key: cord-0217041-52rw8zxs
authors: Chao, Hanqing; Fang, Xi; Zhang, Jiajin; Homayounieh, Fatemeh; Arru, Chiara D.; Digumarthy, Subba R.; Babaei, Rosa; Mobin, Hadi K.; Mohseni, Iman; Saba, Luca; Carriero, Alessandro; Falaschi, Zeno; Pasche, Alessio; Wang, Ge; Kalra, Mannudeep K.; Yan, Pingkun
title: Integrative Analysis for COVID-19 Patient Outcome Prediction
date: 2020-07-20
journal: nan
DOI: nan
sha: d56a5009775dbd3b170039f03cb26067afb79d32
doc_id: 217041
cord_uid: 52rw8zxs

While image analysis of chest computed tomography (CT) for COVID-19 diagnosis has been intensively studied, little work has been performed for image-based patient outcome prediction. Management of high-risk patients with early intervention is a key to lower the fatality rate of COVID-19 pneumonia, as a majority of patients recover naturally. Therefore, an accurate prediction of disease progression with baseline imaging at the time of the initial presentation can help in patient management. In lieu of only size and volume information of pulmonary abnormalities and features through deep learning based image segmentation, here we combine radiomics of lung opacities and non-imaging features from demographic data, vital signs, and laboratory findings to predict need for intensive care unit (ICU) admission. To our knowledge, this is the first study that uses holistic information of a patient including both imaging and non-imaging data for outcome prediction. The proposed methods were thoroughly evaluated on datasets separately collected from three hospitals, one in the United States, one in Iran, and another in Italy, with a total 295 patients with reverse transcription polymerase chain reaction (RT-PCR) assay positive COVID-19 pneumonia. Our experimental results demonstrate that adding non-imaging features can significantly improve the performance of prediction to achieve AUC up to 0.884 and sensitivity as high as 96.1%, which can be valuable to provide clinical decision support in managing COVID-19 patients. Our methods may also be applied to other lung diseases including but not limited to community acquired pneumonia.

Coronavirus disease 2019 (COVID- 19) , which results from contracting an extremely contagious beta-coronavirus, is responsible for the latest pandemic in human history. The resultant lung injury from COVID-19 pneumonia can progress rapidly to diffuse alveolar damage, acute lung failure, and even death [1, 2] . Given the highly contagious nature of the infection, the burden of COVID-19 pneumonia has imposed substantial constraints on the global healthcare systems. In this paper, we present a novel framework of integrative analysis of heterogeneous data including not only medical images, but also patient demographic information, vital signs and laboratory blood test results for assessing disease severity and predicting intensive care unit (ICU) admission of COVID-19 patients. Screening out the high-risk patients, who may need intensive care later, and monitoring them more closely to provide early intervention may help save their lives.

Reverse transcription polymerase chain reaction (RT-PCR) assay with detection of specific nuclei acid of SARS-CoV-2 in oral or nasopharyngeal swabs is the preferred test for diagnosis of COVID-19 infection. Although chest computed tomography (CT) can be negative in early disease, it can achieve higher than 90% sensitivity in detecting COVID-19 pneumonia but with low specificity [3] . For diagnosis of COVID-19 pneumonia, CT is commonly used in regions with high prevalance and limited RT-PCR availability as well as in patients with suspected false negative RT-PCR. CT provides invaluable information in patients with moderate to severe disease to assess the severity and complications of COVID-19 pneumonia [4] . Prior clinical studies with chest CT have reported that qualitative scoring of lung lobar involvement by pulmonary opacities (high lobar involvement scores) can help assess severe and critical COVID-19 pneumonia. Li et al. [5] showed that high CT severity scores (suggestive of extensive lobar involvement) and consolidation are associated with severe COVID-19 pneumonia. Zhao et al. [6] reported that extent and type of pulmonary opacities can help establish severity of COVID-19 pneumonia. The lung attenuation values change with the extent and type of pulmonary opacities, which differ in patients with more extensive, severe disease from those with milder disease. Most clinical studies focus on qualitative assessment and grading of pulmonary involvement in each lung lobe to establish disease severity, which is both timeconsuming and associated with interobserver variations [6, 7] .

