Incidence and Outcomes of Bloodstream Infections in Critically Ill Patients with COVID-19: A Nationwide Multicenter Study in South Korea

Abstract

Purpose

Coronavirus disease 2019 (COVID-19) increases the risk of bloodstream infections (BSI), but associated risk factors remain incompletely characterized. We aimed to evaluate BSI risk factors in patients admitted to the intensive care unit for COVID-19.

Materials and Methods

This nationwide multicenter study, conducted across 22 university-affiliated hospitals in South Korea, included COVID-19 patients treated with high-flow nasal cannula (HFNC) therapy or mechanical ventilation. Baseline characteristics were compared between patients who developed BSI within 30 days and those who did not. Pathogens were analyzed, risk factors were evaluated using multivariable logistic regression, and supplementary Cox proportional hazards analyses were performed to address time-dependent outcomes.

Results

Of 573 included patients, 206 (35.9%) developed at least one BSI episode. Staphylococcus species were the most frequent isolates. Gram-negative bacteria became increasingly prevalent over time, while fungal infections were associated with prolonged hospitalization. Multivariable analysis identified chronic neurological disease, higher baseline Sequential Organ Failure Assessment score, mechanical ventilation, prone positioning, and extracorporeal membrane oxygenation (ECMO) use prior to infection as independent BSI predictors. Notably, although crude in-hospital mortality was significantly higher in the BSI group (41.5% vs. 31.7%, p=0.012), BSI was not an independent predictor of in-hospital mortality in either logistic regression or Cox analyses.

Conclusion

In critically ill patients with COVID-19, high baseline severity, chronic neurological disease, and pre-infection advanced life support (mechanical ventilation and ECMO) were independent predictors of BSI. Rather than acting as an independent cause of death, BSI primarily serves as a clinical marker of extreme illness severity and prolonged invasive interventions.

Graphical Abstract

INTRODUCTION

The coronavirus disease 2019 (COVID-19) pandemic has underscored the substantial burden of secondary infections in critically ill patients, particularly those requiring intensive care unit (ICU) admission.¹ Bloodstream infections (BSIs) are serious complications in this population, strongly associated with prolonged hospitalization, increased antimicrobial resistance, and higher crude mortality.¹,² Early studies during the pandemic highlighted gram-negative pathogens as the predominant etiological agents of BSIs in COVID-19 ICUs, consistent with prepandemic ICU infection patterns.³ However, emerging multicenter evidence suggests a dynamic epidemiological shift, with gram-positive cocci, including Staphylococcus spp. and Enterococcus faecium, increasingly implicated in early-onset BSIs.⁴,⁵

Previous studies have reported BSI incidence rates of 10%–50% among ICU patients with COVID-19, with mortality exceeding 40% in affected individuals.⁶,⁷ While traditional risk factors—such as mechanical ventilation (MV), renal replacement therapy, and corticosteroid use—are well-recognized in general critical care settings,¹,⁸,⁹ their specific roles and temporal dynamics within the unique immunopathological landscape of COVID-19 remain to be thoroughly elucidated. Furthermore, previous studies have largely focused on gram-negative pathogens, leaving gaps in knowledge about temporal shifts in pathogen distribution and modifiable risk factors unique to COVID-19 critical care.¹⁰

A key unresolved question is whether the distinct immunopathological profile of severe COVID-19—characterized by prolonged lymphopenia, corticosteroid-induced immunosuppression, and mucosal barrier injury—predisposes patients to unique microbial patterns.¹,¹¹,¹²

Despite prior insights, important knowledge gaps remain. First, the temporal evolution of BSI pathogens during different phases of ICU admission has not been adequately characterized, limiting the development of time-sensitive antimicrobial strategies. Second, the impact of chronic comorbidities, particularly neurological disorders, on BSI susceptibility in patients with COVID-19 remains underexplored, despite their known associations with immune dysregulation and aspiration risk.

