Predicting Clinical Outcomes of SARS-CoV-2 Infection During the Omicron Wave Using Machine Learning (preprint)

The Omicron SARS-CoV-2 variant continues to strain healthcare systems. Developing tools that facilitate the identification of patients at highest risk of adverse outcomes is a priority. The study objectives are to develop population-scale predictive models that: 1) identify predictors of adverse outcomes with Omicron surge SARS-CoV-2 infections, and 2) predict the impact of prioritized vaccination of high-risk groups for said outcome. We prepared a retrospective longitudinal observational study of a national cohort of 192,984 patients in the U.S. Veteran Health Administration who tested positive for SARS-CoV-2 from January 15 to August 15, 2022. We utilized sociodemographic characteristics, comorbidities, vaccination status, and prior COVID-19 infections, at time of testing positive for SARS-CoV-2 to predict hospitalization, escalation of care (high-flow oxygen, mechanical ventilation, vasopressor use, dialysis, or extracorporeal membrane oxygenation), and death within 30 days. Machine learning models demonstrated that advanced age, high comorbidity burden, lower body mass index, unvaccinated status, prior SARS-CoV-2 infection, and oral anticoagulant use were the important predictors of hospitalization and escalation of care. Similar factors predicted death. However, prior SARS-CoV-2 infection was associated with lower 30-day mortality, and anticoagulant use did not predict mortality risk. The all-cause death model showed the highest discrimination (Area Under the Curve (AUC) = 0.895, 95% Confidence Interval (CI): 0.885, 0.906) followed by hospitalization (AUC = 0.829, CI: 0.825, 0.834), then escalation of care (AUC=0.805, CI: 0.795, 0.814). Assuming a vaccine efficacy range of 70.8 to 78.7%, our simulations projected that targeted prevention in the highest risk group may have reduced 30-day hospitalization, care escalation, and death in more than 2 of 5 unvaccinated patients.

Are South Korean College Students Benefitting from Digital Learning?

With COVID-19, the advancement of mobile devices (e.g., smartphones, laptops, tablets) has brought a welcoming change to education: digital learning. This study addresses the relationship between mobile device use and academic performance through three different models by controlling demographic data, technological infrastructure conditions, and daily total multi-tasking time. The first model emphasized the daily total mobile device use time. The second model divided the daily total mobile device use time into academic and non-academic oriented uses. The final model divided the overall mobile device use into seven specific usage types. The study found that an increase in the daily total mobile device use time negatively affected GPA;only non-academic purpose use time had a negative significance toward GPA;none of the seven usage types were significant in predicting GPA. Based on the findings, suggestions on improvements for the future digital learning policy were proposed.

Moral foundations, ideological divide, and public engagement with u.S. Government agencies' covid-19 vaccine communication on social media

Guided by moral foundation theory, this study examined how moral framing interacted with local constituents' ideological leaning to affect public engagement outcomes of government agencies' COVID-19 vaccine communication on Facebook. We analyzed a dataset of over 5,000 U.S. government agencies' Facebook posts on COVID-19 vaccines in 2021 (N = 70,671), assessed their use of moral language using a newly developed computational method, and examined how political divide manifests itself at the collective level. Findings from both fixed and random effects models suggest that: 1) the use of moral language is positively associated with public engagement outcomes on government agencies' social media accounts;2) five types of moral foundations have distinct effects on three types of public engagement (affective, cognitive, and retransmission);3) moral foundations and local politics interact to affect public engagement, in that followers of government agencies in liberal states/counties prefer messages emphasizing the care/harm and fairness/cheating dimensions while those in conservative states/counties prefer the loyalty/betrayal dimension. The study demonstrates how a strategic employment of moral language can contribute to public engagement of government agencies' mass communication campaigns. (PsycInfo Database Record (c) 2022 APA, all rights reserved)

Nursing Intervention Analysis in COVID-19 Negative Pressure Isolation Wards and General Wards: Observational study (preprint)