To address the urgent clinical needs, artificial intelligence (AI), especially deep learning, has been applied to COVID-19 CT image analysis [8] . AI has been used to differentiate COVID-19 from community acquired pneumonia (CAP) on chest CT images [9, 10] . To unveil what deep learning uses to diagnose COVID-19 from CT, Wu et al. [11] proposed an explainable diagnosis system by classifying and segmenting infections. Gozes et al. [12] developed a deep learning based pipeline to segment lung, classify 2D slices and localize COVID-19 manifestation from chest CT scans. Shan et al. [13] went on to quantify lung infection of COVID-19 pneumonia from CT images using deep learning based image segmentation.

Among the emerging works, a few AI based methods target at severity assessment from chest CT. Huang et al. [14] developed a deep learning method to quantify severity from serial chest CT scans to monitor the disease progression of COVID-19. Tang et al. [15] used random forest to classify pulmonary opacity volume based features into four severity groups. By automatically segmenting the lung lobes and infection areas, Gozes et al. [16] suggested a "Corona Score" to measure the progression of disease over time. Zhu et al. [17] further proposed to use AI to predict if a patient may develop severe symptoms of COVID-19 and how long it may take if that is the case. Although promising results have been presented, the existing methods primarily focus on the volume of pulmonary opacities and their relative ratio to the lung volume for severity assessment. The type of pulmonary opacities (e.g. ground glass, consolidation, crazy-paving pattern, organizing pneumonia) is also an important indicator of the stage of the disease and is often not quantified by the AI algorithms [18] .

Furthermore, in addition to measuring and monitoring the progression of severity, it could be life-saving to predict mortality risk of patients by learning from the clinical outcomes. Since majority of the infected patients will recover, managing the high-risk patients is the key to lower the fatality rate [19, 20, 21] . Longitudinal study analyzing the serial CT findings over time in patients with COVID-19 pneumonia shows that the temporal changes of the diverse CT manifestations follow a specific pattern correlating with the progression and recovery of the illness [22] . Thus, it is promising for AI to perform this challenging task.

In this paper, our objective is to predict outcome of COVID-19 pneumonia patients in terms of the need for ICU admission with both imaging and nonimaging information. The work has two major contributions.

1. While image features have been commonly exploited by the medical image analysis community for COVID-19 diagnosis and severity assessment, nonimaging features are much less studied. However, non-imaging health data may also be strongly associated with patient severity. For example, Yan et al. [23] showed that machine learning tools using three biomarkers, including lactic dehydrogenase (LDH), lymphocyte and high-sensitivity Creactive protein (hs-CRP), can predict the mortality of individual patients. Thus, we propose to integrate heterogeneous data from different sources, including imaging data, age, sex, vital signs, and blood test results to predict patient outcome. To the best of our knowledge, this is the first study that uses holistic information of a patient including both imaging and non-imaging data for outcome prediction. 2. In addition to the simple volume measurement based image features, radiomics features are computed to describe the texture and shape of pul-monary opacities. A deep learning based pyramid-input pyramid-output image segmentation algorithm is used to quantify the extent and volume of lung manifestations. A feature dimension reduction algorithm is further proposed to select the most important features, which is then followed by a classifier for prediction.

It is worth noting that although the presented application on COVID-19 pneumonia, the proposed method is a general approach and can be applied to other diseases.

The proposed method was evaluated on datasets collected from teaching hospitals across three countries, These datasets included 113 CT images from Firoozgar Hospital (Tehran, Iran)(Site A), 125 CT images from Massachusetts General Hospital (Boston, MA, USA)(Site B), and 57 CT images from University Hospital Maggiore della Carita (Novara, Piedmont, Italy)(Site C). Promising experimental results for outcome prediction were obtained on all the datasets with our proposed method, with reasonable generalization across the datasets. Details of our work are presented in the following sections.

