Background
Since the emergence of the first case of coronavirus disease (COVID-19) in 2019, more than 676 million cases have been reported worldwide [
1,
2]. Significant efforts have been made to develop vaccines and treatments, such as systemic corticosteroids and remdesivir [
3,
4], to improve disease prognosis. However, there have been recurring outbreaks in Japan [
5] and may acquire stronger virulence in the future, forcing us to prepare for the next pandemic. Because there are heterogeneous disease courses of COVID-19, early identification of patients at risk of severe disease is important for the appropriate use of medical resources.
Chest computed tomography (CT) is extensively used to diagnose COVID-19 and to predict its severity [
6,
7]. In the early years of the COVID-19 pandemic, qualitative assessments of pneumonia by radiologists were useful for predicting disease severity [
8‐
10]; however, the reproducibility of evaluation by radiologists is problematic [
11]. In fact, semi-quantitative analysis of pneumonia and well-aerated lung volumes using CT density has shown a better prediction of severity than qualitative assessments by radiologists [
12,
13]. Furthermore, artificial intelligence (AI) -based CT analyses have recently become useful in diagnosing and predicting the severity of COVID-19 [
14‐
16]. The advantage of AI-based analysis is that it can quickly, easily, and reproducibly quantify pneumonia without intra- or interobserver variability. However, reports on the usefulness of AI-based CT analysis in predicting COVID-19 severity are limited to a small number of cases at a single center over a short duration [
15,
17]. Only a few studies have compared the detailed clinical characteristics or complications during hospitalization of patients based on the volume of pneumonia quantified using AI-based analyses [
6].
Since the latter half of 2020, various systemic symptoms have been reported to persist after the acute phase of COVID-19 [
18] and to cause long-term lung sequelae [
19]. The pathogenesis of the sequelae is not well understood, and the analysis of structural changes using CT may be important for understanding them. Qualitative CT evaluation by radiologists has shown that lung sequelae are frequently accompanied by residual shadows three months after COVID-19 [
20]; however, only a few studies have performed AI-based CT quantification of residual lesions [
21,
22]. Moreover, no study has examined the clinical characteristics of patients with residual lesions using AI-based CT quantification. As there is no established management strategy to improve lung sequelae, further understanding of the underlying structural changes in relation to the clinical features is warranted.
It was hypothesized that AI-based CT quantification of pneumonia would not only be useful for predicting outcomes in the acute phase but also for the objective assessment of lung sequelae. This study aimed to investigate (1) the usefulness of AI-based CT quantification of COVID-19 pneumonia in predicting critical outcomes using a multicenter database with long-term duration, and (2) the longitudinal change in AI-based CT quantification of residual lesions and the clinical characteristics of patients with persistent pneumonia after the acute phase of infection.
Discussion
This is the first study to investigate the usefulness of AI-based CT quantification of COVID-19 pneumonia for predicting critical outcomes and complications of COVID-19 and to evaluate the longitudinal change in quantification using a large retrospective cohort database. Pneumonia cases with a high percentage of lung lesion classified using AI-based CT quantification was strongly associated with critical outcomes adjusted for other known predictors, including the presence of pneumonia in qualitative assessments and other complications during hospitalization. In addition, longitudinal follow-up of the quantification revealed a high percentage of residual lesions at three months in more cases with critical outcomes. These results show that AI-based CT quantification is a valuable tool that can sufficiently predict the severity of COVID-19 pneumonia and that this tool highlights populations associated with COVID-19 sequelae by assessing persistent pneumonia.
The strength of this study is the use of a novel AI-based CT quantification of pneumonia in a large, multicenter, and long-term cohort. Many previous studies have shown that AI-based CT quantification can predict disease severity in a small number of patients [
15,
17]. The large sample size in this study allowed for multivariable analysis and showed that AI-based CT quantification predicts critical outcomes independently of many prognostic factors. In routine clinical practice, especially during a pandemic, it is difficult to perform CT imaging under the same conditions. Although this study included multiple CT scanner models and imaging conditions, it was clinically significant in predicting disease severity. Multiple waves of the COVID-19 pandemic have been confirmed worldwide, including in Japan [
30]. Because the characteristics of the viral strain and various other factors, such as the development of therapeutic drugs and vaccines, are involved, the clinical characteristics of patients differ depending on the epidemic period. In this study, we investigated the stratification of epidemic waves and observed that AI-based CT quantification was useful for predicting the severity of COVID-19 regardless of the epidemic waves. This suggests a clinically significant application in the event of future epidemics.