To address these gaps, we conducted a nationwide, multicenter, retrospective cohort study of 1755 critically ill patients with COVID-19. Using comprehensive microbiological and clinical data, our objectives were to 1) characterize temporal changes in the distribution of BSI pathogens during ICU admission and 2) identify modifiable risk factors associated with early- versus late-onset BSIs. We hypothesized that the interplay between prolonged immunomodulatory therapy, cumulative invasive procedures, and microbiota disruption in severe COVID-19 would contribute to the predominance of gram-positive and fungal pathogens in later stages of ICU stay, with MV and renal replacement therapy acting as key drivers of microbial selection.

MATERIALS AND METHODS

Study population

This nationwide multicenter, retrospective observational cohort study in South Korea utilized data from patients with COVID-19 admitted to the ICUs of 22 university-affiliated hospitals between January 2020 and August 2021. Eligible participants were hospitalized adults (age ≥19 years) with polymerase chain reaction-confirmed severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2) infection who received high-flow nasal cannula (HFNC) oxygen therapy, MV, or extracorporeal membrane oxygenation (ECMO). Patients were excluded if they were admitted only for monitoring, received low-flow oxygen therapy, or lacked sufficient blood culture data. Since the participating study sites act as national tertiary referral centers for severe COVID-19, our cohort included patients transferred from other facilities. Specifically, 13.1% (75/573) of the included patients were directly transferred from outside healthcare facilities, including other acute care hospitals, convalescent hospitals, and nursing homes, with several already on MV or ECMO.

Data collection and definitions

Trained coordinators at each center collected the following data: 1) demographics [age, sex, body mass index (BMI)], laboratory findings, comorbidities, Clinical Frailty Scale, COVID-19 vaccination status, timing of COVID-19 diagnosis, and Sequential Organ Failure Assessment (SOFA) score at admission; 2) initial corticosteroid dose and duration of therapy; 3) ICU treatments and information regarding the use of rescue therapies, including remdesivir, tocilizumab, vasopressors, ECMO, and continuous renal replacement therapy (CRRT); 4) cultured pathogens during hospitalization; and 5) clinical outcomes, including in-hospital mortality and hospital/ICU length of stay (LOS). Immunocompromised status was defined as HIV infection with a CD4 count <200 cells/mm³, neutropenia with an absolute neutrophil count <1000 cells/mm³, or the use of immunosuppressive therapy after solid organ transplantation. Chronic neurological disease was defined as the presence of stroke, neuromuscular disease, epilepsy, or movement disorders. BSI was defined as a positive blood culture for one or more bacterial or fungal organisms. Two investigators reviewed cultures to determine whether isolates were clinically significant or contaminant. Isolates from blood cultures were excluded if they were considered contaminants and not the true causes of BSI based on clinical and microbiological data, following CDC criteria. The initial steroid dose was standardized to a dexamethasone-equivalent dose. Extended corticosteroid use was defined as a cumulative treatment duration of more than 10 days, exceeding the standard course established by the RECOVERY trial.¹³