The data used in our work were acquired from three sites. All the CT imaging data were from patients who underwent clinically indicated, standard-of-care, non-contrast chest CT without intravenous contrast injection. Age and gender of all patients were recorded. For datasets from Sites A and B, lymphocyte count and white blood cell count were also available. For datasets of Sites A and C, peripheral capillary oxygen saturation (SpO2) and temperature on hospital admission were recorded. Information pertaining patient status (discharged, deceased, or under treatment at the time of data analysis) was also recorded as well as the number of days of hospitalization to the outcome.

Site A Dataset. We reviewed medical records of adult patients admitted with known or suspected COVID-19 pneumonia in Firoozgar Hospital (Tehran, Iran) between February 23, 2020 and March 30, 2020. Among the 117 patients with positive RT-PCR assay for COVID-19, three patients were excluded due to presence of extensive motion artifacts on their chest CT. With one patient who neither admitted to ICU nor discharged, 113 patients are used in this study.

Site B Dataset. We reviewed medical records of adult patients admitted with COVID-19 symptom in MGH between March 11 and May 3, 2020. 125 RT-PCR positive admitted patients underwent unenhanced chest CT are selected to form this dataset.

Site C Dataset. We reviewed medical records of adult patients admitted with COVID-19 pneumonia in the Novara Hospital (Piedmont, Italy) between March 4, 2020 and April 6, 2020. We collected clinical and outcome information of 57 patients with positive RT-PCR assay for COVID-19.

Two experienced thoracic subspecialty radiologists evaluated all chest CT examinations and recorded opacity type, distribution and extent of lobar involvement. Information on symptom duration prior to hospital admission, duration of hospital admission, presence of comorbid conditions, laboratory data, and outcomes (recovery or death) was obtained from the medical records. Entire lung volume was segmented on thin-section DICOM images (1.5-2 mm) to obtain whole-lung analysis. Statistics of the datasets are shown in Tables 3-5 in Section 3.4.

Two chest radiologists (SRD with 16 years of experience, MKK with 14 years of experience) reviewed all CTs unware of the information on clinical features, laboratory data, and outcomes information. On an imaging workstation (MicroDicom DICOM Viewer, Sofia, Bulgaria), images were reviewed in lung windows (window center -600 HU; window width 1500 HU). The radiologists noted the opacities type (pure ground-glass, mixed ground-glass and consolidation, pure consolidation, organizing pneumonia, nodular and ground-glass with interlobular thickening (crazy-paving pattern)) and extent of individual lobe involvement of right upper, left upper, right middle, right lower and left lower lobes. The latter was assessed on a 6-point scale (0% lobar involvement = score 0; <5% involvement of lobar volume = score 1; 5-25% involvement of lobe = score 2; 26-50% lobar involvement = score 3; 51-75% lobar involvement = score 4; >75% lobar involvement = score 5) [13] . Addition of individual extent scores provided the overall extent of pulmonary opacities with a minimum possible score of 0 and maximum score of 25 [9, 13] .

In order to predict the need for ICU admission of patients with COVID-19 pneumonia, we use three types of imaging and non-imaging features. Our adopted features include hierarchical lobe-wise quantification features (HLQ), whole lung radiomics features (WLR), and features from demographic, vital signs, and blood examination (DVB). Figure 1 shows an overview of the overall framework of the presented work. In the rest of this section, we first introduce the details of these features. Since it is challenging to fuse the large number of inhomogeneous features together, a feature selection strategy is proposed, followed by random forest based classification [24] .

