We collected samples from 46 patients who visited People’s Hospital of Yicheng City between January 16, 2020, and March 4, 2020, and were diagnosed with COVID-19 according to the Chinese Government Diagnosis and Treatment Guideline (Trial 5th version; Medicine, 2020). For a confirmed diagnosis of COVID-19, nucleic acid was extracted from sputum or throat swab samples using a nucleic acid extractor (EX3600, Shanghai Zhijiang Biotechnology Co.) and a nucleic acid extraction reagent (No. P20200201, Shanghai Zhijiang Biotechnology Co.).

Fluorescence-based quantitative polymerase chain reaction (PCR; ABI7500) and SARS-CoV-2 nucleic acid detection kit (triple fluorescence PCR, No. P20200203, Shanghai Zhijiang Biotechnology Co.) were used for nucleic acid detection. This kit uses a one-step reverse transcription–PCR combined with Taqman technology to detect RNA-dependent RNA polymerase (RdRp), envelope (E), and nucleocapsid (N) genes. PCR results were concluded to be positive if (1) RdRp gene was positive (cycle threshold [Ct]<43) and either E or N gene was positive (Ct<43), or (2) if two sequential tests for RdRp were positive and those for E and N genes were negative. The 46 study patients with COVID-19 were classified into 2 types: (1) nonsevere, comprising patients showing mild symptoms without radiological manifestations of pneumonia, fever, or respiratory tract symptoms with radiological manifestations of pneumonia, and (2) severe, comprising patients meeting any of the following criteria—respiratory rate ≥30 breaths/min, pulse oxygen saturation ≤93% in resting state, partial pressure of arterial oxygen ≤300 mm Hg (1 mm Hg=0.133 kPa), respiratory failure requiring mechanical ventilation, shock incidence, and admission to intensive care unit with other organ failure. In total, 10 patients were categorized as severe cases and 36, as nonsevere cases. The last follow-up of these patients was on March 10, 2020.

Approval for studies on CT screening and clinical test results was obtained from the Medical Ethics Committee of The People’s Hospital of Yicheng City, China (2020Yc002)

We collected and reviewed clinical information of 46 patients with COVID-19 after admission, including clinical signs and symptoms, comorbidities, travel history, laboratory tests, and CT scans. To consolidate all patients’ records into a single table, missing records for a given day were noted as “NA” (not available). In all, we obtained 178 records (27 severe and 151 nonsevere cases) from 105 different laboratory tests and chest CT images. Note that throughout the clinical course, each patient had more than one record variably classified as severe or nonsevere. Patients with at least one severe record were classified as severe cases.

We identified 44 laboratory tests that had more than 50% missing values (NA), and we then imputed the NAs with the mean values. Related laboratory tests were identified based on the criterion that the P value (Mann-Whitney U test) between the severe and nonsevere groups is smaller than .05. In all, we found that 36 laboratory tests were related to the detection of COVID-19 severity. The patients’ CT images were processed using a pretrained CNN with a U-Net structure [16] to segment the lung lobes from the background. The intensities were then normalized to grayscale for all patients before further analysis. We then analyzed the intensities of the 3D CT volumes within lung masks to obtain CT features for each record.

Prediction models were developed to predict patient severity based on laboratory and CT signatures collected at corresponding dates. Each patient record was considered a sample for a model; as a result, 178 samples were evaluated using those models. Before using model prediction, we used random forest importance score, mutual information, and fold change as possible approaches to select important model features while avoiding potential overfitting. We found mutual information to be the most robust approach. We considered different candidate machine learning models, including random forest classifier, gradient boost classifier, XGB classifier, logistic classifier, and supported vector machine. Random forest was found to be the best classifier, and model parameters were optimized using a genetic algorithm (Tree-Based Pipeline Optimization Tool). The area under the curve of the receiver operating characteristic (AUROC) and F₁ scores were used to evaluate model accuracy considering the dataset imbalance. All models were trained with 5-fold cross-validation with stratified train-test splits that preserve the percentage of samples in severe and nonsevere groups. All cross-validated results were averaged over 20 runs.

