Introduction

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) causes a wide spectrum of clinical manifestations, named coronavirus disease 19 (COVID-19), which range from asymptomatic infections to severe interstitial pneumonia. Unfortunately, the COVID-19 pandemic has led to a large number of critically ill patients in a very short time and is currently overwhelming the healthcare systems worldwide [1].

In this scenario, delays in intensive care unit (ICU) transfers [2,3,4] and in patient intubations [5, 6] have been associated with increased mortality among COVID-19 patients with hypoxemic acute respiratory failure (ARF). Thus, correct triage and prompt ICU allocation of patients scheduled to receive intubation in case of oxygen therapy or noninvasive ventilation failure are crucial to achieve an effective COVID-19 pandemic response [7].

Potential prognostic factors of intra-hospital mortality and intubation in patients hospitalized due to COVID-19-related hypoxemic ARF include older age [8, 9], male gender [9], diabetes [9], prothrombin time and D-dimer level [10], lymphocytopenia [11,12,13], leukopenia [11,12,13], C-reactive protein [11,12,13], lactate dehydrogenase [11,12,13], sequential organ failure assessment (SOFA) score [14], arterial oxygen pressure on inspired oxygen fraction ratio (PaO2/FiO2) [8, 14], and acute kidney injury [14].

The Fleischner Society Statement on Chest Imaging and COVID-19—issued on April 7th, 2020—recommends chest computerized tomography (CT) imaging for the triage of (i) patients with suspected COVID-19 presenting with moderate to severe clinical features and a high pretest probability of disease, (ii) COVID-19 patients with worsening respiratory status, and (iii) patients with functional impairment and/or hypoxemia after recovery from COVID-19 [15, 16]. Fittingly, a single-center study has recently shown how the compromised lung volume estimated by quantitative CT analysis is a strong predictor of the need for oxygenation support and intubation among COVID-19 patients [17]. Thus, an algorithm combining demographic and early clinical characteristics together with laboratory findings and chest CT analysis results may favor the early identification of COVID-19 patients requiring invasive mechanical ventilation (IMV) or at increased risk of mortality.

Our primary aim was to retrospectively identify the risk factors associated with intubation and intra-hospital mortality in a cohort of COVID-19 patients admitted to the hospital for hypoxemic ARF by analyzing their demographic characteristics, early clinical and laboratory variables, and quantitative chest CT analysis results. As a secondary endpoint, we sought to determine the performance of a newly developed algorithm based on the aforementioned variables in predicting the probability of IMV and intra-hospital mortality in our study population.

Methods

Patients and data

The present investigation is an observational retrospective single-center study. Ethical approval was issued by the Comitato Etico Interaziendale Novara, Italy (Chairperson Prof. G. Zulian) on May 20th, 2020 (Ethics Committee No. CE 121/20). The requirement for informed consent was waived due to the retrospective nature of the study.

We analyzed 187 consecutive patients with COVID-19 pneumonia diagnosed with real-time reverse transcriptase-polymerase chain reaction (RT-PCR) nasopharyngeal swabs, subjected to chest CT images, and admitted to the subintensive and intensive care units of our hospital between March 1st and April 30th, 2020. Patients with poor-quality chest CT images were excluded. The study was reported in accordance with STROBE guidelines.

Clinical and laboratory characteristics

Demographic information, body mass index (BMI), time from first-symptoms, comorbidity, date of admission to hospital, clinical laboratory on admission including the first PaO2/FiO2 ratio, arterial oxygen saturation (SpO2), blood cell counts (i.e., leukocyte and lymphocyte count), biomarkers of inflammation (i.e., lactate dehydrogenase, ferritin, C-reactive protein, procalcitonin, fibrinogen), glycemia, and troponin were collected. Furthermore, we recorded the type of oxygen assistance administered—i.e., standard oxygen therapy (low-flow oxygen nasal cannula, Venturi mask, non-rebreathing mask), noninvasive ventilation [continuous positive airway pressure (CPAP) or bilevel positive airway pressure (BiPAP)], or invasive mechanical ventilation. All data were derived from both electronic hospital records and digitization of paper documents.

Criteria for intubation

Criteria for intubation were cardiac or respiratory arrest; inability to protect the airway; coma or psychomotor agitation; unmanageable secretions or uncontrolled vomiting; life-threatening arrhythmias or electrocardiographic signs of ischemia; hemodynamic instability, defined as systolic arterial pressure < 90 mmHg despite adequate filling or use of vasoactive agents; intolerance to all interfaces; dyspnea during noninvasive ventilation administered as CPAP or BiPAP; respiratory rate > 30 breaths/min; SpO2 < 92% during CPAP or BiPAP; and acidosis.