In our work, we employed deep neural networks to segment both lungs, five lung lobes and pulmonary opacities (as regions of infection) from non-contrast chest CT examinations. For training purpose, we semi-automatically labeled 71 CT volumes using 3D Slicer [25] . For lung lobe segmentation, we adopted the automated lung segmentation method by Hofmanninger et al. [26] . The pre-trained model 1 was fine-tuned with a learning rate of 1 × 10 −5 using our annotated data. The tuned model was then applied to segment all the chest CT volumes. Segmentation of pulmonary opacities was completed by our previously proposed method, Pyramid Input Pyramid Output Feature Abstraction Network (PIPO-FAN) [27] with publicly released source code 2 . Figure 2 shows the segmentation results of lung lobes and pulmonary opacities. From axial and 3D view, we can see that the segmentation models can smoothly and accurately predict isolate regions with pulmonary opacities.

Based on the segmentation results in Section 3.1, we then compute the ratio of opacity volume over different lung regions, which is a widely used measurement to describe the severity [15, 17] . The lung regions include the whole lung, the left lung and the right lung and 5 lung lobes (lobe# 1-5) as shown in . These four HU ranges correspond to normal lungs, ground glass opacity (GGO), consolidation, and regions with pulmonary calcification, respectively. As a result, each CT image was partitioned to 32 components (8 ROIs × 4 ranges/ROI). We extracted two quantitative features from each part, i.e., volumes of pulmonary opacities (VPO) and ratio of pulmonary opacities to the corresponding component (RPO), as defined below:

where x is a selected component (among the 32 components). Segment(x) denotes the pulmonary opacities in the selected component x based on the segmentation of pulmonary opacities in Figure 2 . V (·) denotes the volume of the selected part.

To more comprehensively describe information in the CT image, we also extracted multi-dimensional radiomics features [28] of all pulmonary opacities. Compared with HLQ feature, although they are all image-based features, they describe the pulmonary opacities from different aspects. HLQ features focus on the pulmonary opacities volume and position of region of interest, while WLR focus on their shape and texture.

For each chest CT volume, we first masked out non-infection regions based on the infection segmentation results, then four kinds of radiomics features are calculated on the volume, i.e., shape, first-order, second-order and higherorder statistics features [29] . Shape features describe the geometric information. First-order, second-order and higher-order statistics features all describes texture information. First-order statistics features describe the distribution of individual voxel values without concerning spatial correlation. Second-order features describe local structures, which provide correlated information between adjacent voxels and statistical measurement of intra-lesion heterogeneity. Second-order features include those extracted using gray level dependence matrix (GLDM), gray level co-occurrence matrix (GLCM), grey level run length matrix (GLRLM), grey level size zone matrix (GLSZM), and neighboring gray tone difference matrix (NGTDM). Higher-order statistics features are computed using the same methods as second-order features but after applying wavelets and Laplacian transform of Gaussian(LoG) filters. The higher-order features help identify repetitive patterns in local spatial-frequency domain in addition to suppressing noise and highlighting details of images.

We used the Pyradiomics package [30] to extract the above described radiomics features from COVID19 chest CT images. For each chest CT volume, a total of 1691 features are extracted. The number of radiomics features for each feature type are summarized in Table 1 . Based on the description above, these features can be categorized into two main groups, i.e., 17 shape features and 93 texture features.

To extracted various features, different image filters are applied before feature extraction. Table 2 shows the details of all 18 image filter types used in our work, including no filter, square filter, square-root filter, logarithm filter, exponential filter, wavelet filter and LoG filter. The image filtered by a 3D wavelet filter has eight channels, including HHH, HHL, HLH, LHH, HLL, LHL, LLH and LLL. The Laplacian of Gaussian (LoG) filters have a hyper-parameter σ which is the standard deviation of the Gaussian distribution. We used five different σ values in our study, i.e., {0.5, 1.5, 2.5, 3.5, 4.5}. Note that shape features are only extracted from the original images (no filter was applied). In addition to features extracted from images, we incorporated features from demographic data (contained by all three datasets), vital signs (from Sites A and B), and laboratory data (from Sites A and C) (DVB). Specifically, such features include patients' age, gender, white blood cell count (WBC), lymphocyte count (Lym), Lym to WBC ratio (L/W ratio), temperature and blood oxygen level (SpO2). These data are highly correlated with the ICU admission of patients when they were admitted to a hospital. Table 3 -5 show the statistics of the above features in Site A, Site B, and Site C datasets respectively. Non-imaging features are not all available for some patients. The number of the collected data for each feature is listed in the last column of the tables.

