Baseline characteristics of participants
The study cohort consisted of COVID-19 patients admitted between December 30th, 2019 and April 17th, 2020, to 21 hospitals in Hubei, China that were designated to treat COVID-19 patients. There were 12,862 patients with COVID-19 who met the eligibility criteria for analysis. Among the study population, 3,254 (25.3%) received corticosteroid therapy (corticosteroid group) and 9,608 (74.7%) without corticosteroid therapy (non-corticosteroid group). The clinical characteristics of patients enrolled in the study were shown in Table 1. Participants receiving corticosteroid were older (61 [IQR, 49-69] years versus 57 [IQR, 45-67] years, P < 0.001) and had a higher number of males (56.2% versus 45.8%) than those in the non-corticosteroid group (Table 1). There was a higher prevalence of pre- existing hypertension (37.2% versus 32.0%, P < 0.001), diabetes mellitus (18.3% versus 15.3%, P < 0.001), chronic obstructive pulmonary disease (COPD;1.6% versus 1.0%, P = 0.005) and heart failure (0.9% versus 0.4%, P = 0.004) in the corticosteroid group than those without corticosteroid therapy (Table 1). The frequencies of neutrophil and leukocyte count increase, lymphocyte count decrease, C- reactive protein (CRP) increase, procalcitonin increase, creatine kinase (CK) increase, creatinine increase, and alanine transaminase (ALT) increase were significantly higher in the corticosteroid group than the non-corticosteroid group (Table 1).
Blood cells strongly associated with the high risk of mortality and corticosteroid treatment
Accumulated evidence from the clinical trial suggested favorable immune response modulation by low-to-moderate dose corticosteroids could be beneficial to patients with critical conditions17-19. However, the critical condition wasn’t exactly identified. To identify an indicator associated with disease severity and also can determine the therapeutic efficacy of corticosteroids in patients with COVID-19, we first identified blood cell parameter has the strongest association with 60-day in- hospital mortality in patients with COVID-19; After blood cell parameter was identified, we selected three candidate threshold values for the parameter with different capabilities to predict clinical outcome in patients with COVID-19; Finally, we further defined an appropriate severity threshold values for initiating corticosteroids therapeutic in patients with COVID-19.
In the first step, to determine indicators significantly associated with 60-day in-hospital mortality, ten parameters from complete blood cells were included in a multivariate logistic regression analysis. Using a multivariate logistic regression model on the dataset of 12,862 patients with COVID-19, the lymphocyte count decrease (OR, 5.17; 95%CI, 4.33-6.17; P < 0.001) and neutrophil count increase (OR, 5.00; 95%CI, 3.96-6.29; P < 0.001) were the top two factors significantly and positively associated with mortality (Table 2). Remarkedly, using LASSO regression, a regularization method creates parsimonious model, the circulating immune cells were key determinants associated with the mortality. Among immune cells, neutrophil and lymphocyte counts were also identified as the top two factors (Table 2). These findings were consistent with previous studies20,21 and indicated that systemic immune response was among the most critical factors related to the clinical outcomes in this cohort of patients with COVID-19.
To further estimate whether the levels of immune cells were also associated the use of corticosteroid in clinic, we conducted the same analyses on the entire cohort. Intriguingly, we found that neutrophil increase and lymphocyte decrease were also the top two covariates significantly associated with use of corticosteroid during hospitalization.
Considering neutrophil and lymphocyte counts indicated immune status and their significant association with clinical outcomes, we adapted an existing index neutrophil to lymphocyte ratio (NLR), a well-known index for patient's inflammatory status and risk of death22-24, and to further explore its potential as an indicator guided the use of corticosteroid. The C-statistic analysis showed continuous NLR has a high prognostic significance for death, with the area under of receiver-operating characteristic curve [AUROC] of 0.87 (95 % CI, 0.86-0.88) in the dataset with 12,862 patients with COVID-19 (Table 3).
NLR value as an indicator of clinical outcome in COVID-19
Since the NLR values exhibited a robust prognostic effect in our study population, we further investigated the candidate NLR threshold values in C-statistic analysis regarding their specificity and sensitivity in predicting mortality in all patients. When the model sensitivity (above 0.9) was considered over specificity, a threshold value of NLR at 3.13 demonstrated a high sensitivity (0.91), but a low specificity (0.62) and accuracy (0.64[95 % CI, 0.63-0.65]) in predicting mortality in our cohort of COVID-19 patients; In contrast, when the model specificity (above 0.9) were considered over sensitivity, a threshold value of NLR at 8.31 had the high specificity (0.91) ) and accuracy (0.89[95 % CI, 0.89-0.90]), but inadequate sensitivity (0.64). Finally, when specific and sensitivity were both considered, a threshold of NLR at 6.12 was identified with the highest Youden index while both good sensitivity, specificity, and accuracy were achieved (sensitivity, 0.75; specificity, 0.85; accuracy, 0.85) (Table 3).
