Addendum: Early triage of critically ill COVID-19 patients using deep learning

Abstract

The original Article did not reference our previous study 1 , which also presents a clinical prediction model for COVID-19 mortality/ critical illness risk. We wish to clarify the similarities and differences between the two studies in terms of cohort populations, methodology and clinical use.By using the COVID-gram, we can predict the probability of a patient to develop critical illness or death based on the risk score that integrates chest radiographic abnormality, age, hemoptysis, dyspnea, unconsciousness, number of comorbidities, cancer history, neutrophil-to-lymphocyte ratio, lactate dehydrogenase and direct bilirubin. This model was based on a traditional logistic model to predict whether or not a patient will develop into critical illness during the hospital stay.However, in the real practice, knowing the speed at which the patient will develop critical illness will also influence decision-making regarding treatment. Thus, after finishing the COVID-gram model, we further requested the time information from the data center at China National Health Commission to establish a time-dependent model based on Cox regression. During model selection, we found that deep-learning methods outperformed a traditional linear Cox model. Thus, in our article, we established the deep-learning model to predict critical illness or death at different time points (i.e. 5 days, 10 days, 30 days etc.). In this model, the panel of independent prognostic factors also differs from those used for COVID-gram and includes X-ray abnormalities, age, dyspnea, COPD (chronic obstructive pulmonary disease), number of comorbidities, cancer history, neutrophil/lymphocytes ratio, lactate dehydrogenase, direct bilirubin and creatine kinase. In addition, based on the unique advantages of deep-learning methods through which inherent data relationship can be modelled, we proposed a new imputation scheme to handle missing data automatically, making prediction more widely available in different areas and more robust for errors in data acquisition. 1. Liang, W. et al. Development and validation of a clinical risk score to predict the occurrence of critical illness in hospitalized patients with COVID-19. JAMA Intern. Med. 180, 1081-1089 (2020).

Harnessing Machine Learning in Early COVID-19 Detection and Prognosis: A Comprehensive Systematic Review

“…It appeared that 30 studies used chest X-ray images, 20 used chest CT images, one used chest X-ray and CT images, and one used chest ultrasound (US) frames. Thirteen of the included studies used ML models to predict COVID-19 disease severity [ 39 , 40 , 46 , 51 , 56 , 59 , 63 , 77 , 86 , 87 , 89 , 92 ]. Severity outcome measures included the need for ICU transfer, hospital stay time, mechanical ventilation, and death (Appendix 1).…”

“…FCONet based on Inception-v3: Sensitivity = 88.3%; Specificity = 97.9%; Accuracy = 94.9%; AUC = 0.97 Li et al [ 57 ] China Unnested case-control study Radiology Hospital-based Primary dataset Six unspecified medical centers in China Total 4,356 chest CT images: 1,296 COVID-19, 1,735 community-acquired pneumonia, 1,325 normal 90% training, 10% testing To detect COVID-19 pneumonia on chest CT Diagnosis of COVID-19 N/A Chest CT images New model based on existing backbone ML; DL; CNN COVNet (a CNN model based on pretrained RestNet50 model as a backbone) Sensitivity = 90% (95% CI:83%, 94%); Specificity = 96% (95% CI:93%, 98%); AUC = 0.95 (95% CI: 0.94, 0.99) Li et al [ 58 ] USA Unnested case-control study Radiology Hospital-based Primary dataset and secondary dataset (1) Massachusetts General Hospital (internal data); (2) Stanford Hospital (external data for training and validation) Total 581 chest X-rays: 314 for training and validation (internal dataset); 154 for testing (internal dataset); 113 for testing (external dataset) 54% training/validation, 46% testing To develop a pulmonary X-ray severity score that predicts the severity of pulmonary disease N/A Intubation death Pulmonary X-ray, severity score New model based on existing backbone ML; DL; CNN Convolutional Siamese Neural Network (a CNN model based on DensNet121 underlying subnetwork with initial pretraining on ImageNet) Internal dataset: Pearson's correlation coefficient = 0.86 (95% CI: 0.80-0.90); Spearman's rank correlation coefficient = 0.84 (95% CI: 0.77-0.88). External dataset: Pearson's correlation coefficient = 0.86 (95% CI: 0.79-0.90); Spearman's rank correlation coefficient = 0.78 (95% CI: 0.67-0.85) Liang et al [ 59 ] China Retrospective cohort Internal medicine Hospital-based Primary dataset National Health Commission (NHC) of the People's Repu...…”

Utilization of Artificial Intelligence in Disease Prevention: Diagnosis, Treatment, and Implications for the Healthcare Workforce

“…All these algorithms are helpful in the prediction of accuracy in various infectious diseases. Several prediction models have also been developed for the identification of new COVID-19-infected patients [26][27][28]. According to a 2018 Deloitte survey of 1100 US managers, whose organizations were already exploring AI, 63 percent of the companies surveyed were employing machine learning in their operations [9].…”