In our work, random forest (RF) [24] classifier, a widely-used ensemble learning method consisting of multiple decision trees, is chosen for predicting ICU admission due to its several nice properties. First, RF is robust to small data size. Second, it can generate feature importance ranking and is thus highly interpretable. Aggregating all the features introduced above, we have 1,762 features in total. Due to the limited data size, the model would easily overfit with all features as input. Thus, we first used RF to rank all the features, then we selected the top K features for our final prediction.

We ranked the feature based on their Gini importance [31] . It is calculated during the training of RF by averaging the decrease of Gini impurity over all trees. Due to the randomness of RF, Gini importance of features may vary when RF model is initialized with different random seeds. Therefore, in our study, feature ranks are computed 100 times with different random seeds. Each time every feature will get a score being its rank. The final feature rank is obtained by sorting the total summed score of each feature.

Based on the rank of all the features, we select top K ∈[1,100] features to train the RF model and calculate the prediction performance in terms of AUC.

This section presents the experimental results of the developed methods. We show the effectiveness of our proposed method on the three datasets separately through both ablation studies and comparison with other state-of-the-art approaches. We did not merge the datasets because of two reasons. First, not all the non-image features were available from the participating sites. Second, the treatment and admission criteria at the participating sites were likely different from each other. Thus, it would not be rigorous to merge the datasets. First, on each dataset the proposed methods with different combinations of features is compared with other state-of-the-art approaches. In this part of experiments, we also included results of a deep learning network which take the best feature combination as input. The deep learning network is composed by three fully connected layers with output dimension of 64, 16 and 2 respectively. Dropout was applied on the first two layers with p= 50%. The cross entropy loss is used to train the network. In five fold cross validation, 3 folds are used to train the network, one fold is leaved as validation set, and the other one is used as test set. Second, the generalization ability of the feature combination learned by our model is studied across three datasets.

Several recent works have shown the importance of using machine learning models to predict patients' outcomes based on lobe-wise quantification features. The infection volume and infection ratio of the whole lung, right/left lung, and each lobe/segment are calculated as quantitative features in [15] . Random forest classifier is used to select the top-ranking features and make the severity assessment based on these features. In another work by Zhu et al. [17] , the authors present a novel joint regression and classification method to identify the severity cases and predict the conversion time from a non-severe case to the severe case. Their lobe-wised quantification features include the infection volume, density feature and mass feature. As we mentioned earlier, all existing image analysis-based outcome prediction works use only image features. We take the features in the two papers as baseline to compare with our work. Receiver Operating Characteristics (ROC) curves of the feature combinations are shown in Figure 3 . For each feature combination, the features are selected only from the feature categories available in the combination using the approach introduced in Section 3.5. For example, HLQ+DVB indicates that only features from these two groups, HLQ and DVB, are selected and used. The number of features K used to obtain the best results are listed in Table 6 . To alleviate the stochasticity of the results, for each feature combination, five RF models with different random seeds are trained and tested with five fold cross validation. The curves shown here are thus the mean results of the five models. The figure legend gives the mean Area Under the Curves (AUCs) of the feature combinations as well as the standard deviation (mean±std). It can be seen that the combination of all three kinds of features, WLR+HLQ+DVB, obtained the best result with an AUC of 0.88±0.01. The variation of AUC along with number of selected features on Site A dataset is presented in Fig. 4 . As marked by the light blue dash line in Fig. 4 , AUC reaches the maximum value when the top 52 features are selected. Details of the 52 selected features are presented in Table 11 at the end of this paper due to its large size. One-tailed t-test is used to evaluate the statistical significance between a method and the best performer. Table 6 summarizes the AUC values and sensitivity with significance test p values and 95% confidence interval (95% CI). The classification threshold is selected by control the positive prediction value (PPV) to be 70%. The combination of WLR+HLQ+DVB significantly exceeds other reported methods [15, 17] with p ≤ 0.001. Further, we achieved a sensitivity of 84.3% while retaining a PPV at 70%. It suggests that our model can rapidly prioritize over 80% patients who would develop into critical conditions, if we allow 3 false positive cases in every 10 positive predictions. With such prediction, hospitals may allocate limited medical resources more efficiently to potentially prevent such conversion and save more lives. Under the same setting, the sensitivity of the model is 79.4%, the accuracy is 81.2%. The deep learning model compared in Table 6 also used all three kinds of features.

