Association of Initial Symptoms or Comorbidities With Pneumonia Lesions in COVID-19 Patients: Based on Articial Intelligence-Enabled CT Quantitation

Background: Coronavirus disease 2019 (COVID-19) patients with a larger ratio of pneumonia lesions are more likely to progress to acute respiratory distress syndrome and death. This study aimed to investigate the relationship of baseline parameters with pneumonia lesions on admission, as quantied by an articial intelligence (AI) algorithm using computed tomography (CT) images. Methods: This retrospective study quantitatively assessed lung lesions on CT using an AI algorithm in 1630 consecutive patients conrmed with COVID-19 on admission and classied the patients into none (0%), mild (>0–25%), intermediate (>25–50%), and severe (>50%) groups, according to the lesion ratio of the whole lung. A multivariate linear regression model was established to explore the relationship between the lesion ratio and laboratory parameters. The baseline parameters associated with lung lesions, including demographics, initial symptoms, and comorbidities, were determined using a multivariate ordinal regression model. Results: The 1630 patients conrmed with COVID-19 had a median whole lung lesion ratio of 4.1%, and the right lower lung lobe had the most lesions among the ve lung lobes based on the evaluation of CT using AI algorithm. The whole lung lesion ratio was associated with the levels of plasma brinogen (r=0.280, p<0.001), plasma D-dimer (r=0.248, p<0.001), serum α-hydroxybutyrate dehydrogenase (r=0.363, p<0.001), serum albumin (r=-0.300, p<0.001), and peripheral blood leukocyte count (r=0.194, p<0.001). Among the four patients groups categorised by whole lung lesion ratio, the highest frequency of cough (p<0.001) and shortness of breath (p<0.001) were found in the severe group, and the highest frequency of hypertension (p<0.001), diabetes (p<0.001) and anemia (p=0.039) were observed in the intermediate group. Based on baseline ordinal regression analysis, cough (p=0.009), shortness of breath (p<0.001), hypertension (p=0.002), diabetes (p=0.005), and anemia (p=0.006) were independent risk factors for more severe lung lesions. Conclusions: Based on AI-enabled CT quantitation, patients with initial symptoms of cough/shortness of breath, or with comorbidities of hypertension, diabetes, or anemia, had a higher risk for more severe lung lesions on admission in COVID-19 patients.


Introduction
Since the outbreak of Coronavirus disease 2019  in December 2019, more than 27 million con rmed cases and 894,241 deaths were reported until 9 September 2020. The trend of the world pandemic shows no sign of reversing and heavily burdens the health system all around the world. In contrast to mild symptoms caused by common cold coronaviruses, severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection leads to lethal pneumonia, and acute respiratory distress syndrome (ARDS) is the primary cause of death in COVID-19 (1). As the lung was the main organ attacked by SARS-CoV-2, the COVID-19 patients with a larger proportion of lung lesions on computed tomography (CT) in early days had a higher likelihood of developing ARDS, admitted to the intensive care unit (ICU), and progress to death (2)(3)(4)(5). However, the baseline determinants of early COVID-19 pneumonia lesions have not been illustrated yet.
SARS-CoV-2 invades target cells by binding to angiotensin-converting enzyme 2, which is highly expressed in the lung and widely distributed in the heart, kidney, gastrointestinal tract, and even active in all organs (6,7). Accordingly, the COVID-19 symptom pro les vary, including respiratory and gastrointestinal symptoms, such as cough, expectoration, shortness of breath, diarrhea, nausea, and other symptoms, namely, headache, chest distress, and myalgia (8,9). With the exception of asymptomatic cases, COVID-19 patients suffered from diverse symptoms in the early stage of COVID-19, although efforts have been devoted to the management of COVID-19 symptoms, none of the research has focussed on the association of initial symptoms with COVID-19 pneumonia lesions (10,11). Previous studies reported that in COVID-19 patients, comorbidities were associated with a higher rate of ICU admission, invasive ventilation, and death, however, the association of comorbidities with early COVID-19 pneumonia is not clear and worthy of investigation (12).
Emerging arti cial intelligence (AI) technology has already been employed in the evaluation of CT images, with deep learning, the AI owned human-level performance in the classi cation of brotic lung disease (13). In the COVID-19 pandemic, AI assessment of lung CT images has been widely used in the diagnosis of COVID-19, evaluating drug e cacy and assessing the prognosis (2). AI performed with a similar accuracy in the diagnosis of COVID-19 compared to the experienced radiologist and provided an objective and instant way to quantitively evaluate the pneumonia lesion on CT, however, the traditional way of evaluating pneumonia lesion on CT was based on semi-quantitative estimation by the radiologist (2,14,15). Thus, we take advantage of AI algorithm to quantify pneumonia lesions on CT to explore the potential baseline determinants associated with the severity of lung lesions in COVID-19 patients on admission.