Statistical analysis

Continuous variables were presented as means±standard deviations for normally distributed data, and as medians [interquartile ranges (IQRs)] for non-normally distributed data. Categorical variables were presented as frequencies and percentages. Comparisons between groups were performed using the Student’s t-test or Mann–Whitney U test for continuous variables, as appropriate, and the χ² test or Fisher’s exact test for categorical variables. Multivariable logistic regression was performed to examine associations between patient characteristics and BSI occurrence. Variables for the multivariable models were selected based on clinical relevance and a univariable significance level of p<0.1. The model included clinically relevant variables [age, sex, BMI, presence of chronic neurological disease, length of hospital stays prior to ICU admission, Clinical Frailty Scale score (prior to admission), prolonged dexamethasone use, SOFA score, tocilizumab use, use of MV, prone positioning, ECMO use, CRRT use, and tracheostomy]. A backward stepwise selection process was used to finalize the model while retaining clinically important variables. A separate multivariable logistic regression model was used to identify risk factors for in-hospital mortality. Similarly, variables with p-values<0.1 in the univariable analysis and those with established clinical importance were included. Included variables were age, sex, BMI, presence of chronic kidney disease, presence of chronic neurological disease, presence of solid malignant tumors, BSI, length of hospital stays prior to ICU admission, Clinical Frailty Scale score (prior to admission), prolonged dexamethasone use, SOFA score, tocilizumab use, use of MV, prone positioning, ECMO, CRRT, and tracheostomy. We used the Hosmer–Lemeshow test to evaluate the goodness-of-fit of the logistic regression model. The test results indicated an adequate fit between the model and the observed data. Goodness-of-fit was computed to assess the relevance of the logistic regression model. Odds ratios (ORs) and the corresponding 95% confidence intervals (CIs) were calculated. To prevent time-dependent exposure bias and reverse causation in identifying predictors for BSI, strict temporal criteria were applied to time-varying interventions (e.g., MV, ECMO, CRRT, and tracheostomy). Specifically, these life-supporting interventions were considered as pre-existing risk factors only if they were initiated strictly prior to the onset of BSI, which was defined as the exact date of the first positive blood culture. Interventions initiated on or after the day of BSI diagnosis were reclassified as non-exposures for the multivariable logistic regression model to ensure chronological plausibility. Furthermore, to assess potential multicollinearity among predictors, variance inflation factors (VIFs) were calculated for all variables in the multivariable models. Sequential modeling was also performed to demonstrate how the inclusion of advanced intervention variables influenced the independent predictive power of the baseline SOFA score.

To further address the time-dependent nature of the outcomes, supplementary Cox proportional hazards (PH) analyses were performed using the identical covariate sets as in the primary logistic regression models. For BSI risk (Model 1), time-to-BSI was modeled as the outcome, with patients who did not develop BSI censored at hospital discharge. The same pre-BSI temporal adjustment was applied to time-varying interventions. For in-hospital mortality (Model 2), BSI was incorporated as a time-varying covariate using the counting process formulation: each patient contributed a pre-BSI interval (BSI=0) and, upon BSI onset, a post-BSI interval (BSI=1), thereby accounting for immortal time bias and the dynamic temporal relationship between BSI and death. Robust standard errors were used in Model 2 by clustering on individual patients. Hazard ratios (HRs) and the corresponding 95% CIs were calculated, and model discrimination was assessed using Harrell’s concordance index (C-index).

The probabilities of in-hospital survival for each group were estimated using the Kaplan–Meier method and compared using the log-rank test. All tests were two-sided, and p-values<0.05 were considered statistically significant. All statistical analyses were performed using SPSS software (IBM, Armonk, NY, USA).

Ethics statement

The Severance Hospital Institutional Review Board approved the study protocol (IRB No. 4-2022-0019). The study complied with the principles of the Declaration of Helsinki and Good Clinical Practice Guidelines. The need for informed consent was waived due to the retrospective study design.

RESULTS

Baseline characteristics

The registry initially included 1755 adult patients with confirmed SARS-CoV-2 infection by polymerase chain reaction who were admitted to the ICU. Of these, 641 patients were excluded as they received only low-flow oxygen therapy and basic monitoring, and 541 were excluded due to insufficient blood culture data. In total, 573 patients were enrolled in this study (Fig. 1). The baseline characteristics of the patients are presented in Table 1. During hospitalization, BSI occurred in 206 patients (35.9%). Among comorbidities, the proportion of patients with chronic neurological diseases (stroke, neuromuscular disease, epilepsy, or movement disorder) was higher in the BSI group than in the non-BSI group (20.3% vs. 12.6%, p=0.014). The Clinical Frailty Scale score (after discharge) was also higher in the BSI group (5.3±1.9 vs. 4.4±2.0, p<0.001). Although there were no differences in the length of hospital stay before ICU admission (0.0 [0.0–0.0] days vs. 0.0 [0.0–0.0] days, p=0.370) or SOFA score between the groups, both ICU LOS (34.0 [17.2–50.8] days vs. 18.0 [12.0–31.0] days, p<0.001) and hospital LOS (40.0 [24.0–66.0] days vs. 25.0 [17.0–45.0] days, p<0.001) were significantly longer in the BSI group. Furthermore, both ICU mortality (40.3% vs. 26.7%, p<0.001) and in-hospital mortality (41.5% vs. 31.7%, p=0.012) were higher in the BSI group.