In the comparison among different combinations of the features, we can see that the results are generally improved with more feature sources added. Comparison between WLR+HLQ (line 6) and WLR+HLQ+DVB (the last line) shows that, on this dataset, introducing non-imaging features can significantly improve the performance (p<0.001 for AUC and p=0.008<0.05 for sensitivity), which further indicates that non-imaging features and image based features are complementary. On the other hand, the comparison with WLR+DVB (line 7) and HLQ+DVB (line 8) shows that the improvement of WLR+HLQ+DVB was not significant (p>0.05). It suggests that different kinds of image based features may contain redundant information and adding more features from the same source only results in marginal improvement. The same set of experiments were repeated on the Site B dataset. Table 7 and Fig. 5 shows the results. The number of features K used to obtain the best results for each combination are listed in Table 7 . It can be seen that, on Exponential-GLRLM-ShortRunLowGL 6 LoG(σ = 4.5)-NGTDM-Contrast 7

Exponential-GLRLM-ShortRun 8 LoG(σ = 3.5)-GLSZM-ZoneEntropy 9

LoG(σ = 4.5)-NGTDM-Busyness 10 Sqrt-NGTDM-Strength 11 LoG(σ = 3.5)-GLCM-Imc2 12 LLH-GLSZM-GLNonUnif Table 7 used the combination of WLR and HLQ. Table 8 lists the 12 WLR+HLQ features used to obtain the best results. Table 9 : Comparison among the exist state-of-the-art methods and the proposed methods with different feature combinations on Site C dataset. As image features in [15] couldn't achieve 70% PPV, its sensitivity was not included. Results on Site C dataset are shown in Table 9 and Fig. 5 . The non-imaging (DVB) features alone also didn't achieve well performance. It might be because many petients' DVB features are missing or incomplete as shown in Table. 5 Yet, the introduce of DVB features significantly improves the AUC performance of HLQ features (p = 0.014 < 0.05). In this experiment, the best AUC value, 0.840, is achieved by merging all three kinds of features. While maintaining a PPV of 70%, it achieved a sensitivity of 94.4%, a specificity of 33.3% and an accuracy of 71.9%. The deep learning model compared in Table 9 used all three kinds of data. Table 12 shows the 35 features used to obtain the best results at the end of this paper due to its large size.

In this section, we further evaluate if feature combinations learned from one site can be generated to other sites. Experiments on all 6 Table 10 .

There were tremendous differences in the geographic distribution and scanner technologies used for imaging patients at the three participating sites. Despite this, we achieved AUC values as high as 0.754 for WLR+HLQ features. Some variations in the AUCs and performance of our model across different sites is expected due to challenges associated with acquisition of consistent data (20) variables and practices. Our study stresses the need to combine imaging findings with clinical, laboratory and management variables which can improve the model performance, aid in better performance statistics on each dataset. On the other hand, complexities of disease and its outcomes are tied to local factors and stress the importance of tweaking the best models based on rich local or institutional level factors rather than a single one-type-fit-all model.