Introduction: With the global spread of Coronavirus disease (COVID-19) and public health crisis, appropriate allocation of healthcare human resources has been necessitated. Although nursing practice takes up a larger part of medical practice in hospitals, the quantitative assessment of nursing care has not been investigated for human resource allocation in the medical field. The objective of this study to explore the time spent for each nursing intervention, and compared provided amount of nursing intervention between negative pressure isolation wards (NPIWs) and general wards (GWs) provided by COVID-19 hub hospitals. Methods: This research is a time-motion (TM) observational study. Three trained external observers recorded their observations for every minute in 19 different work schedules in 2 NPIWs and 2 general respiratory wards. Observation items were chosen based on the standard operating guidelines of Integrated Nursing and Caring Services developed by the Ministry of Health and Welfare and National Health Insurance Service. The average nursing workload per shift was compared by calculating the sum of the spent time of three nurses staffed in each shift in each ward between two groups. In addition, to compare the amount of directed nursing care for patients between two types of wards, nursing work category was divided into directed and undirected nursing interventions. Results: In the comparison of demographic characteristics of nursing workforce between two groups, there was no statistically significant difference (p>0.05 respectively). In both groups, the most time-consuming nursing work category was recording in three work shifts. The average duration of those work tasks was 312.5 minutes in NPIWs and 307 minutes (per 3 nurses) in GWs, having no significant difference (p>0.05). Of all nurse duties, the second most time-consuming work category was others (including changing to protective clothing) in NPIWs, and medication administration and transfusion in GWs. The mean duration of performing the category for others that include wearing PPE was 308 minutes in NPIWs and 160 minutes (per 3 nurses) in GWs, showing a significant difference (p<0.05). The greater amount of time was taken for hygiene management in isolation wards. Medication administration and transfusion and nursing assessment were more frequently performed in GWs, demonstrating a statistical significance. In the aggregated spent time for all duties including directed and undirected nursing care, the time spent for directed nursing care was 654 minutes longer in GWs than in NPIWs (per 3 nurses) in each work shift, displaying a significant difference. Conclusion: This study provides the quantitative difference in time-consuming nursing works between NPIWs and GWs by direct observation. Recording was the most time-consuming nursing work category in both NPIWs and GWs. Considering nurses in each duty in GWs provided more directed nursing care than nurse in NIPWs, careful considerations are required in allocation of nursing workforce.

Incidence of febrile seizures in children with COVID-19 (preprint)

Purpose This study aimed to investigate the incidence of febrile seizures (FS) associated with coronavirus disease 2019 (COVID-19) in children and examine the variation in these incidences after the Omicron surge. Methods The number of confirmed COVID-19 cases aged below 5 years residing in the Jeonbuk province from January 2020 to June 2022 was obtained from official data released by the Korean government. During the same period, data regarding FS patients with COVID-19 were obtained from all local hospitals capable of FS treatment in Jeonbuk. The data were analyzed retrospectively. Results The number of children under 5 years of age in Jeonbuk was 62,772, of which 33,457 (53.2%) were diagnosed with COVID-19 during the study period. Of these, 476 patients (1.4%) required hospitalization and 64 (0.19%, 44 boys: 20 girls) developed FS. Until 2021, before the Omicron surge, 23.4% of the patients (89 of 381) required hospitalization, but no children with COVID-19 were hospitalized for FS. However, after the Omicron surge in 2022, 16.5% of hospitalized children (64 of 387) experienced FS, despite the decline in hospitalization rates among COVID-19 patients (1.2%). Twenty-five patients (39.1%) had complex FS, and one (1.6%) presented with febrile status epilepticus. Forty-two patients (65.6%) experienced first-time FS, with an average of 1.5 convulsive events. Conclusions During the COVID-19 pandemic, the incidence of FS was approximately 0.19%; however, after the emergence of the Omicron variant, FS occurred more frequently and became more complex.

Sensitive and reproducible detection of SARS-CoV-2 using SERS-based microdroplet sensor.