Quantitative CT analysis

CT scan was performed within 1 day from admission. CT images were independently reviewed by two radiologists with 10 and 14 years of clinical experience: all radiologists were blinded to the clinical status of the patients. The lung parenchyma segmentation was performed through a software-based evaluation on a dedicated workstation using the open-source 3D Slicer software (Fig. 1). More details can be found in the Supplementary Information.

Fig. 1
figure 1

Quantitative CT analysis of a 74-year-old male COVID-19 patient. a Non-contrast chest CT on admission, showing a characteristic bilateral and subpleural ground glass opacity (GGO). b Well-aerated parenchyma segmented semi-automatically by a 3D slicer; the blue area is the result of the subtraction of the all parenchyma (HU − 1100; − 250) with GGO (− 700; − 250) + consolidation (− 250; + 150). c The GGO area, obtained by semi-automatic segmentation (HU − 700; − 250). d Manual segmentation of consolidation areas (HU − 250; + 150)

Full size image

Statistical analysis

A sample size was computed to ensure a predictive ability of the RF model close to 0.8 with a margin of error in the sample estimates d = 0.05. More details can be found in the Supplementary Information. Descriptive statistics were reported as median and interquartile range for continuous variables and percentages (absolute numbers) for categorical variables. Missing values were handled leaving null the estimate. The logistic regression model, odds ratio (OR) together with the 95% confidence intervals (95% CI), and p values were reported for each predictor, considering separately their association with intubation and intra-hospital mortality. The analyses were performed using R software (version 0.2) with the packages caret and rms.

Random forest predictive tool

Model

To identify predictors of intra-hospital intubation or mortality, a random forest (RF) algorithm was employed. The variables having less than 10% of missing values were included in the predictive tool. More details can be found in the Supplementary Information.

Results

Patient characteristics

The main characteristics of the 187 patients included in the study are summarized in Table 1. One hundred forty patients (∼75%) were males, with a median age of 64 years, a median BMI of 28 kg/m2, and a median PaO2/FiO2 of 258 mmHg on admission. The median symptom onset was 7 days prior to admission. The most frequent comorbidities were hypertension (51%) and diabetes (24%). Main laboratory and chest CT findings on admission are also reported in Table 1. Forty-five patients (24%) received standard oxygen therapy, 86 patients (46%) received noninvasive ventilation, and 56 patients (30%) were admitted to ICU and required IMV.

Table 1 Patient characteristics and clinical, laboratory, and CT findings relative to alive and death patients
Full size table

Risk factors

The patients’ demographic, clinical, and laboratory characteristics along with chest CT analysis findings relative to both the whole study population and alive vs death subjects are listed in Table 1. When stratifying patients according to mortality, variables independently associated with mortality were age, other cardiovascular diseases except for coronary artery disease, C-reactive protein, lactate dehydrogenase levels, and IMV. Furthermore, on quantitative chest CT examination, we found a significant positive association between GGO and consolidation and fibrosis. Conversely, female sex, PaO2/FiO2, SpO2, lymphocytes count, standard oxygen therapy, and evidence of well-aerated lung parenchyma on chest CT scan were all inversely associated with mortality.

Table 2 enlists the patients’ demographic, clinical, and laboratory characteristics along with the chest CT analysis findings relative to both the whole population and intubated vs non-intubated patients. Among our cohort of 187 patients, 29 patients were excluded from this analysis because classified as “do-not-intubate” subjects, i.e., patients deemed ineligible for intubation in case of CPAP or BiPAP failure. Variables independently associated with intubation were C-reactive protein, lactate dehydrogenase levels, and white blood cell count. On quantitative chest CT analysis, GGO, consolidation, and fibrosis resulted positively associated with intubation, while PaO2/FiO2, SpO2, lymphocyte count, and well-aerated parenchyma were inversely associated with intubation. Lastly, for each logistic regression and for the continuous variables the cut points maximizing the best predictive value along with their corresponding area under the curve (AUC) are shown in Tables 1 and 2.