Study participants
A total of 1630 adult patients (≥ 18 years old) diagnosed with COVID-19 in Huo Shen Shan Hospital, Wuhan, China, from 4 February 2020 to 9 March 2020 and underwent chest CT scans with AI quantitative assessment of lung lesions on admission were included in the retrospective study. This study was approved by the Human Ethics Committee of Huo Shen Shan Hospital (No. HSSLL023), with a waiver of informed consent.

Data collection
Standardised clinical electronic medical records were investigated by the research team, and all the extracted data related to patients' demographic, initial symptoms, comorbidities, laboratory tests, and chest CT images were crosschecked by two experienced doctors. Demographic data, such as age, sex, smoking status, and drinking status were collected. Initial symptoms and comorbidities were obtained according to the patients' self-reports. The values of laboratory parameters and the proportion and volume of lung lesions based on chest CT images were recorded according to the examination results at admission.

Laboratory testing
In order to con rm the SARS-CoV-2 infection, real-time reverse transcriptase-polymerase chain reaction was used to detect SARS-CoV-2 from throat swab specimens. Laboratory tests included complete blood count, coagulation pro le, biochemical pro le (including renal and liver function, lactate dehydrogenase), cardiac function test, procalcitonin, and high-sensitivity C-reactive protein (hsCRP). Laboratory test results at admission were selected.

CT acquisition and AI-based quantitation of lung lesion
The COVID-19 con rmed patients holding their breath at the end of inspiration in the supine position underwent unenhanced CT scans using multi-detector CT (uCT 760, United Imaging, Shanghai, China) at admission. The parameters adopted in the CT scans were as follows: tube voltage, 120 kV; tube currenttime product, 145 mAs; collimation, 0.625 mm; reconstructed slice thickness, 10 mm; matrix size, 512×512.
Based on CT scans, the lung lesions in COVID-19 patients were quanti ed using AI technology (DAMO Academy Medical AI, Alibaba, China). The reported parameters of the lung lesions included the lesion volume (mL) and the lesion ratio (%), calculated as the ratio of the lesion volume to the total volume of the corresponding lung or lobes. The severity of lung lesions was determined by the level of whole lung lesion ratio, and the patients were categorised into none (0%), mild (>0-25%), intermediate (>25-50%), and severe (>50%) groups (2).

Statistical analysis
According to the severity of lung lesions, the patients were categorised into none, mild, intermediate, and severe groups. Categorical variables were presented as the number of patients and percentages and were compared using χ 2 tests or Fisher's exact tests, as appropriate. According to the results of the Kolmogorov-Smirnov tests, the continuous variables with skewed distribution are presented as median with interquartile range (IQR) and the differences between the four groups were compared using the Kruskal-Wallis H test. The association between the lesion ratio of the whole lung and laboratory ndings was checked by univariate linear regression analysis, and the variables associated with the lesion ratio of the whole lung (p<0.05) were included in a stepwise multivariate linear regression model. The baseline determinants (including demographics, initial symptoms, and comorbidities) of lung lesion severity on admission were investigated using univariate ordinal logistic regression analysis. A test of parallel lines was performed to determine the feasibility of ordinary regression. Variables with p-values <0.05 and pvalues of parallel line test >0.05 were included in a multivariate ordinal logistic model. A two-sided α-value less than 0.05 was considered statistically signi cant. Statistical analyses were performed using the SPSS software (Version 22.0, IBM Corp., Armonk, NY, USA).