Fig. 1. Studyflow diagram.

Table 1. Baseline Characteristics.

Variables		BSI (n=206)	Non-BSI (n=367)	p
Age (yr)		68.8±12.7	70.3±13.1	0.182
Sex				0.302
	Male	120 (58.3)	206 (56.3)
	Female	86 (41.7)	160 (43.7)
HTN		119 (57.5)	221 (60.4)	0.498
DM		77 (37.2)	135 (36.9)	0.941
Cardiovascular disease		31 (15.0)	44 (12.0)	0.314
Chronic lung disease		13 (6.3)	30 (8.2)	0.403
Chronic neurological disease		42 (20.3)	46 (12.6)	0.014
Chronic kidney disease		23 (11.1)	34 (9.3)	0.484
Chronic liver disease		5 (2.4)	11 (3.0)	0.680
Immunocompromised*		6 (2.9)	9 (2.5)	0.752
Connective tissue disease		2 (1.0)	5 (1.4)	0.675
Hematological malignancies		3 (1.4)	6 (1.6)	0.860
Solid malignant tumors		11 (5.3)	31 (8.5)	0.164
Length of hospital stay before ICU admission (days)		0.0 [0.0–0.0]	0.0 [0.0–0.0]	0.370
Clinical Frailty Scale score (before admission)		3.2±1.6	3.3±1.9	0.613
Clinical Frailty Scale score (after discharge)		5.3±1.9	4.4±2.0	<0.001
SOFA score		4.1±2.1	3.9±2.4	0.366
Outcomes
	ICU length of stay (days)	34.0 [17.2–50.8]	18.0 [12.0–31.0]	<0.001
	ICU mortality	83 (40.3)	97 (26.7)	<0.001
	In-hospital length of stay (days)	40.0 [24.0–66.0]	25.0 [17.0–45.0]	<0.001
	In-hospital mortality	86 (41.5)	116 (31.7)	0.012
Treatment
	Remdesivir	140 (67.6)	273 (74.6)	0.081
	Duration of remdesivir use (days)	5.0 [5.0–5.0]	5.0 [5.0–5.0]	0.793
	Dexamethasone	203 (98.1)	349 (93.4)	0.110
	Duration of dexamethasone use (days)	21.0 [10.0–34.5]	16.0 [10.0–29.0]	0.063
	Tocilizumab	15 (7.2)	39 (10.7)	0.233
Life supporting interventions
	HFNC	19 (9.2)	124 (33.9)	<0.001
	Duration of HFNC (days)	1.0 [0.0–5.0]	4.0 [1.0–9.0]	<0.001
	MV	127 (61.4)	195 (53.3)	0.061
	Duration of MV (days)	30.0 [14.0–48.5]	14.0 [8.0–33.0]	<0.001
	Prone positioning	69 (33.3)	88 (24.0)	0.019
	Duration of prone positioning (days)	3.0 [2.0–9.0]	4.0 [2.0–7.0]	0.701
	ECMO	61 (29.5)	47 (12.8)	<0.001
	Duration of ECMO (days)	27.5 [17.8–46.2]	18.0 [10.0–38.0]	0.049
	NIV	13 (6.3)	38 (10.4)	0.126
	Inhaled NO	17 (8.2)	14 (3.8)	0.034
	Hemoperfusion	2 (1.0)	4 (1.1)	>0.999
	CRRT	60 (29.0)	48 (13.1)	<0.001
Tracheostomy		101 (48.8)	87 (23.9)	<0.001

BSI, bloodstream infection; HTN, hypertension; DM, diabetes mellitus; ICU, intensive care unit; SOFA, Sequential Organ Failure Assessment; HFNC, high-flow nasal cannula; MV, mechanical ventilation; ECMO, extracorporeal membrane oxygenation; NIV, non-invasive ventilation; CRRT, continuous renal replacement therapy.

