A total of 106 patients with ARDS caused by COVID-19 were included in this retrospective analysis and 53 (50%) of patients were defined as survivors. Characteristics in this cohort are summarized in Table 1. The majority were diagnosed with moderate to severe ARDS on admission (P/F ratio median [IQR]: 112 [87-148] mmHg). In total, 32 (30%) had a previous T2D diagnosis and 10 patients with HbA1c ≥ 6.5% (48 mmol/mol) had newly diagnosed T2D. We identified 58 (54%) patients with hyperglycemia (FPG >140mg/dl) on admission to the ICU, of which 33 (56%) had no prior diagnosis of T2D.
Utilizing a Cox-PH model (adjusted for age, sex and history of T2D) for FPG on admission as a continuous covariate, we replicated previous findings that high FPG on admission is a predictor of mortality. This model showed statistical significance for FPG and the model (HR: 1.00 [95% CI, 1.00-1.01], Fasting plasma glucose (mg/dl) (P)<0.001, LR(P)=0.002) and demonstrated a linear increase in HR with admission FPG. However, HbA1c did not show statistical significance in an equally adjusted model and additionally failed the proportional hazards assumption test.
Furthermore, we evaluated several established metrics of variability SD, CoefVar, MSSD, rMSSD and DGV of FPG as continuous parameters of FPG variability in multivariable Cox-PH models. While these models include age, sex and history of T2D as a covariate for adjustment, neither were statistically significant in any of the models. Out of these variability metrics, we selected the statistically significant metric with the highest C-Index and lowest AIC, where median DGV outperformed all the other contenders (Supplement Fig. 1).
We created an age and sex adjusted Cox-PH model (LR(P) <0.001) to demonstrate the change of HR by the range of DGV values. In this model, age and sex were not statistically significant (HR: 1.01 [95% CI, 0.99-1.04], P=0.3; HR: 0.94 [95% CI, 0.51-1.71], P=0.833, respectively) in contrast to DGV (HR: 1.02 [95% CI, 1.02-1.03], P<0.001). The model demonstrated a proportional increase in HR and at HR=1 DGV was 34.63mg/dl in males and 27.35 mg/dl in females (Fig. 1).
To determine an outcome-based cut-off for DGV we fitted a regression tree model (25.5 mg/dl) and compared it to a cut-off based on a hazard ratio of 1 derived from a Cox-PH model which was adjusted for age, sex and history of T2D (31 mg/dl). The regression tree-based cut-off demonstrated a higher AUC (0.729 vs. 0.689) in 30-day survivalROC curves, therefore we used the cut-off DGV value of 25.5 mg/dl in further models and testing rather than the Cox-PH based cut-off of DGV 31 mg/dl. In the group with DGV < 25.5mg/dl (low DGV group), significantly more patients (n=41 (69%)) survived to discharge compared to only 12 patients (25%) with DGV ≥ 25.5mg/dl (high DGV group) (Table 1). Furthermore, we analysed differences in clinical characteristics, comorbidities, laboratory parameters and medication on admission of patients in high and low DGV groups. There was no significant difference in age, gender, P/F ratio and likelihood of dialysis, ECMO and dexamethasone therapy among both groups. Similarly, there were no statistically significant differences in symptoms reported at or prior to admission between both groups. However, patients in high DGV group were significantly more obese (Median [IQR]: 28.4 [26.3—31.1] vs. 31.2 [27.5—38.1], P=0.01) and had a lower fever at admission to the ICU (Median [IQR]: 38.1 [37.3—38.6] vs. 37.6 [36.9—38.1], P<0.05). Comparing inflammatory markers on admission to the ICU, we found ferritin to be significantly lower in the high DGV group, whereas leukocytes, lymphocytes, CRP, procalcitonin and IL-6 remained not significantly different between the two groups (Table 1).
Additionally, Kaplan-Meier estimators showed a significantly (P<0.0001) longer median survival of patients in low DGV group of 87 days in comparison to patients in high DGV group, which had a median survival of 25 days (Fig. 2).
Based on these findings, we generated an unadjusted Cox-PH model for mortality in DGV (HR: 1.02, (P)<0.001, LR(P)<0.001). This model was still significant after adjusting for clinically predetermined confounders (HR: 1.016, (P)=0.001, LR(P)<0.001). This model demonstrated that median DGV remained independently associated with adverse outcome (Fig. 3).
In order to examine a potentially predictive use of median DGV, we calculated median DGV for the first three days of ICU admission. The results of an equally adjusted model (using the three-day median FPG instead of the median FPG over the entire ICU admission) indicated nonetheless significant association of three-day median DGV with adverse outcome (Supplement Fig. 2).