Table 2 Patient characteristics and clinical, laboratory, and CT findings relative to intubated and non-intubated patients
Full size table

RF algorithm

We next evaluated the importance of the variables encompassed in the RF algorithms for intra-hospital mortality and intubation prediction. In the RF model for mortality prediction, the most important nodes (importance > 50) were C-reactive protein, age, PaO2/FiO2, glycemia, SpO2 on admission, a well-aerated parenchyma, lactate dehydrogenase, GGO, lymphocytes, other consolidation/fibrosis, and symptom onset (Fig. 2A). In the RF tree model for intubation prediction, the most important nodes (importance > 50) were PaO2/FiO2, a well-aerated parenchyma, C-reactive protein, GGO, other consolidation/fibrosis, glycemia, lactate dehydrogenase, lymphocytes, and age (Fig. 2B). The variables having more than 10% of missing values were excluded from the predictive tool (i.e., BMI, fibrinogen, procalcitonin, ferritin, and troponin). The balanced accuracy in predicting intra-hospital mortality was 0.89 (κ value = 0.72; AUC = 0.73) (Fig. 3A), whereas the balanced accuracy in predicting intubation was 0.9 (κ value = 0.75; AUC = 0.74) (Fig. 3B). When the quantitative CT analysis variables were removed from both RF models, the accuracy of the model predicting intra-hospital mortality dropped to 0.75, whereas that of the model predicting intubation fell to 0.69 (Fig. 4A, B).

Fig. 2
figure 2

Important predictor variables (importance > 50) in the random forest (RF) models for intra-hospital mortality (a) and intubation prediction (b)

Full size image
Fig. 3
figure 3

AUC of the RF models. a) The RF model balanced accuracy in predicting intra-hospital mortality is 0.89 (κ = 0.72; AUC = 0.73). b The RF model balanced accuracy in predicting intubation is 0.9 (κ = 0.75; AUC = 0.74)

Full size image
Fig. 4
figure 4

RF model performance without CT parameters. a The RF model balanced accuracy in predicting intra-hospital mortality is 0.75 (κ = 0.1). b The RF model balanced accuracy in predicting intubation is 0.69 (κ = 0.29)

Full size image

Discussion

The main findings of our investigation can be summarized as follows: (i) in our cohort of COVID-19 patients, elderly male subjects with comorbidities, such as other cardiovascular diseases, intubated for severe hypoxemic ARF associated with a major inflammatory response and a widespread pulmonary involvement on chest CT were at increased risk of intra-hospital mortality; (ii) excluding those patients classified as “do-not-intubate,” subjects with severe hypoxemic ARF experiencing increased inflammatory response and poor lung aeration on chest CT were at increased risk of intubation; and (iii) our novel RF algorithms performed well in predicting intra-hospital mortality and intubation in our study population.

The intra-hospital mortality rate of critically ill patients admitted for COVID-19 has been reported to range from 17 to 67% [18]. Well recognized risk factors associated with low survival or poor outcome in ICU are: male gender, increasing age, comorbidities such as diabetes, hypercholesterolemia, chronic obstructive pulmonary disease, IMV at high positive end-expiratory pressure, low PaO2/FiO2 on ICU admission, high SOFA score, acute kidney injury, reduced respiratory system compliance, late pulmonary infections, and cardiovascular complications [14, 19].

In our setting, intra-hospital mortality rate was 31%, in line with previous reports [18]. Risk factors for intra-hospital mortality identified in our cohort confirmed all previously reported predictors [10,28]. Also, the symptom onset duration was confirmed to be a poor outcome predictor, being a fever lasting more than 7 days from onset of illness associated with increased ICU admission [29].

Although our study confirms with an innovative approach (i.e., RF) the risk factors of intubation and mortality found in the recent literature, it has several limitations. First, due the retrospective nature of the present single-center study, our results lack of generalizability. Second, no power sample was estimated for RF model accuracy assessment. Thus, no definitive conclusions can be drawn on the precision of our algorithm for intubation and intra-hospital mortality prediction. Third, some variables were excluded during the algorithm construction due to missing data occurrence > 10% and the fact that other laboratory values, such as PaCO2, creatinine, and D-Dimer were not collected. Lastly, our prediction models were not validated before the present investigation. Therefore, a future prospective investigation addressing the validation of our models is clearly needed.

Conclusions

In our cohort of COVID-19 patients, we confirmed all demographic, clinical, hemato-chemical, and radiologic predictors of adverse outcomes previously reported. In addition, our innovative RF models based on the risk factors identified in our subset showed a good level of accuracy in predicting intra-hospital mortality and intubation. Thus, this approach may help to accurately identify patients at higher risk of deterioration, which would be particularly important in case of disproportion between resources and demand. Our results await further confirmation in larger multicentric prospective studies.