Forecasting models were built to forecast patient severity based on laboratory and CT signatures collected from nonsevere cases at admission. In these models, instead of the patients’ records, the patients themselves were considered as samples to build forecasting relationships. CT records were not collected as frequently as laboratory tests were performed, and initial, nonsevere CT records were not available for 3 severe cases. Therefore, we built two separate random forest models based on CT features and laboratory tests with 7 and 10 severe cases, respectively. Other model details were identical to those of the severity prediction models.

We collected clinical data of 46 patients with COVID-19 who were admitted at the People’s Hospital of Yicheng City, between mid-January and early-March 2020. We recorded 305 biochemical test results from 105 different tests, based on the clinical reports of all 46 study patients (Multimedia Appendix 1). General patient information is shown in Table 1. The general trend that older patients with COVID-19 tend to develop more systemic symptoms was not observed in our study [17]. However, patients with comorbidities, especially diabetes and hypertension, tended to develop more severe symptoms than others. Moreover, patients with severe COVID-19 typically experienced fatigue, anorexia, malaise, chest congestion, and shortness of breath.

In all, 52% (24/46) patients had a travel history to or from Wuhan within the past 1 month, and 20% (9/46) patients had clear exposure history in the local city (Table 1). According to the patients’ medical records, 80% of the severe cases had one or more comorbidities, whereas only 34% of the nonsevere cases had comorbidities. This finding is consistent with that of previous studies [18]. Moreover, 81% (37/46) patients had cough and only 20% (9/46) patients reported sputum production. Fever was the most common symptom; however, severe cases (7/10, 70%) had a higher proportion of mid- to high-grade fever (ie, >38.9°C) than the nonsevere cases (15/36, 42%). More than half of the patients (25/46, 54%) experienced fatigue, and approximately three-quarters of them had anorexia (33/46, 72%) or malaise (34/46, 74%); these symptoms were observed in almost all severe cases (fatigue, 9/10, 90%; anorexia, 9/10, 90%; and malaise, 10/10, 100%). Headache, nausea, diarrhea, and dyspnea were rarely observed in both severe and nonsevere cases. Moreover, less than half of all patients reported chest congestion (16/46, 35%) or shortness of breath after activity (19/46, 41%), and these symptoms were approximately 20% more common in severe cases than in nonsevere cases.

Data processing yielded 61 laboratory tests results, 36 of which were significantly related to severity. Eight related laboratory tests that showed the largest fold change are illustrated in Figure 1A. Among these tests, D-dimer, LDH, and lymphocytes were found to be associated with mortality risk [17]. Principal component analysis results clearly showed separation between the severe and nonsevere groups, indicating that the COVID-19–related laboratory tests can be used to identify disease severity (Figure 1B). To build a statistical model to predict severity, we first selected the most important laboratory tests to avoid overfitting. Three different approaches—fold change, random forest importance score, and mutual information—were considered the top-ranking laboratory features. In fact, the three approaches led to very similar ranking, and the top features obtained from mutual information resulted in the largest intersection with those obtained from the other two approaches (Figure 1C). This finding suggests that mutual information is the most robust feature selector among the three abovementioned approaches; therefore, we used mutual information to select laboratory features to be used in the model. We used a random forest model with hyperparameters optimized by a genetic algorithm (see Methods) to predict severity based on laboratory features. Our results suggested that the prediction accuracy does not further increase with an increase in the number of laboratory features beyond 12. As a result, a signature of the top 12 laboratory features was considered, and the corresponding model yielded a cross-validated AUROC score of 0.88 and an F₁ score of 0.69 (Figure 1D).