In this paper, we propose to combine size and volume information of the lungs and manifestations, radiomics features of pulmonary opacities and nonimaging DVB features to predict need for ICU admission in patients with COVID-19 pneumonia. Metrics related to ICU admission rates, need and availability are key markers in management of individual patients as well as in resource planning for managing high prevalence diseases. To the best of our knowledge, this is the first study that uses holistic information of a patient including both imaging and non-imaging data to predict patient outcome.

Although promising results were achieved, the study has a few limitations. First of all, due to the limited size of our datasets, we could not conduct more fine-grained outcome predictions. The size of the available datasets could also be the reason that more complex models, such as deep neural networks, did not perform well in our experiments. Once larger datasets are available, our model can be rapidly adapted to assess generalization ability and to establish implications on datasets from other sites. Efforts are underway (such as within Radiological Society of North America) to establish such imaging datasets of COVID-19 pneumonia. Second, variations in performance of different imaging and clinical features on datasets from three sites underscore the need for careful local vetting of deep learning predictive models. Future models should take into account regional bias introduced from different criteria on imaging use, underlying patient comorbidities, and management strategies, so that more robust models can be built. This also goes beyond the generalization ability of machine learning in medical applications. The best and most relevant results likely require regional, local or even site-specific tuning of predictive models. This is especially true in context of the three sites, which are under very different healthcare systems as in our study. We believe that this limitation is not unique to our model. Last but not the least, another limitation of our study pertains to the lack of access to the specific treatment regimens at the three sites; their inclusion could have further enhanced the accuracy of our algorithm. However, it also suggests that this generic approach can be trained on data from a hospital to create a customized predictive model for clinical decision support.

In summary, our integrative analysis machine learning based predictive model can help assess disease burden and forecast meaningful patient outcomes with high predictive accuracy in patients with COVID-19 pneumonia. Many patients with adverse outcomes from COVID-19 pneumonia and cardiorespiratory failure develop diffuse alveolar damage and adult respiratory distress syndrome (ARDS), which are also well-known end stage manifestations of other pulmonary diseases such as from other infections and lung injuries. Although we did not test our model in patients with ARDS from non-COVID causes, given the overlap in imaging and clinical features of respiratory failure, we expect that the methods of quantifying pulmonary opacities used in our approach will extend beyond COVID-19 pneumonia. Further studies will help assess such applications beyond the current pandemic of COVID-19 pneumonia. Org-GLDM-LowGray Original-Shape-LeastAxisLength 4 LoG(σ = 2.5)-NGTDM-Complexity 5

Lobe#2 RPO HU4

Lobe#5 RPO HU3 7 LLL-GLRLM-LongRunLG 8 Left Lung RPO HU4 9 Lym count 10 Lobe#2 VPO HU3

Lobe#2 VPO HU2 12 LLL-GLRLM-LGRun

Strength 16 Lobe#5 VPO HU3

SmallAreaHG 24 LoG(σ = 2.5)-GLCM-MCC 25 WBC 26 Org-Shape-LeastAxisLength

LongRunLG 28 LoG(σ = 1.5)-GLCM-Corr

Lobe#2 Infection Ratio HU4 30 HLL-FirstOrder-Mean

Left Lung VPO HU2 32 LoG(σ=2.5)-GLSZM-LALG

Lobe#2 VPO HU4 36 HLH-GLSZM-Zone%

Lobe#4 VPO HU4 40 LLL-GLCM-SumAverage

LoG(σ = 2.5)-NGTDM-Busyness 42 LLH-FirstOrder-Skewness

Lobe#2 RPO HU3 44 LoG(σ = 1.5)-GLSZM-ZonVar

LoG(σ = 2.5)-GLCM-Imc2

LoG(σ = 1.5)-GLSZM-LALG

SumAverage 50 Left Lung VPO HU3

JointAverage 52 Left Lung RPO HU3

Red text indicates non-imaging features. Green text indicates lobe-wise quantification features, HU1-HU4 are the four HU intervals

Reninangiotensinaldosterone system inhibitors in patients with Covid-19

Renin-angiotensin system blockers and the COVID-19 pandemic: at present there is no evidence to abandon reninangiotensin system blockers