Surface-enhanced Raman scattering (SERS)-based assays have been recently developed to overcome the low detection sensitivity of severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). SERS-based assays using magnetic beads in microtubes slightly improved the limit of detection (LoD) for SARS-CoV-2. However, the sensitivity and reproducibility of the method are still insufficient for reliable SARS-CoV-2 detection. In this study, we developed a SERS-based microdroplet sensor to dramatically improve the LoD and reproducibility of SARS-CoV-2 detection. Raman signals were measured for SERS nanotags in 140 droplets passing through a laser focal volume fixed at the center of the channel for 15 s. A comparison of the Raman signals of SERS nanotags measured in a microtube with those measured for multiple droplets in the microfluidic channel revealed that the LoD and coefficient of variation significantly improved from 36 to 0.22 PFU/mL and 21.2% to 1.79%, respectively. This improvement resulted from the ensemble average effects because the signals were measured for SERS nanotags in multiple droplets. Moreover, the total assay time decreased from 30 to 10 min. A clinical test was performed on patient samples to evaluate the clinical efficacy of the SERS-based microdroplet sensor. The assay results agreed well with those measured by the reverse transcription-polymerase chain reaction (RT-PCR) method. The proposed SERS-based microdroplet sensor is expected to be used as a new point-of-care diagnostic platform for quick and accurate detection of SARS-CoV-2 in the field.

Outcomes of SARS-CoV-2 infection among children and young people with pre-existing rheumatic and musculoskeletal diseases.

OBJECTIVES: Some adults with rheumatic and musculoskeletal diseases (RMDs) are at increased risk of COVID-19-related death. Excluding post-COVID-19 multisystem inflammatory syndrome of children, children and young people (CYP) are overall less prone to severe COVID-19 and most experience a mild or asymptomatic course. However, it is unknown if CYP with RMDs are more likely to have more severe COVID-19. This analysis aims to describe outcomes among CYP with underlying RMDs with COVID-19. METHODS: Using the European Alliance of Associations for Rheumatology COVID-19 Registry, the Childhood Arthritis and Rheumatology Research Alliance (CARRA) Registry, and the CARRA-sponsored COVID-19 Global Paediatric Rheumatology Database, we obtained data on CYP with RMDs who reported SARS-CoV-2 infection (presumptive or confirmed). Patient characteristics and illness severity were described, and factors associated with COVID-19 hospitalisation were investigated. RESULTS: 607 CYP with RMDs <19 years old from 25 different countries with SARS-CoV-2 infection were included, the majority with juvenile idiopathic arthritis (JIA; n=378; 62%). Forty-three (7%) patients were hospitalised; three of these patients died. Compared with JIA, diagnosis of systemic lupus erythematosus, mixed connective tissue disease, vasculitis, or other RMD (OR 4.3; 95% CI 1.7 to 11) or autoinflammatory syndrome (OR 3.0; 95% CI 1.1 to 8.6) was associated with hospitalisation, as was obesity (OR 4.0; 95% CI 1.3 to 12). CONCLUSIONS: This is the most significant investigation to date of COVID-19 in CYP with RMDs. It is important to note that the majority of CYP were not hospitalised, although those with severe systemic RMDs and obesity were more likely to be hospitalised.

The structure of a novel antibody against the spike protein inhibits Middle East respiratory syndrome coronavirus infections

Middle East respiratory syndrome coronavirus (MERS-CoV) is a zoonotic virus, responsible for outbreaks of a severe respiratory illness in humans with a fatality rate of 30%. Currently, there are no vaccines or United States food and drug administration (FDA)-approved therapeutics for humans. The spike protein displayed on the surface of MERS-CoV functions in the attachment and fusion of virions to host cellular membranes and is the target of the host antibody response. Here, we provide a molecular method for neutralizing MERS-CoV through potent antibody-mediated targeting of the receptor-binding subdomain (RBD) of the spike protein. The structural characterization of the neutralizing antibody (KNIH90-F1) complexed with RBD using X-ray crystallography revealed three critical epitopes (D509, R511, and E513) in the RBD region of the spike protein. Further investigation of MERS-CoV mutants that escaped neutralization by the antibody supported the identification of these epitopes in the RBD region. The neutralizing activity of this antibody is solely provided by these specific molecular structures. This work should contribute to the development of vaccines or therapeutic antibodies for MERS-CoV.