Positive predictive values (PPV) and negative predictive values (NPV) for each of these three NLR threshold values indicated an imbalanced dataset in terms of predictive accuracy. Thus, balanced accuracies were calculated. The results showed a similar pattern that a threshold value of NLR 6.12 had the highest balanced accuracy for predicting clinical mortality in patients with COVID-19 (Table 3).
Primary outcomes based on NLR-dependent patient stratifications
To evaluate the efficiency of each NLR cutoff value in differentiating clinical benefits of corticosteroid therapy, we test the association between corticosteroid therapy and clinical outcomes across the different sub-cohorts in our patient population stratified based on NLR threshold values of 3.13, 6.12 and 8.31 as described earlier. The clinical characteristics of patients in the sub-cohorts stratified based on NLR values were listed in Table 4. Consistent with the observations made in the complete cohort, the patients treated with corticosteroid were more severe and had higher percentages of comorbidities than the patients not received corticosteroid treatment in each NLR based sub-cohorts.
We applied two models, i.e., the Cox time-varying model and the marginal structural model (MSM), to account for immortal time bias and time-varying confounders by indications plus immortal time bias, respectively, to estimate the associations between corticosteroid use and 60-day all-cause death in each NLR based sub-cohorts. The imbalanced confounders between the corticosteroid and the non-corticosteroid groups were adjusted.
When the study cohort was stratified using a cut-off value of NLR 3.13 into two sub-cohorts with NLR ≤ 3.13 and NLR > 3.13, the Cox time-varying model (adjusted HR, 0.62; 95%CI, 0.49-0.77; P < 0.001) showed a significantly lower risk of COVID-19 death in the corticosteroid group than the non- corticosteroid group in the sub-cohort of patients with NLR > 3.13 (Table 5). However, this association was not supported by the MSM model. In contrast, there were no significant associations between corticosteroid therapy with the risk of 60-day all-cause death in the sub-cohort of patients with an NLR ≤ 3.13 using either Cox time-varying or MSM model (Table 5). Therefore, NLR 3.13 were not a suitable threshold for differentiating clinical benefits of corticosteroid therapy in patients with COVID-19.
When the study cohort was stratified using an NLR cut-off value of 8.31, in the sub-cohort with NLR > 8.31 corticosteroid treatment was associated with a significantly lower incidence and risk of 60-day all-cause death compared to the non-corticosteroid users by both Cox time-varying model (adjusted HR, 0.55; 95%CI, 0.42-0.70; P < 0.001) and MSM (adjusted HR, 0.52; 95%CI, 0.38-0.70; P < 0.001). In the sub-cohort with NLR ≤ 8.31, corticosteroid treatment was not significantly associated with the risk of COVID-19 death using either the Cox time-varying model or MSM model (Table 5). These results suggested that COVID-19 patients with NLR >8.31showed significant benefit from corticosteroid treatment.
Finally, when patients were divided into two subgroups by an NLR at 6.12, a cutoff with the highest balanced accuracy, after adjusted for baseline and time-varying confounders, both Cox time- varying model (adjusted HR, 0.58; 95%CI, 0.46-0.73; P < 0.001) and MSM analysis (adjusted OR, 0.56; 95%CI, 0.42-0.73; P < 0.001) indicated that the in-hospital use of corticosteroid was associated with a lower risk of 60-day all-cause death compared to corticosteroid nonuse (Table 5). Once the NLR was lower than 6.12, there was no significant association between corticosteroid use with the risk of all-cause death (Table 5). Thus, the NLR at 6.12 not only showed optimal sensitivity and specificity to predict the risk of mortality, but also can stratify patients with the different therapeutic outcome from corticosteroid therapy.
These results demonstrated that the clinical benefits and therapeutic outcome of corticosteroid therapy for COVID-19 were significantly influenced by the patient’s immune status. Importantly, given the highest balanced accuracy for predicting mortality risk, the consistency observed from both models, and, in particular the remarkable differences in the opposite clinical outcome observed in the stratified sub-cohorts, a cut-off value of NLR at 6.12 appeared to be the optimal level for patient stratification for survival benefits from corticosteroid therapy.