Diagnostic Performance of CT and Reverse Transcriptase-Polymerase Chain Reaction for Coronavirus Disease

= 2.5)-GLRLM-ShortRunLowGL

LoG(σ = 3.5)-firstorder-Range 12 LoG(σ = 3.5)-GLDM-LowGL

GLCM-Corr 14 Age 15 LoG(σ = 2.5)-GLRLM-LowGLRun 16 LoG

GLCM-MaxProb 18 LoG(σ = 4.5)-GLCM-Idmn

Mean 20 LoG(σ = 2.5)-GLDM-SmallDepLowLG

LoG(σ = 4.5)-FirstOrder-Range 22 LoG(σ = 2.5)-GLDM-LargeDepLowLG

LHL-FirstOrder-Maximum 24 LoG(σ = 2.5)-GLCM-LDMN

LoG(σ = 1.5)-NGTDM-Strength 26 LoG(σ = 4.5)-NGTDM-Complexity

LoG(σ = 2.5)-GLSZM-LowGray 28 LHL-FirstOrder-Kurtosis

LoG(σ = 1.5)-GLCM-MCC 30 LoG(σ = 3.5)-GLCM-IMC2

LoG(σ = 2.5)-GLCM-SumAverage 34 LoG(σ = 2.5)-GLCM-LDN

Kurtosis Red text indicates non-imaging features. Green text indicates lobe-wise quantification features, HU1-HU4 are the four HU intervals

Chest CT severity score: An imaging tool for assessing severe covid-19

The Clinical and Chest CT Features Associated with Severe and Critical COVID-19 Pneumonia., Investigative radiologyPublisher

Relation Between Chest CT Findings and Clinical Conditions of Coronavirus Disease (COVID-19) Pneumonia: A Multicenter Study

Correlation of Chest CT and RT-PCR Testing in Coronavirus Disease 2019 (COVID-19) in China: A Report of 1014 Cases

Review of artificial intelligence techniques in imaging data acquisition, segmentation and diagnosis for covid-19

Artificial Intelligence Distinguishes COVID-19 from

Adaptive Feature Selection Guided Deep Forest for COVID-19 Classification with Chest CT

JCS: An Explainable COVID-19 Diagnosis System by Joint Classification and Segmentation

Coronavirus Detection and Analysis on Chest CT with Deep Learning

Lung Infection Quantification of COVID-19 in CT Images with Deep Learning

Serial quantitative chest ct assessment of covid-19: Deep-learning approach

Severity Assessment of Coronavirus Disease 2019 (COVID-19) Using Quantitative Features from Chest CT Images

Rapid AI Development Cycle for the Coronavirus (COVID-19) Pandemic: Initial Results for Automated Detection & Patient Monitoring using Deep Learning CT Image Analysis

Joint Prediction and Time Estimation of COVID-19 Developing Severe Symptoms using Chest CT Scan

CT imaging features of 2019 novel coronavirus (2019-nCoV)

Likelihood of survival of corona virus disease

Intensive care management of coronavirus disease 2019 (covid-19): challenges and recommendations

Clinical observation and management of covid-19 patients

Temporal changes of CT findings in 90 patients with COVID-19 pneumonia: a longitudinal study

An interpretable mortality prediction model for COVID-19 patients

Random forests

3d slicer: a platform for subjectspecific image analysis, visualization, and clinical support, in: Intraoperative imaging and image-guided therapy

Automatic lung segmentation in routine imaging is a data diversity problem, not a methodology problem

Multi-organ segmentation over partially labeled datasets with multi-scale feature abstraction

Radiomics: images are more than pictures, they are data

Radiomics: the facts and the challenges of image analysis

Computational radiomics system to decode the radiographic phenotype

Classification and regression trees

This work was partially supported by National Institute of Biomedical Imaging and Bioengineering (NIBIB) under award R21EB028001 and National Heart, Lung, and Blood Institute (NHLBI) under award R56HL145172.