AI can evolve without labels: self-evolving vision transformer for chest X-ray diagnosis through knowledge distillation (preprint)

Although deep learning-based computer-aided diagnosis systems have recently achieved expert-level performance, developing a robust deep learning model requires large, high-quality data with manual annotation, which is expensive to obtain. This situation poses the problem that the chest x-rays collected annually in hospitals cannot be used due to the lack of manual labeling by experts, especially in deprived areas. To address this, here we present a novel deep learning framework that uses knowledge distillation through self-supervised learning and self-training, which shows that the performance of the original model trained with a small number of labels can be gradually improved with more unlabeled data. Experimental results show that the proposed framework maintains impressive robustness against a real-world environment and has general applicability to several diagnostic tasks such as tuberculosis, pneumothorax, and COVID-19. Notably, we demonstrated that our model performs even better than those trained with the same amount of labeled data. The proposed framework has a great potential for medical imaging, where plenty of data is accumulated every year, but ground truth annotations are expensive to obtain.

Host Chitinase 3-like-1 is a Universal Therapeutic Target for the Delta, Omicron and Other SARS-CoV-2 Viral Variants in COVID 19 (preprint)

COVID 19 is the disease caused by severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2; SC2) which has caused a world-wide pandemic with striking morbidity and mortality. Evaluation of early SC2 strains suggested limited viral genetic diversity. However, genetic and epidemiologic investigations in the interim have revealed impressive genetic variability. Many of these viral variants are now defined as variants of concern (VOC) based on genetic alterations in their spike (S) and other proteins that cause enhanced transmissibility, decreased susceptibility to antibody neutralization or therapeutics and or their ability to induce severe disease. The delta {delta} and omicron (o) variants are particularly problematic based on their impressive and unprecedented transmissibility and ability to cause break through infections. The delta variant also accumulates at high concentrations in host tissues and has caused waves of lethal disease. SC2 infection is mediated by S protein binding to cellular ACE2 receptors and subsequent S protein protease processing. Because studies from our laboratory have demonstrated that chitinase 3-like-1 (CHI3L1) stimulates ACE2 and S priming proteases, studies were undertaken to determine if interventions that target CHI3L1 are effective inhibitors of SC2 viral variant infection. Here we demonstrate that CHI3L1 augments epithelial cell infection by pseudoviruses that express the alpha, beta, gamma, delta or omicron S proteins and that the CHI3L1 inhibitors anti-CHI3L1 and kasugamycin inhibit epithelial cell infection by these VOC pseudovirus moieties. Thus, CHI3L1 is a universal, VOC-independent therapeutic target in COVID 19.

Coffee shop corporate social responsibility (CSR) and reuse intention using triple bottom line theory

PurposeThe purpose of this study is to identify the attributes that statistically affect reason intention. The triple bottom line, a theoretical framework of corporate social responsibility (CSR) consisting of economic, social and environmental subdimensions, is used as the theoretical foundation.Design/methodology/approachIn this study, price fairness, quarantine and hygiene, and eco-friendliness represent economic, social and environmental CSR, respectively. Amazon Mechanical Turk is used for data collection. The valid number of observations is 474. Structural equation modeling is implemented to test the research hypotheses.FindingsThe results indicate that price fairness, quarantine and hygiene positively affect the reuse intention of coffee shops. However, eco-friendliness appears to be an attribute that does not significantly affect reuse intention.Originality/valueThis study theoretically contributes to the literature by demonstrating the explanatory power of triple bottom line theory for café customer intention.