Validating NLR cutoff at 6.12 in subgroup and sensitivity analyses
To further validate the performance of NLR based sub-cohort stratification using 6.12 cut-off value for the clinical benefits of corticosteroid treatment, we performed subgroup and sensitivity analyses using the following patient cohorts: (1) patients on mechanical ventilation with versus without corticosteroid therapy; (2) patients receiving a most commonly applied corticosteroid, methylprednisolone (methylprednisolone group), versus those not taking any forms of corticosteroids (non-methylprednisolone group); and (3) patients taking corticosteroids versus those not taking corticosteroids after patients from two randomly selected hospitals were removed. Both Cox time- varying and MSM models were conducted on those subgroups for sensitivity analyses.
We first analyzed the associations between corticosteroid use and the risk of death among the patients who received mechanical ventilation during hospitalization. From both Cox time-varying model and MSM analysis, corticosteroid therapy was consistently associated with a significantly reduced risk of 60-day mortality among patients with an NLR > 6.12 at admission with adjusted HR of 0.53 (95%CI, 0.36-0.78; P = 0.001) in the Cox time-varying model and adjusted OR of 0.59 (95% CI, 0.37-0.94; P < 0.001) (Table 6). Notably, among patients with NLR ≤ 6.12, corticosteroid therapy was not associated with any significant change in risk of 60-day all-cause death compared to no- corticosteroid treatment (Table 6).
Methylprednisolone is the most commonly prescribed corticosteroids in the study cohort. We further conducted subgroup analysis to compare the incidences and risks of 60-day all-cause death between patients taking methylprednisolone (methylprednisolone group) and those not taking any forms of corticosteroids (non-methylprednisolone group). Consistent with the earlier analysis using the entire cohort, for patients with NLR above 6.12, methylprednisolone therapy was associated with a significantly lower risk of 60-day all-cause death compared to the patients in the non- methylprednisolone group based on the Cox time-varying model (adjusted HR, 0.52; 95% CI, 0.39- 0.68; P < 0.001) and MSM (adjusted OR, 0.46; 95% CI, 0.31-0.69; P < 0.001) (Table 6). Again, no significant associations between methylprednisolone use and 60-day all-cause death risk were observed among patients with NLR ≤ 6.12 (Table 6).
In a sensitivity analysis by randomly removing two hospital sites, for patients with NLR > 6.12, corticosteroid treatment demonstrated a lower risk of 60-day all-cause death than no corticosteroid therapy from both the Cox time-varying model and MSM (Table 6). No significant associations between corticosteroid use and mortality risk were found in the patients with NLR ≤ 6.1 (Table 6). Thus, all these findings in subgroup and sensitivity analyses were consistent with our observations in the entire cohort, and further supported the accuracy and robustness of NLR cutoff at 6.12 to guide patient stratification for corticosteroid treatment.
Corticosteroid use in patients stratified based on NLR cutoffs
Among the 3,254 individuals who received corticosteroids, the major type of corticosteroids was methylprednisolone, which accounted for 97.1% of the corticosteroids treated patients, followed by prednisolone (10.7%) and hydrocortisone (1.0%) (Table 7). Methylprednisolone was also the most frequently administered corticosteroid in each subgroup of patients separated by different NLR cutoffs (Table 7).
The median duration of corticosteroids treatment was 8.0 (95%CI, 5.0-15.0) days. Notably, the initiated time of corticosteroid therapy was as early as at 1.0 (95%CI, 0.0-4.0) day after admission. The median level of the daily dosage was relatively low, at 40.0 (95%CI, 31.1-40.0) mg of methylprednisolone-equivalent dosage with the median accumulated dose at 320.0 (95%CI, 180.0- 580.0) mg (Table 7). Intriguingly, the accumulated dosages of corticosteroid in patients with an NLR above each cutoff were gradually increased along with the elevation of NLR values (Table 7). It is apparent that patients with an NLR above the cutoff had earlier initiating time, higher daily and accumulated doses, and longer corticosteroid treatment duration than those below the cutoffs (Table 7).
Adverse effects of corticosteroid treatment in patients with an NLR > 6.12
In previous clinical trials, the adverse effects observed from corticosteroids therapy included hyperglycemia, gastroduodenal bleeding, hypernatremia, and infection19,25. Here, we analyzed the incidences and risks of those potential adverse effects related to corticosteroid treatment in the NLR stratified sub-cohorts. Among the subgroup with an NLR > 6.12 at admission, corticosteroid treatment was associated with higher incidences of hyperglycemia requiring treatment (IRR, 2.33 [1.95-2.78]; IR, 0.79 versus 0.34; P < 0.001), infection requiring acceleration of antibiotics (IRR, 2.51 [2.18-2.90]; IR, 1.29 versus 0.51; P < 0.001) and fungal infection requiring antifungal treatment (IRR, 1.67 [1.21-2.31]; IR, 0.19 versus 0.11; P = 0.002) (Table 8). The higher risks of hyperglycemia requiring treatment and infection needing antibiotic acceleration in the corticosteroid treatment group were consistently observed in both the Cox time-varying model and the MSM (Table 8). Thus, despite a significantly lower risk of 60-day all-cause death in patients taking corticosteroid, the adverse effects of hyperglycemia and infection should be closely monitored in patients with NLR > 6.12.