To extract CT features, we first segmented the lungs from the CT volumetric images using a deep CNN, U-Net. Because the CNN was pretrained with several annotated datasets, including a COVID-19 dataset from MedSeg [16], we directly transferred the trained CNN to segment CT images of the study patients. The CT slices acquired across a clinical course of a patient are shown in Figure 2A. Right after onset, the patient was diagnosed with nonsevere disease, with no apparent opacity visible in lung CT. The patient was classified as severe on Day 4, and this continued for 2 weeks thereafter, with increasing amounts of ground-glass opacity and patchy consolidation. The ground-glass opacity and consolidation started to fade away from Day 27, and on Day 30, the patient was confirmed to be asymptomatic. As seen in Figure 2A, the opacity of the segmented lung lobes is associated with disease severity. The opacity can be represented by the intensity distribution within the segmented lung volumes (Figure 2B). Note that only the slices in the middle are shown in Figure 2A for illustration purposes; all slices were considered, however, to determine intensity distribution. As the symptoms became severe, the background became increasingly opaque, as indicated by the peak locations and peak heights of the intensity distribution. The distribution also changed from unimodal to bimodal. Therefore, we considered peak location and height as well as the first four moments of the intensity distribution (ie, mean, standard deviation, skewness, and kurtosis) as CT features. Since the intensity distribution can become bimodal, we also added the Otsu threshold to reflect the bimodality and entropy to supplement standard deviation. Three exemplary CT features observed along the clinical course of the patient are shown in Figure 2D. Thus, Otsu threshold is an excellent predictor for severity based on visualization. We then analyzed all 178 CT records and determined the corresponding intensity distributions (Figure 3A and 3B). We found that distributions of severe and nonsevere cases were in direct contrast in terms of peak height and skewness. Principal component analysis also showed improved separation between the 2 groups (Figure 3C). Among the 8 CT features examined, peak location and entropy were not significantly related to severity, whereas all the other 6 CT features showed a statistically significant relation (Figure 3C). Mean and standard deviation, as well as skewness and kurtosis, were highly correlated; therefore, standard deviation and kurtosis were not considered as CT features.

The CT feature extraction enables quantitative prediction with signatures of both CT and laboratory features. We first analyzed the Spearman correlation between the CT and laboratory features (Figure 4A). Most features were not significantly correlated; however, lymphocyte, neutrophil, D-dimer, and platelet–large cell ratio showed good correlation with several CT features. Similarly, we used mutual information to select features to be used in the model. We used a random forest model with optimized hyperparameters to predict severity from CT and laboratory features. We selected a signature of 16 features from the feature number analysis (Figure 4B). The corresponding prediction model yielded a cross-validated AUROC score of 0.93 and an F₁ score of 0.81 (Figure 4C), which are considerably improved from the corresponding scores of the model with laboratory tests only (Figure 1D). The signature includes CT peak height, CT intensity mean, CT intensity skewness, CT Otsu threshold, lymphocyte percentage, gamma-glutamyl transpeptidase, LDH, C-reactive protein, white blood cell, D-dimer, cholinesterase, neutrophil percentage, hemoglobin, tricyclic antidepressant, albumin, and chloride.

Forecasting disease severity has significant clinical importance, as it allows clinicians to better prepare for treatment course. In addition to predicting severity based on CT and laboratory signatures, we also developed a statistical model to forecast severity from patient records upon admission when they were considered nonsevere. Although CT features are excellent predictors of severity, they are not as good for forecasting, yielding an AUROC of 0.68. In contrast, the random forest model based on laboratory tests yielded an AUROC of 0.81, indicating excellent forecasting predictability (Figure 5A). Other metrics considered for forecasting are presented in Table 2. This statistical model comprised 8 laboratory tests, among which lymphocyte and neutrophil counts (percentage) showed the highest fold change. In addition to comorbidity, we identified 8 laboratory tests that could be used for severity forecasting: individual counts of lymphocyte, neutrophil, monocyte, and eosinophil; red blood cell distribution width; hemoglobin; procalcitonin; and platelet–large cell ratio.