Extending the Ring Theory of Personhood to the Care of Dying Patients in Intensive Care Units

It is evident, in the face of the COVID-19 pandemic that has physicians confronting death and dying at unprecedented levels along with growing data suggesting that physicians who care for dying patients face complex emotional, psychological and behavioural effects, that there is a need for their better understanding and the implementation of supportive measures. Taking into account data positing that effects of caring for dying patients may impact a physician’s concept of personhood, or “what makes you, ‘you’”, we adopt Radha Krishna’s Ring Theory of Personhood (RToP) to scrutinise the experiences of physicians working in intensive care units (ICU) using a fictional scenario that was inspired by real events. The impact of death and dying, its catalysts, internal constituents, external factors, dyssynchrony, and buffers, specific to ICU physicians, were identified and explored. Such a framework allows for ramifications to be considered holistically and facilitates the curation of strategies for conflict resolution. This evaluation of the RToP acknowledges the experience and wide-ranging effects it has on ICU physicians. As such, our findings provide insight into their specific needs and highlight the importance of support on a personal and organisational level. Although further research needs to be conducted, the RToP could serve as the basis for a longitudinal assessment tool supported by the use of portfolios or mentorship due to their provision of personalised, appropriate, specific, timely, accessible and long-term support.

Capturing the dynamics of the China crude oil futures: Markov switching, co-movement, and volatility forecasting

The launch of the China's Shanghai International Energy Exchange (INE) oil futures market in 2018 has shed new light on the role of China in international crude oil market. Understanding the dynamics of the newly arrived RMB denominated crude oil futures market not only facilitates the international market participants in risk management and hedging, but also provides helpful information for policy-makers, especially for those from emerging countries, to financialize its energy market, along with the currency internationalization and financial market liberalization. However, the literature on the China crude oil futures is quite scant compared with abundant literature on international benchmarks. In this regard, we make attempts to capture the dynamics of the China crude oil futures by: (1) adopting Markov switching analysis to uncover the regime switching of the INE crude oil futures market;(2) investigating the dynamic connectedness of INE, WTI, and Brent crude oil futures;(3) forecasting the realized volatility of INE crude oil futures. The results have shown that: (1) the outbreak of global pandemic at the beginning of 2020 has switched the crude oil futures market from a stable regime to a volatile regime;(2) the increasing financial uncertainty originated from the world, U.S., other advanced countries and emerging countries could significantly negatively affect the movement of crude oil futures. However, China suffers the least;(3) the dynamic conditional correlations between INE vs WTI, and INE vs Brent are high but lower and more volatile than that of WTI vs Brent;(4) Brent crude oil futures contribute to improving the accuracy of volatility forecasting of INE crude oil futures;and more importantly (5) we highly recommend carrying out volatility forecasting by incorporating intraday realized measures into mixed data sampling approach, in particular, when intraday transactions present extremely different behavior across the time.

TET1 controls Cxcl1 induction by DNA demethylation and promotes neutrophil recruitment during acute lung injury (preprint)

Neutrophils are rapidly recruited from the peripheral blood to the inflammatory site to initiate inflammatory response against pathogenic infections. The process to recruit neutrophils must be properly regulated since the abnormal accumulation of neutrophils can cause organ damage and dysfunction. The acute respiratory distress syndrome (ARDS)/acute lung injury (ALI) is a common cause of respiratory failure that is characterized by the infiltration of neutrophils and epithelial integrity disruption. Indeed, recent studies suggest a pathogenic role of neutrophils in the clinic severity of the coronavirus disease 2019 (COVID-19) ARDS. The chemokine CXCL1, which is rapidly induced by inflammatory stimuli, plays a key role in neutrophil influx during lung inflammation. The molecular basis of Cxcl1 induction is not fully understood. Here we report that TET1, a member of the ten eleven translocation (TET) methylcytosine dioxygenase protein family, displays a striking specificity in the regulation of gene expression in macrophages. RNA sequencing (RNA-seq) analysis showed that Tet1 disruption significantly altered the expression of only 48 genes that include Cxcl1 and several other genes known to be important for cell migration and trafficking in bone marrow derived macrophages (BMDMs) in response to LPS stimulation. TET1 regulates the induction of Cxcl1 by facilitating the DNA demethylation of the Cxcl1 promoter. In Tet1 −/− mice, the induction of Cxcl1 was suppressed, resulting in defective neutrophil recruitment to the lung during LPS-induced acute lung injury. Our results identify a novel epigenetic mechanism that selectively controls Cxcl1 induction and neutrophil recruitment during acute lung injury. Key Points TET1 has a striking specificity in macrophage gene regulation and controls Cxcl1 induction by inflammatory stimuli via DNA demethylation Neutrophil recruitment is defective in Tet1 deficient mice during acute lung injury