The advantage of AI-based CT analysis is that it enables a more reproducible, quick, and quantitative evaluation of pneumonia than qualitative evaluation by radiologists, and achieves faster analysis than semi-quantitative analysis. The accuracy and promptness of AI-based CT quantification are useful tools for identifying patients with early exacerbation of COVID-19 in clinical practice, although there are several infected patients during the pandemic. In this study, AI-based CT quantification of pneumonia volume was predictive of various clinical outcomes independent of the clinician’s visual assessment of pneumonia. This finding extends previous studies showing that AI-based CT quantification analysis is superior to density-based qualitative analysis for predicting disease severity [
31,
32]. Previous reports about pneumonia quantification indicated that the thresholds of % volume of pneumonia were 6.0-8.2% for requiring oxygen support, 23% for requiring IMV support, and 36–40% for mortality, respectively [
12,
33‐
36]. Therefore, we considered the threshold of 16.0% in this study to be an appropriate cut-off value for predicting critical outcomes. Although a large-scale study has proposed severity prediction models using lung CT analysis, multiple clinical data, and radiomics analysis [
6,
37], we believe that our simple AI-based CT examination of pneumonia volume (pneumonia with % lung lesions on CT images) can provide prognostic information with similar accuracy and should be more clinically relevant in terms of routine clinical applications.
The pneumonia group with a high % lung lesions showed a higher incidence of multiorgan complications in this study. This is consistent with previous findings of poor outcomes after hospitalization, including bacterial infection [
38], renal failure [
39], heart failure [
40], thrombosis [
41], and liver failure [
42]. In this context, our data are important as they suggest that AI-based measurement of pneumonia volume can predict complications during hospitalization and allow for early intervention in selected high-risk patients.
Factors related to COVID-19 severity have been proposed to account for the diversity in the natural course of COVID-19 [
25,
28,
29]. These previous findings are consistent with our findings that poor prognostic factors, including BMI, diabetes, and CKD, were more frequently observed in patients with a high percentage of pneumonia. Moreover, our data reaffirm previous findings that inflammatory biomarkers, including CRP, procalcitonin, and ferritin, are associated with severe disease [
43] and pneumonia [
17,
44] because COVID-19 causes a cytokine storm and systemic inflammation with severe disease [
45].
COVID-19 causes persistent sequelae, referred to as Long-COVID [
46]. A review of CT qualitative evaluations by radiologists showed that lung lesions remained in 70–94% of patients with COVID-19 at 3 months [
20,
47‐
49]. Furthermore, patients with COVID-19 who had residual lung lesions 3 months after COVID-19 onset had poor pulmonary function and severely decreased oxygen saturation during the 6-minute walk test [
48]. However, no study has examined pulmonary sequelae using an AI-based CT quantification analysis. In this study, despite the overall improvement in pneumonia volume 3 months after onset, many patients showed residual lesions and more residual lesions were associated with older age, female sex, history of hypertension, and higher severity of COVID-19. These findings are consistent with those of previous reports (a review of radiologists’ qualitative assessments), in which cases of observed residual lesions were associated with older age [
48], length of hospitalization, and the need for IMV support [
50]. In the present study, we predicted the % volume of residual lesions using the % volume of lung lesions on CT images at admission. AI-based CT quantification analysis at the time of onset is more useful than qualitative analysis for predicting the severity of COVID-19 [
31,
32] and may also prevail over the long-term disease course. Longitudinal AI-based CT quantification analysis may be a predictor of long-COVID symptoms, and future studies are warranted to clarify its association with these symptoms.
This study has some limitations. First, the number of cases evaluated 3 months after onset was small owing to drop out. This study used follow-up CT scans obtained 3 months after onset based on previous reports showing that 3 months should be reasonable to evaluate residual lesions, [
51] and show a correlation between residual lesions at 3 months and physiological sequelae, including dyspnea [
47,
48]. In our study, the severity of the clinical condition on admission was greater in the patients who underwent follow-up CT than in those who did not. This may be a cause of selection bias that the severe patients were more likely to be included in the analysis after 3 months. Second, we did not evaluate the extrapulmonary organs using CT in this study. Several organ abnormalities outside the lungs, including muscle, fat, and coronary artery calcification, are associated with COVID-19 [
52‐
54]. Combining these analyses of extrapulmonary organs with pneumonia volume may lead to improved accuracy in predicting disease severity, and future studies are warranted.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.