Quantitative assessment of lung lesions
The AI algorithm was employed to quantitatively evaluate the lung lesion on CT, and the distribution of lung lesions in COVID-19 patients is shown in Table 1. The median lesion ratio and median lesion volume of the whole lung were 4.1% (IQR 0.5-12.9) and 153.9 ml (IQR, 20.8-426.2), respectively. A larger proportion and volume of lung lesions were observed in the right lung, rather than in the left lung. Of the total ve lung lobes, the highest proportion and volume of lung lesions were found in the right lower lobe. The median ratio and the median volume of lung lesions in the right lower lobe were 5.9% (IQR 0.3-22.8%) and 48.5 ml (IQR 2.7-149.6), respectively.  (2). The demographics, initial symptoms, and comorbidities of the patients are presented in Table 2. Of all 1630 cases, the median age was 62 years (IQR 52-69), and 839 patients (51.5%) were male. The proportions of smokers and drinkers were 5.6% and 3.2%, respectively. No differences were found in sex and smoking status among the four groups. The highest median age was observed in the intermediate group.
With the gradually worsening of lung lesions in the four groups, the frequency of fever, cough, shortness of breath, fatigue, and chest distress increased accordingly, and the highest proportion of myalgia was found in the intermediate group.
Discrepancies were not observed in the other initial symptoms. The highest proportions of hypertension, diabetes, anemia, and stroke were found in the intermediate group.
Other comorbidities among the four groups showed no signi cant differences. Laboratory ndings on admission in COVID-19 patients with different levels of lung lesion ratios were studied (Table 3). With the augment of lung lesion ratio in the four groups, the levels of plasma brinogen, plasma D-dimer, serum lactate dehydrogenase, serum α-hydroxybutyrate dehydrogenase, serum high-sensitive cardiac troponin I, serum myoglobin, serum B-type natriuretic peptide, serum procalcitonin, serum hsCRP, peripheral blood leukocyte count, peripheral blood neutrophil percentage and prothrombin time increased gradually. Accordingly, the levels of peripheral blood hemoglobin, serum total protein, serum albumin, activated partial thromboplastin time, and peripheral blood lymphocyte percentage gradually decreased. The highest levels of serum alanine aminotransferase and serum creatinine kinase-myocardial band were observed in the severe group and the highest levels of serum aspartate transaminase, serum γ-glutamyltransferase, serum urea nitrogen, serum cystatin C and thrombin time were found in the intermediate group.   Table S1).  .995) Not selected CI, con dence interval; COPD, Chronic obstructive pulmonary disease; "Not applicable" indicates the variable not applicable to the ordinal logistic regression for the p value of parallel line test < 0.05. The lesion ratio of the right lung, left lung, and the whole lung in patients with or without cough, shortness of breath, hypertension, diabetes, or anemia are shown in Fig. 2. Patients with one of the aforementioned initial symptoms or comorbidities had higher lesion ratios in the whole lung, right lung, and right lung, compared to those in the patients without these initial symptoms or comorbidities.

