Introduction

The presence of a new severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) was first reported in China and has subsequently spread to all regions of the world, straining the health systems of many countries.¹ Viral pneumonia associated with SARS-CoV-2 has been officially denominated as coronavirus disease 2019 (Covid-19).

Approximately 5 to 33% of patients with Covid-19 pneumonia will be admitted to an intensive care unit (ICU).^2,3,4 Previous studies have identified several risk factors associated with a severe course of Covid-19 pneumonia and its progression to acute respiratory distress syndrome.^5,6 These risk factors may be categorized into patient characteristics obtained through the medical interview (e.g. age, comorbidities and symptoms);^5,6.7.8.9.10 vital signs obtained from triage evaluation (e.g. respiratory rate and arterial pressure);^11,12 laboratory abnormalities including inflammatory, coagulation and organ-specific studies (e.g. lactate dehydrogenase, D-dimer, fibrinogen, cardiac troponins, liver function tests, among others),^{5,6,8,9,10,13} and lung imaging findings (e.g. number of affected lobes and estimated pneumonia extension).^14,15

In several countries, including Mexico, hospital reconversion and temporary care centers have been implemented to cope with a large number of Covid-19 patients.¹⁶ However, many of these centers are prepared to care for patients with supplemental oxygen requirements in general wards, but just to a few patients who require admission to an intensive care unit (ICU) and mechanical ventilation.

The study aimed to derive and validate simple risk prediction models to anticipate the need for admission to an ICU. These models include clinical, laboratory, and image findings obtained at the first-contact evaluation to aid triage, decision-making, and timely referral to maintain healthcare system capacity.

Materials and methods

This is a prospective observational cohort study. All consecutive adult (>18 years) patients hospitalized at Instituto Nacional de Ciencias Médicas y Nutrición Salvador Zubirán, a referral center for Covid-19 patients at Mexico City, were evaluated for this analysis. The study was approved by the local Research and Ethics Board (CAI-3368-20-20-1) that waived the use of a written informed consent form due to the study’s nature.

All patients hospitalized between March 16^thand April 30^thwere assigned to the development cohort, and patients hospitalized between May 1^stand May 21^thwere assigned to the validation cohort. The study population was segregated into two groups: 1) patients who required admission to an ICU at any time during their hospitalization, and 2) patients admitted to hospital general wards who were discharged from hospitalization without ever been considered for admission to an ICU. Patients were censored for each group once they were admitted to the ICU or by the date of discharge from hospitalization to home, respectively. All patients who remained hospitalized in general wards by the end of the study were not included in this analysis to avoid bias, as it was considered that their risk for ICU admission was still active.

All patients had Covid-19 pneumonia diagnosed by chest computed tomography (CT), and SARS-CoV-2 infection confirmed in respiratory specimens using real-time reverse-transcriptase polymerase chain reaction (RT-PCR) assay by local testing and confirmed at a central laboratory. Patients were considered positive if the initial test results were positive, or if it were negative but by repeat the test were positive.

The following variables were obtained by the time of the triage and emergency department (ED) evaluation: demographic variables, previous medical and medication history including smoking; symptoms; physical examination including weight, height, and vital signs; laboratory evaluation including arterial blood gas analysis, inflammatory biomarkers, troponin-I levels, complete blood count and blood chemistries; and chest CT-scan findings. All patients underwent chest CT-scans that were evaluated by experienced specialists. Lung involvement was semi-quantitatively classified as mild (<20%), moderate (20-50%) or severe (>50%).

For each patient, we calculated the Charlson comorbidity index (an index that predicts 10-year survival in patients with multiple comorbidities),¹⁷ National Early Warning Score (NEWS) 2,¹⁸ Sequential Organ Failure Assessment (SOFA) score,¹⁹ CURB-65 score for pneumonia severity,²⁰ the MuLBSTA score for viral pneumonia mortality,²¹ the ROX index to predict the risk of intubation,²² the CALL score model for prediction of progression risk in Covid-19,²³ and the COVID-GRAM score for the risk of critical illness in Covid-19.²⁴ All scores were calculated at admission to the ED. Although some of these scores were not derived to evaluate the risk of admission to an ICU, we evaluated their value if repurposed for this end.

The need for admission to critical care was determined by the medical team in charge of the patient and included the need for mechanical ventilation or high-dose vasopressors. Patients hospitalized in medical wards were treated with supplementary oxygen (nasal cannula or non-rebreathing oxygen mask) but were never considered for intubation and ICU admission. All patients in the latter group were censored once discharged from hospitalization to home after improvement. Clinical outcomes for patients admitted to ICU were monitored up to May 25^th, 2020.

