3.1 Patient characteristics
A total of 189 patients with COVID-19 with pneumonia were hospitalized, and 173 patients met the inclusion criteria (Figure 1). The clinical characteristics and laboratory findings are shown in Table 1 (development group=120, validation group=53). All patients were ethnically Asian. The 28-day mortality was 23.7% (41 of 173). The median age of these patients was 74 (range, 23-95 years). A total of 4.6% (8 of 173) of the total patients were younger than 40 years, while 87.9% (152 of 173) were above the age of 60 years. A total of 70.5% were males. The proportion of severe/critical patients was 57.8% (100/173) on admission. The 1st and 2nd most common underlying diseases in our study were hypertension (58.4%, 101/173) and diabetes mellitus (31.8%, 55/173), respectively.
We analysed the correlation between KL-6 level and RSG or CPV sign, and showed that KL-6 was positively correlated with RSG (r=0.734, P<0.001) and CPV (r=0.387, P<0.001) in the development group. The KL-6 level in patients with a high RSG (≥3) was higher than that in patients with a low RSG (<3) (872(577-1039) vs 318 (210-455), U/ml, P<0.001) in development group. And the KL-6 level in patients with CPV sign was higher than that in patients without CPV signs (591(398-905) vs 325(207-567), U/ml, P<0.001) in the development group.
3.2 Nomogram construction and calibration
The 6 variables were significantly correlated with 28-day mortality including disease severity, serum KL-6, P/F ratio, NLR, IL-6 and RSG in both the development group and validation group (P<0.05, Table 1). Univariate Cox regression analysis was performed, and showed that these variables significantly different (P<0.05, Table 2). Then, we further performed a multivariable Cox regression analysis based on KL-6 and RSG, which confirmed that KL-6 and RSG were independent predictors (Figure 2.1). A nomogram predictive of 28-day survival was developed based on KL-6 and RSG (Figure 2.2). The C-index of the nomogram in the development group and validation group was 0.790 (95% CI: 0.702-0.878) and 0.874 (95% CI: 0.798-0.950), respectively. The calibration curve showed that the nomogram performed well when compared to the actual results (Figure 2.3, 2.4). Findings from the decision curve analysis (DCA) showed that using the nomogram model to predict 28-day survival probability added more net benefit than either having all patients or none patients treated by this model, suggesting that the nomogram model was clinically useful (Figure 2.5).
3.3 Predictive performance of the KL-6, RSG and combined scores
Survival ROC curve analysis was applied to analyse the predictive performance of KL-6, RSG and the combined score (Figure 3.1). The optimal cut-offs and relevant sensitivity and specificity are listed in Table 3. The combined score in the development group yielded an AUC of 0.869 (95% CI: 0.776-0.942, P<0.001) with 76.9% sensitivity and 87.3% specificity, which indicated a good predictive value of the two biomarkers. The combined score in the validation group yielded an AUC of 0.932 (95% CI: 0.862-0.997, P<0.001) with 84.6% sensitivity and 87.5% specificity, suggesting that the model had good performance in predicting patient 28-day mortality.
On the basis of the optimal cut-off based on the ROC curve analysis, all indicators were transformed into binary variables, and risk ratios were calculated. The risk ratios of KL-6, RSG, and combined scores in the development group and validation group were associated with a higher incidence of 28-day mortality, with risk ratios of 17.858 (95% CI: 7.130-44.729, P<0.001), 7.279 (95% CI: 2.932-18.069, P<0.001), 12.107 (95% CI: 4.861-30.154, P<0.001), 18.826 (95% CI: 4.140-85.588, P<0.001), respectively.
3.4 Kaplan - Meier curve of subgroups stratified by KL-6, RSG and combined scores
All 173 inpatients were followed in this retrospective study. The primary endpoint occurred in 41 inpatients (23.7%). Patients were categorized into two subsets according to the optimal cut-off value of KL-6, RSG and combined score in the development group (Table 3). Kaplan-Meier curves showed that differences between the two subsets were statistically significant (Figure 3.2, 3.3, 3.4) in the development group. In the validation group, the model showed that patients with a high combined score had poor outcomes compared to those with a low combined score (Figure 3.5).