Secreted SARS-CoV-2 ORF8 modulates the cytokine expression profile of human macrophages (preprint)

Severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is still adapting to its new human host. Attention has focussed on the viral spike protein, but substantial variation has been seen in the ORF8 gene. Here, we show that SARS-CoV-2 ORF8 protein undergoes signal peptide-mediated processing through the endoplasmic reticulum and is secreted as a glycosylated, disulphide-linked dimer. The secreted protein from the prototype SARS-CoV-2 virus had no major effect on viability of a variety of cell types, or on IFN or NF-{kappa}B signalling. However, it modulated cytokine expression from primary CSF1-derived human macrophages, most notably by decreasing IL-6 and IL-8 secretion. Furthermore, a sequence polymorphism L84S that appeared early in the pandemic associated with the Clade S lineage of virus, showed a markedly different effect, of increasing IL-6 production. We conclude that ORF8 sequence polymorphisms can potentially affect SARS-CoV-2 virulence and should therefore be monitored in sequencing-based surveillance.

IL10RB as a key regulator of COVID-19 host susceptibility and severity (preprint)

Background Recent efforts have identified genetic loci that are associated with coronavirus disease 2019 (COVID-19) infection rates and disease outcome severity. Translating these genetic findings into druggable genes and readily available compounds that reduce COVID-19 host susceptibility is a critical next step. Methods We integrate COVID-19 genetic susceptibility variants, multi-tissue genetically regulated gene expression (GReX) and perturbargen signatures to identify candidate genes and compounds that reverse the predicted gene expression dysregulation associated with COVID-19 susceptibility. The top candidate gene is validated by testing both its GReX and observed blood transcriptome association with COVID-19 severity, as well as by in vitro perturbation to quantify effects on viral load and molecular pathway dysregulation. We validate the in silico drug repositioning analysis by examining whether the top candidate compounds decrease COVID-19 incidence based on epidemiological evidence. Results We identify IL10RB as the top key regulator of COVID-19 host susceptibility. Predicted GReX up-regulation of IL10RB and higher IL10RB expression in COVID-19 patient blood is associated with worse COVID-19 outcomes. In vitro IL10RB overexpression is associated with increased viral load and activation of immune-related molecular pathways. Azathioprine and retinol are prioritized as candidate compounds to reduce the likelihood of testing positive for COVID-19. Conclusions We establish an integrative data-driven approach for gene target prioritization. We identify and validate IL10RB as a suitable molecular target for modulation of COVID-19 host susceptibility. Finally, we provide evidence for a few readily available medications that would warrant further investigation as drug repositioning candidates.

Effect of prone positioning on oxygenation and static respiratory system compliance in COVID-19 ARDS versus non-COVID ARDS (preprint)