Results

Patient characteristics at admission

A total of 705 patients with Covid-19 pneumonia and SARS-CoV-2 positive test were hospitalized during the study period. Eighty-three patients who remained hospitalized in general wards, 28 patients referred to a convalescence center, 22 early referred to other institutions, and three patients who were discharged by discharge against medical advice were excluded from the analysis. The development cohort included 329 patients and the validation cohort 240 patients (figure 1).

ICU: intensive care unit

Figure 1: Consort flow diagram showing all subjects included in the study. Mexico City, 2020 

Derivation of the ABC-GOALS scores

A predictive point score was constructed based on the weighted OR’s from each logistic regression model. The derived scores were defined as the ABC-GOALS (Arterial pressure, Breathlessness, Charlson, Glucose, Obesity, Albumin, LDH and S/F ratio) and were labeled as clinical (ABC-GOALS_c), clinical plus laboratory (ABC-GOALS_cl) and clinical plus laboratory plus x-ray (ABC-GOALS_clx) scores (figure 2).

All three ABC-GOALS score variations demonstrated good accuracy in estimating the risk of admission to an ICU, with an area under the curve of 0.79 (95%CI 0.74-0.83), 0.86 (95%CI 0.82-0.90) and 0.88 (95%CI 0.84-0.92) for the ABC-GOALS_c, ABC-GOALS_cland ABC-GOALS_clxscores, respectively (supplementary figure 1²⁵). Calibration plots showed good agreement between the estimated and observed scores (supplementary figure 2²⁵).

By segregating the scores into three levels of risk of admission to ICU, the ABC-GOALS_cclassified patients into low- (0-2 points, mean risk 8%, 95%CI 7-9%), moderate- (3-7 points, mean risk 32%, 95%CI 30-34%) and high-risk (≥8 points, mean risk 73%, 95%CI 70-75%) of admission ICU. The respective groups for the ABC-GOALS_clwere low- (0-3 points, mean risk 4%, 95%CI 3-4%), moderate- (4-9 points, mean risk 30%, 95%CI 27-32%) and high-risk (≥10 points, mean risk 80%, 95%CI 78-82%) for ICU admission. Finally, the groups for the ABC-GOALS_clxwere low- (0-4 points, mean risk 6%, 95%CI 5-6%), moderate- (5-9 points, mean 32%, 95%CI 29-34%) and high-risk (≥10 points, mean 80%, 95%CI 78-82%) for ICU admission (figure 2).

ICU: intensive care unit; LDH: lactate dehydrogenase; ABC-GOALS clinic: score including only clinical variables; ABC-GOALS clinic+laboratory: score including both clinical and laboratory variables; ABC-GOALS clinic+lab+x-ray: score including clinical, laboratory and x-ray findings

Figure 2 Guide to ABC-GOALS scores calculation. Mexico City, 2020 

The ABC-GOALS variants outperformed other scores that were repurposed for the prediction of admission to ICU, based on the area under the curve values (table III). As previously stated, most of these scores were not created to predict this outcome. For practicality, the three ABC-GOALS scores are planned to be implemented into an application for mobile devices.

Discussion

In this study, we derived and validated a prognostic score that, evaluated upon patient admission to the ED, helps predict the probability of admission to ICU during patient hospitalization. This tool may prove useful for patient assessment at first-contact evaluation and to timely refer patients to other units in case of a lack of or overcrowding of local ICUs. The ABC-GOALS scores predicted the need for admission to an ICU with greater accuracy than other reported Covid-19 prognostic scores, as well as other scores that were created for the prediction of other outcomes in sepsis or pneumonia, such as mortality.

To date, more than 11 million SARS-CoV-2 infections have been reported worldwide with more than 500 000 total deaths.²⁶ The fast spread of the disease²⁷ and the large number of patients admitted to hospitals have strained and overwhelmed local health systems. Among hospitalized patients, 5 to 33% will require admission to an ICU and 75% to 100% of them will require support with mechanical ventilation.^2,3,4,7 An effective evaluation of patients with severe disease may be critical to maintaining the healthcare system functional for as long as possible.

The ABC-GOALS scores summarize many of the previously reported factors that have been shown to independently associate with severe Covid-19 disease. We found that age and comorbidities predictive performance is enhanced when integrated into the Charlson comorbidity index. It has been reported in other studies that there may be interactions between age and some comorbidities such as diabetes.²⁸ Also, the Charlson comorbidity index offers the advantage of integrating several comorbidities into one score that has been previously associated with survival.¹⁷ Among clinical symptoms, dyspnea or shortness of breath is the most consistently reported symptom associated with Covid-19 severity and mortality^5,8,11,29 and was independently associated with ICU admission in this study. Obesity is a long-recognized factor for severe pulmonary infections³⁰ and has been consistently associated with adverse outcomes in Covid-19.³¹ Low systolic arterial pressure and respiratory rate at triage evaluation complete the clinical ABC-GOALS.