Continuous variables are presented as mean±standard deviation or median [interquartile range]. Categorical variables are presented as numbers (%).

^*Immunocompromised was defined as: HIV infection (CD4 <200 cells/mm³), neutropenia (neutrophils <1000 cells/mm³), or ongoing immunosuppressive therapy following solid organ transplantation.

Compared with the non-BSI group, a smaller proportion of patients in the BSI group required HFNC therapy alone (9.2% vs. 33.9%, p<0.001), and the duration of HFNC use was significantly shorter (1.0 [0.0–5.0] days vs. 4.0 [1.0–9.0] days, p<0.001). Conversely, higher proportions of patients in the BSI group underwent prone positioning (33.3% vs. 24.0%, p=0.019), ECMO (29.5% vs. 12.8%, p<0.001), CRRT (29.0% vs. 13.1%, p<0.001), and tracheostomy (48.8% vs. 23.9%, p<0.001). The durations of MV (30.0 [14.0–48.5] days vs. 14.0 [8.0–33.0] days, p<0.001) and ECMO (27.5 [17.8–46.2] days vs. 18.0 [10.0–38.0] days, p=0.049) were also significantly longer in the BSI group. BSIs were identified in 37 patients within 2 days and 48 patients within 7 days. The baseline characteristics of patients with BSI within 2 and 7 days are presented in Supplementary Tables 1 and 2 (only online).

BSI pathogens

Table 2 details the distribution of confirmed causative pathogens according to length of hospital stay. Among the 37 cases that occurred within 2 days, 28 (75.7%) were due to gram-positive cocci, 7 (18.9%) to gram-negative bacteria, and 2 (5.4%) to fungi. The most frequently identified gram-positive cocci were Staphylococcus species (17/37; 45.9%), while Acinetobacter species were the most common gram-negative bacteria (2/37; 5.4%). All fungal pathogens were Candida species.

Table 2. Identified Pathogens of BSI.

			≤2 days	3–7 days	8–30 days
Patients at risk in ICU (denominator)			n=573	n=376	n=238
Total BSI events			37 (6.5)	27 (7.2)	161 (67.6)
Bacteria			35 (94.6)	26 (96.3)	139 (86.4)
	Gram-positive cocci		28 (75.7)	15 (55.6)	91 (56.2)
		Staphylococcus species	17 (45.9)	11 (40.7)	61 (37.7)
		Staphylococcus aureus	1 (2.7)	3 (11.1)	5 (3.1)
		Streptococcus species	2 (5.4)	0 (0)	9 (5.6)
		Enterococcus species	2 (5.4)	2 (7.4)	19 (11.7)
	Gram-negative bacteria		7 (18.9)	11 (40.7)	48 (29.6)
		Acinetobacter	2 (5.4)	0 (0)	15 (9.3)
		Escherichia coli	1 (2.7)	3 (11.1)	5 (3.1)
		Klebsiella pneumoniae	1 (2.7)	4 (14.8)	17 (10.5)
Fungus			2 (5.4)	1 (3.7)	22 (13.6)
		Candida species	2 (5.4)	1 (3.7)	22 (13.6)

BSI, bloodstream infection; ICU, intensive care unit.

The denominator for each timeframe represents the number of patients remaining in the ICU at risk of a first BSI episode. Patients who died, were discharged from the ICU, or experienced a prior BSI before the start of the timeframe were excluded from the subsequent at-risk population.