Discussion
To the best of our knowledge, this is the rst study to investigate the association of demographics, initial symptoms, and comorbidities with COVID-19 pneumonia lesions quantitatively evaluated by AI algorithm on CT when admitted in a relatively large population of 1630 patients in Huo Shen Shan Hospital. We found that in COVID-19 patients, the right lung had a larger percentage of lesions compared to that in the left lung, and the right lower lobe had the most lesions among the ve lung lobes. In addition, the levels of plasma brinogen, plasma D-dimer, serum albumin, serum α-hydroxybutyrate dehydrogenase, and peripheral blood leukocyte count were independently correlated with the lesion ratio of the whole lung.
The COVID-19 patients with speci c initial symptoms or comorbidities, such as cough, shortness of breath, fatigue, hypertension, diabetes, or anemia were more likely to develop into more severe lung lesions at admission. Similar to Zhang et al.'s study, the right lower lobe was found to have the most lesions among the ve lung lobes. Additionally, we found that the right lung had a larger percentage of lesions compared to that in the left lung. We reported a larger median lesion ratio and median lesion volume in COVID-19 patients on admission, compared to those observed in COVID-19 patients ready for discharge in the study by Zhang et al (16).
The highly correlated laboratory parameters with the lesion ratio of the whole lung in this study consisted of coagulation pro le, serum enzymes, complete blood count, and hsCRP. These laboratory parameters were multiorgan related and confoundedly associated with the lesion ratio of the whole lung. Consistent with He et al.'s nding, the levels of serum hsCRP and serum lactate dehydrogenase were positively correlated with the lesion ratio of the whole lung, and the level of serum albumin was negatively correlated with the lesion ratio of the whole lung (2). However, adjusted by multivariate linear regression model, we found that only the plasma brinogen level, plasma D-dimer level, serum albumin level, serum α-hydroxybutyrate dehydrogenase level, and peripheral blood leukocyte count were independently correlated with the lesion ratio of the whole lung. This observation shares the same perspective with previous researches that reported elevated plasma D-dimer or serum α-hydroxybutyrate dehydrogenase were associated with poor clinical outcomes in COVID-19 patients (17,18). Laboratory ndings and postmortem ndings in COVID-19 patients revealed activation of coagulation pathways and in ammation responses in the COVID-19 (19)(20)(21). However, D-dimer and brinogen levels were independently correlated with the lesion ratio of the whole lung, highlighting the vital role of the coagulation system in lung injury. Procoagulant state overproduced proin ammatory cytokines via proteinase-activated receptors participating in the lung injury, or it was only the consequence of lung lesion initiated by endothelial cell injury requires further validation (22). The higher level of peripheral blood leukocyte count facilitated the overproduction of cytokines, which promoted the formation of cytokine storm and led to worse severe lung lesions. The deterioration of serum albumin and α-hydroxybutyrate dehydrogenase concomitantly with the severity of the whole lung lesion ratio con rmed the involvement of multiorgan injury in SARS-CoV-2 infection. However, multiple organs that were injured by the invasion of SARS-CoV-2 synchronously or secondary to respiratory failure needs further investigation.
Initial symptoms of fever, cough, shortness of breath, and fatigue were risk factors of more severe COVID-19 pneumonia lesions on admission. It is believed that vascular endothelial cells play a central role in the pathophysiological process of COVID-19 (23). In severe COVID-19 infection, cytokine storms are always induced by the early response of proin ammatory cytokines and injury of the vascular endothelial cells, especially in the lung. The formation of cytokine storms in early stage prompt fever. The invasion of SARS-CoV-2 stimulated cough as a defensive re ex. Patients with coughs are more likely to have infections in the respiratory tract and have more lesions. Shortness of breath as a sign of hypoxia, indicating the injury of the cardiopulmonary system and is often present in severe pneumonia. The higher lung-lesion ratio was shown to be associated with a higher respiratory rate in COVID-19 patients (2). We also found that the patients with initial symptoms of shortness of breath had a larger ratio of lung lesions on admission. In agreement with a previous study, we found that fatigue was associated with poor outcomes. However, a previous nding in the poor prognosis of expectoration was not found in our study (11,24). Different from those studies, our research focussed on the association of expectoration with the severity of lung lesions and carried out a larger study population.
Compared to patients who died with H1N1, SARS-CoV-2 infection prompted endotheliitis and induced the apoptosis of vascular endothelial cells, and 9 times the number of alveolar capillary microthrombi that were found in COVID-19 victims (19). It is believed that the injury of vascular endothelial cells, particularly in the lung, promotes in ammation, induces tissue oedema, and leads to ARDS (25). Thus, hypertension and diabetes with pre-existing injured endothelial cells were associated with poor clinical outcomes in COVID-19 patients (12,25,26). In our study, we found that diabetes and hypertension, other than chronic obstructive pulmonary disease (COPD), were associated with the severity of lung lesions, which suggests that COVID-19 patients with hypertension and/or diabetes might have more severe pneumonia and thus predicts worse clinical outcomes. A recent study reported that initial anemia was correlated with longer hospital stay and higher mortality in COVID-19 patients, and patients with anemia of in ammation (68.8% of all kinds of anemia) had an eight-fold higher likelihood of developing mechanical ventilation (27). Anemia of in ammation was the most common anemia in hospitalized patients (28). A higher level of in ammation in patients with anemia of in ammation may participate in the injury of endothelial cells in the lung. In addition, COVID-19 patients with anemia were more likely to suffer from hypoxia, which might increase cytokines and lead to the apoptosis of endothelial cells, and further progress to more severe pneumonia (29).
Several limitations of this study should be noted. First, this was a retrospective study and was subject to selection bias. Second, ndings in this single-centre study need a multicentre prospective study for validation. Finally, the documented initial symptoms were self-reported by patients on admission, which inevitably involves recall bias.

Conclusions
Based on AI quantitative assessment on CT, the initial symptoms of fever, cough, shortness of breath, fatigue, and comorbidities, such as hypertension, diabetes, and anemia, independently predict a worse severity of lung lesions on admission in COVID-19 patients. This nding provides early-warning indicators of more severe COVID-19 pneumonia and might inform and improve the individualised treatments of COVID-19 patients.