Adverse effects of corticosteroid treatment in patients with an NLR ≤ 6.12
Among patients with an NLR ≤ 6.12, the incidences of corticosteroid-correlated adverse effects, including gastrointestinal hemorrhage (IRR, 2.62 [1.51-4.54]; IR, 0.02 versus 0.01; P < 0.001), hyperglycemia requiring treatment (IRR, 2.84 [2.47-3.27]; IR, 0.30 versus 0.10; P < 0.001), infection requiring acceleration of antibiotics (IRR, 4.63 [4.23-5.07]; IR, 0.88 versus 0.19; P < 0.001), fungal infection needing antifungal treatment (IRR, 3.82 [2.78-5.25]; IR, 0.07 versus 0.02; P < 0.001) and hypernatremia (IRR, 3.09 [2.27-4.20]; IR, 0.06 versus 0.02; P < 0.001), were all significantly more than the non-corticosteroid group (Table 8). The Cox time-varying analysis indicated that treatment of corticosteroid was associated with significantly increased risks of hyperglycemia and infection with adjusted HR of 3.15 (95% CI, 2.61-3.81; P < 0.001) for hyperglycemia requiring treatment, 2.72 (95% CI, 2.38-3.12; P < 0.001) for infection requiring acceleration for antibiotics. The elevated risks of hyperglycemia requiring treatment (adjusted OR, 2.37; 95% CI, 1.32-4.25; P = 0.004) was consistently observed in corticosteroid group using MSM (Table 8). Therefore, among patients with an NLR ≤ 6.12, there were no detectable survival benefits, but significantly increased adverse effects, in particular hyperglycemia associated with corticosteroid therapy.
Associations of corticosteroid therapy with outcomes in patients with diabetes
Given the significantly elevated risk of hyperglycemia by corticosteroid treatment and the high prevalence of pre-existing diabetes in patients with COVID-19, it is clinically important to estimate the influence of corticosteroid use in COVID-19 patients with pre-existing diabetes. This is becoming much more urgent and relevant since patients with diabetes are shown to be at remarkably higher risk of death after SARS-CoV-2 infection26, and that risk can be further exacerbated when the blood glucose is poorly controlled27.
In the current study cohort, 2,066 individuals had pre-existing diabetes, among which 596 cases (aged 65.0 [57.0-72.0] years; 59.2% males) received corticosteroid therapy and the other 1,470 cases (aged 64.0 [57.0-71.0] years; 51.0% males) did not receive corticosteroids. The occurrences of comorbidities in the corticosteroid and the non-corticosteroid groups in this diabetes cohort were comparable. Patients received corticosteroids exhibited more severe pathological conditions with higher frequencies of lymphocyte decrease and leukocyte count, neutrophil count, CRP, procalcitonin, D-dimer, and organ injury marker increases than those received no corticosteroids.
The incidence of 60-day all-cause death was significantly higher in the diabetic patients taking corticosteroids (IRR, 3.21 [2.47-4.18]; IR, 0.42 versus 0.13; P < 0.001) compared to those not taking corticosteroids (Table 9). When we addressed the time-varying confounders and imbalanced variables between the corticosteroid and non-corticosteroid groups, there were no significant associations of corticosteroid use with the risk of COVID-19 death in both Cox time-varying model and MSM model (Table 9).
Regarding the adverse effects, the crude incidences of hyperglycemia requiring treatment (IRR, 2.13 [1.86-2.43], P < 0.001), infection requiring acceleration of antibiotics (IRR, 4.33 [3.63-5.18], P< 0.001), fungal infection (IRR, 3.45 [2.28-5.23], P < 0.001), and hypernatremia (IRR, 2.40 [1.70- 3.39], P < 0.001) were significantly higher in the corticosteroid group than the non-corticosteroid group (Table 9). After adjusting for time-varying and basic imbalanced confounders, the Cox time- varying and MSM models consistently demonstrated a close association between corticosteroid use and significantly increased risks of hyperglycemia requiring treatment, infection requiring acceleration of antibiotics, needs for antifungal treatment and hypernatremia (Table 9). These findings indicated that, for patients with diabetes, corticosteroid use have more significant impacts on blood glycemic levels and concurrent infections in patients with diabetes compared to the general patients. Despite these side effects, our study did not demonstrate an association between use of corticosteroid and increase or reduction of 60-day all-cause death of COVID-19.