Background: Prone positioning is recommended for patients with moderate-to-severe acute respiratory distress syndrome (ARDS) receiving mechanical ventilation. While the debate continues as to whether COVID-19 ARDS is clinically different from non-COVID ARDS, there is little data on whether the physiological effects of prone positioning differ between the two conditions. We aimed to compare the physiological effect of prone positioning between patients with COVID-19 ARDS and those with non-COVID ARDS. Methods: We retrospectively compared 23 patients with COVID-19 ARDS and 145 patients with non-COVID ARDS treated using prone positioning while on mechanical ventilation. Changes in PaO2/FiO2 ratio and static respiratory system compliance (Crs) after the first session of prone positioning were compared between the two groups: first, using all patients with non-COVID ARDS, and second, using subgroups of patients with non-COVID ARDS matched 1:1 with patients with COVID-19 ARDS for baseline PaO2/FiO2 ratio and static Crs. We also evaluated whether the response to the first prone positioning session was associated with the clinical outcome. Results: When compared with the entire group of patients with non-COVID ARDS, patients with COVID-19 ARDS showed more pronounced improvement in the PaO2/FiO2 ratio (adjusted difference 39.3 [95% CI 5.2–73.5] mmHg) and static Crs (adjusted difference 3.4 [95% CI 1.1–5.6] mL/cmH2O). However, these between-group differences were not significant when the matched samples (either PaO2/FiO2-matched or compliance-matched) were analyzed. The improvements in PaO2/FiO2 ratio (subdistribution hazard ratio 1.19, 95% CI 1.08–1.30) and static Crs (subdistribution hazard ratio 1.57, 95% CI 1.29–1.91) after the first prone positioning session were associated with successful discontinuation of mechanical ventilation in patients with COVID-19 ARDS. Conclusions: In patients with COVID-19 ARDS, prone positioning was as effective in improving respiratory physiology as in patients with non-COVID ARDS. Thus, it should be actively considered as a therapeutic option. The physiological response to the first session of prone positioning was predictive of the clinical outcome of patients with COVID-19 ARDS. 

Vision Transformer using Low-level Chest X-ray Feature Corpus for COVID-19 Diagnosis and Severity Quantification (preprint)

Developing a robust algorithm to diagnose and quantify the severity of COVID-19 using Chest X-ray (CXR) requires a large number of well-curated COVID-19 datasets, which is difficult to collect under the global COVID-19 pandemic. On the other hand, CXR data with other findings are abundant. This situation is ideally suited for the Vision Transformer (ViT) architecture, where a lot of unlabeled data can be used through structural modeling by the self-attention mechanism. However, the use of existing ViT is not optimal, since feature embedding through direct patch flattening or ResNet backbone in the standard ViT is not intended for CXR. To address this problem, here we propose a novel Vision Transformer that utilizes low-level CXR feature corpus obtained from a backbone network that extracts common CXR findings. Specifically, the backbone network is first trained with large public datasets to detect common abnormal findings such as consolidation, opacity, edema, etc. Then, the embedded features from the backbone network are used as corpora for a Transformer model for the diagnosis and the severity quantification of COVID-19. We evaluate our model on various external test datasets from totally different institutions to evaluate the generalization capability. The experimental results confirm that our model can achieve the state-of-the-art performance in both diagnosis and severity quantification tasks with superior generalization capability, which are sine qua non of widespread deployment.

Severity Quantification and Lesion Localization of COVID-19 on CXR using Vision Transformer (preprint)

Under the global pandemic of COVID-19, building an automated framework that quantifies the severity of COVID-19 and localizes the relevant lesion on chest X-ray images has become increasingly important. Although pixel-level lesion severity labels, e.g. lesion segmentation, can be the most excellent target to build a robust model, collecting enough data with such labels is difficult due to time and labor-intensive annotation tasks. Instead, array-based severity labeling that assigns integer scores on six subdivisions of lungs can be an alternative choice enabling the quick labeling. Several groups proposed deep learning algorithms that quantify the severity of COVID-19 using the array-based COVID-19 labels and localize the lesions with explainability maps. To further improve the accuracy and interpretability, here we propose a novel Vision Transformer tailored for both quantification of the severity and clinically applicable localization of the COVID-19 related lesions. Our model is trained in a weakly-supervised manner to generate the full probability maps from weak array-based labels. Furthermore, a novel progressive self-training method enables us to build a model with a small labeled dataset. The quantitative and qualitative analysis on the external testset demonstrates that our method shows comparable performance with radiologists for both tasks with stability in a real-world application.