Among the three derived scores in this study, the clinical ABC-GOALS has an inferior performance and probably equivalent to other early warning scores. It was designed to provide a tool for areas without timely access to laboratory results. However, the clinical score remains to be externally validated, as some reported models based exclusively on demographic (but not physical examination) information have been shown to decrease their performance after external validation.¹¹

The scores integrating clinical with laboratory (ABC-GOALS_cl) and imaging (ABC-GOALS_clx) findings seem to perform better for prediction of admission to an ICU. Both models include values of glucose, lactate dehydrogenase, and oxygenation at admission that have been previously reported to be associated with prognosis in Covid-19.^8,10,32,33 Lactate dehydrogenase has been previously integrated into a predictive model (see below). In contrast to other laboratory tests that may increase during the progression of the disease, such as D-dimer and cardiac troponin levels,^34,35 LDH has proved to be the most robust laboratory predictor for admission to ICU when evaluated at first-contact. Glucose and oxygenation were also included in these models. Oxygenation, reflected by the S/F ratio, probably reflects the disease extension along with the imaging findings. It is worth noting that the model that included imaging findings (ABC-GOALS_clx) had an equivalent performance to the clinical and laboratory model, therefore, it may not be necessary to perform imaging in a patient who is planned to be referred based on the clinical plus laboratory model (ABC-GOALS_cl).

We evaluated two scoring models that have been derived to predict the risk of Covid-19 progression: the CALL score²³ and the COVID-GRAM score.²⁴ Both were derived from Chinese populations. As shown, both scores had lower performance than the ABC-GOALS scores in our population. The CALL score²³ includes the presence of any comorbidity, age, lymphocyte count, and LDH levels. Most patients in our study presented with lymphocyte count below the cut-off set in the CALL score, therefore reducing the predictive value of the score and of this specific parameter. For the COVID-GRAM score,²⁴ it includes X-ray abnormalities, which were present in all of our patients by lung CT-scan. It also includes rare manifestations of Covid-19 that were barely documented in our population, such as hemoptysis and unconsciousness. Both, the CALL and COVID-GRAM scores, were derived from populations with a lower risk of critical care admission than our population, as 19.2% and 8.2% of the patients developed critical illness, respectively, compared to 34.9% and 47.9% in our development and validation cohorts, respectively. Thus, specific items included in the scores plus the derivation from a population with a lower risk for ICU admission may explain the lower performance of these scores in our population.

We showed that repurposing of other scores employed in pneumonia may not perform adequately to predict the need for admission to ICU in Covid-19 pneumonia. For example, the MuLBSTA score²¹ has been validated for mortality prediction in viral pneumonia but may have a lower performance in Covid-19 pneumonia. This score includes the number of affected lobes by viral pneumonia, however, all patients in our study showed multi-lobar lung infiltrates reducing the performance of this parameter. Other variables included in this score, such as bacterial coinfection, may not perform well when evaluated at admission, as bacterial coinfection more frequently occurs later in the disease evolution.

This study has some strengths. First, all data were collected prospectively with very few missing data for all collected variables. Second, there was an appropriate number of patients in the group that required admission to an ICU, which allowed the study of multiple predictors. Third, the study is reported based on TRIPOD guidelines.³⁶

There are limitations to this study. The study included a single-center and very sick population with high rates of ICU admission; therefore, the score may perform well in severe hospitalized patients with Covid-pneumonia. There is still a need for external validation. Another limitation is that medical practice and admission criteria to ICU may vary between institutions and countries, especially in stressed health systems. Although a significant number of patients was excluded because they were still undergoing continuing care, it would be expected that their clinical and laboratory parameters would behave as in the patients included with few effect in the prognostic models derived. This score was derived from a population living in Mexico City at >7 300 feet over the sea level. The mean partial pressure of oxygen at this altitude has been estimated at around 66mmHg (estimated normal baseline P/F ratio ≈ 314).^37,38 Therefore, we used the S/F ratio to account for respiratory compensation and the right-shift of the hemoglobin dissociation curve that takes place at higher altitudes. Finally, the thresholds set to define the low-, moderate- and high-risk groups should be adapted to local needs and ICU availability. Therefore, we provided the estimated percentages of ICU admission for each sum of the score.

In summary, the ABC-GOALS score represents a tool to evaluate patients with Covid-19 at admission to the ED, designed to timely predict their risk of admission to an ICU. This score may help early referral and planning of attention during the Covid-19 pandemic.