acute respiratory syndrome definition and in sepsis management guidelines [16,28]. Although PaO 2 /FiO 2 ratio has several limitations [29,30], it has been independently associated with mortality in the other cohort of acute respiratory failures [31]. PaO 2 /FiO 2 and other oxygenation indexes are included in the most used indices for mortality prediction, Sepsis-related Organ Failure Assessment (SOFA) and Acute Physiology and Chronic Health Disease Classification System II (APACHE II) score [32,33]. Respiratory rate is one of the simplest clinical indexes, and is also included in APACHE II and in Q-SOFA (quick-SOFA) score [33,34]. In Covid-19 patients, a respiratory rate over 24 breaths per min has been previously associated with mortality [35], in agreement with our findings.
Pneumonia and septic shock have been closely associated in other populations [36]. This is not confirmed in Covid-19 patients according to our findings and previous published data. Guan et al. [37] reported an incidence of septic shock as low as 1% in Covid-19 patients. This was consistent with our data, yielding no association between hemodynamic data at admission and worsening of clinical conditions. Coronary heart disease, respiratory rate, lymphocyte count, lactate dehydrogenase, creatinine, creatine kinase, troponin I, D-dimer, ferritin, IL-6 and procalcitonin were associated with the risk of clinical deterioration, similarly to what was found by Zhou et al. [8]. Coronary heart disease was a strong predictor of deterioration in our cohort, and has been associated with poor clinical outcomes in other respiratory viral infections [38]. Troponin I was elevated in about one fourth of patients compared to half patients in other reports, and was a strong predictor of ICU transfer or death [8]. Most studies describing critically ill patients very frequently reported an increase in troponin levels without a clear association with myocardial dysfunction or myocarditis in ICU patients [39]. The increase of troponin is not clear evidence of myocarditis or myocardial infarction; it could be related to a direct viral damage of myocardium (through ACE2 expressed by myocardiocytes), a hyperinflammatory syndrome, cardiac microvascular damage or hypoxia-induced myocardial injury [38]. A slight increase of troponin levels can also be attributed to the high frequency of acute kidney injury in Covid-19 patients.
A rise in serum creatinine doubled the hazard for deterioration in our population. Serum creatine at admission is a marker for AKI, which is strongly associated with mortality in hospitalized patients [40]. Zhou et al. [8] reported a 50-fold higher frequency of AKI among non-survivors. The etiology of Covid-19 related AKI is yet to be determined, but we hypothesize that it could be due to dehydration after several days of high fever and diarrhea, to a high level of positive end expiratory pressure, multiorgan failure, or direct viral cytopathic injury of renal cells.
Imaging features obtained through a chest CT scan were strong predictors of outcome, however we decided not to include them in our model because this diagnostic tool may not be available in all countries or hospitals. Accordingly, several societies including the Royal College of Radiology currently report no role for the CT scan in Covid-19 diagnosis [41]. In our multivariable model, inclusion of chest CT scan added a negligible increase in the tool's predictive power.
Our study has limitations. First, since the data were retrospectively collected, we could not assess all laboratory parameters in all patients, including IL-6. Second, we considered eligible all consecutively admitted Covid-19 patients, irrespective of the latency since symptom onset. As a consequence, we also included patients who required ICU transfer very early after admission. We acknowledge that these patients would not benefit from a risk-stratification at admission. However, we were able to substantially verify the validity of our risk-stratification tool after excluding these severely ill patients. Third, we conducted multiple comparisons, and this could increase type I error, thus our analyses should be interpreted with due caution. Finally, our prognostic index possibly needs to be improved by external prospective cohorts, to account for potential differences in healthcare systems, measurement methods, definitions of predictors, and subject characteristics or context.
Our study has several strengths. First, we had at our disposal a database rich in clinical, laboratory and imaging findings, including CT, in more than two hundred patients with few missing data, which may be difficult to find in other settings. We applied appropriate time-to-event methodology and developed a simple and easy-to-apply clinical scoring system; its C-index of 0.845 (0.802-0.887) demonstrates excellent predictive power.
One of the causes of the Italian Covid-19 crisis, is probably the initial massive admission to hospitals of patients with low risk of severe deterioration, coupled with many hospitals being unprepared to identify and contain such a massive flux of patients requiring respiratory support. Our study provides novel information about Covid-19 and its clinical outcomes. There is an urgent need to guide patient management and focus treatment efforts, especially in conditions where the system is overwhelmed. We believe that our results could aid early identification and management of patients at risk of clinical deterioration, who should, therefore, receive additional monitoring and aggressive supportive care.