Within 3–7 days after admission, 27 BSI cases occurred: 15 (55.6%) were caused by gram-positive cocci, 11 (40.7%) by gram-negative bacteria, and 1 (3.7%) by fungi—all Candida species. Overall, Staphylococcus spp. were the most common gram-positive organisms (11/27; 40.7%), whereas Klebsiella pneumoniae was the most frequent gram-negative organism (4/27; 14.8%).

In total, 161 BSI cases were identified between days 8 and 30. Among these, gram-positive cocci accounted for 91 cases (56.2%), gram-negative bacteria for 48 cases (29.6%), and fungi (Candida species) for 22 cases (13.6%). Among gram-positive cocci, Staphylococcus species were most common (61 cases; 37.7%), followed by Enterococcus species (19 cases; 11.7%), and Streptococcus species (5 cases; 3.1%). Among gram-negative bacteria, Klebsiella species were most frequent (17 cases; 10.5%), followed by Acinetobacter species (15 cases; 9.3%), and Escherichia coli (5 cases; 3.1%). These findings demonstrate that gram-positive organisms, particularly Staphylococcus spp., predominate in the early phase of hospitalization (≤2 days). From 3 to 7 days after admission, the proportion of gram-negative bacteria, especially K. pneumoniae, increased markedly. In patients hospitalized for longer durations (8–30 days), Staphylococcus spp. remained the most frequent gram-positive organisms, whereas Klebsiella spp. and Acinetobacter spp. predominated among gram-negative bacteria. Moreover, the proportion of fungal infections, predominantly Candida spp., increased substantially with prolonged hospitalization.

Risk factors for BSI

Risk factors for BSI were evaluated using multivariable logistic regression analysis (Table 3). To prevent time-dependent exposure bias and reverse causation, life-supporting interventions (e.g., MV, ECMO, CRRT, prone positioning, and tracheostomy) were included as exposures only if they were initiated strictly prior to the onset of BSI. The model included the following clinically relevant variables: age, sex, BMI, presence of chronic neurological disease, length of hospital stays before ICU admission, Clinical Frailty Scale score (before admission), extended dexamethasone use, SOFA score, tocilizumab use, use of MV, prone positioning, ECMO, CRRT, and tracheostomy. Independent risk factors for BSI were a history of chronic neurological disease (adjusted OR, 2.36; 95% CI, 1.35–4.13; p=0.003), higher SOFA score (adjusted OR, 1.11; 95% CI, 1.05–1.18; p<0.001), prone positioning (adjusted OR, 1.59; 95% CI, 1.06–2.39; p=0.025), ECMO (adjusted OR, 2.01; 95% CI, 1.14–3.53; p=0.016), and MV (adjusted OR, 2.61; 95% CI, 1.37–4.99; p=0.004). In contrast, tracheostomy and CRRT were not independently associated with BSI after adjustment for the exact timing of the interventions.

Table 3. Multivariable Logistic Regression Analysis of Risk Factors for BSI.

Variables	Univariable		Multivariable
	OR (95% CI)	p	OR (95% CI)	p
Age	0.99 (0.98–1.00)	0.182	1.00 (0.98–1.01)	0.606
Sex, male	1.07 (0.76–1.51)	0.695	0.95 (0.66–1.37)	0.782
BMI	1.01 (0.97–1.05)	0.573	1.00 (0.96–1.05)	0.918
Chronic neurological disease	1.77 (1.12–2.80)	0.015	2.36 (1.35–4.13)	0.003
Length of hospital stay before ICU admission (days)	0.98 (0.89–1.07)	0.612	1.01 (0.97–1.04)	0.647
Clinical Frailty Scale score	1.00 (0.97–1.04)	0.821	0.88 (0.78–1.01)	0.068
Dexamethasone (extended)	1.29 (0.88–1.89)	0.200	1.21 (0.71–2.04)	0.484
Tocilizumab	0.66 (0.35–1.22)	0.182	0.64 (0.28–1.42)	0.271
SOFA	1.04 (0.95–1.14)	<0.001	1.11 (1.05–1.18)	<0.001
MV (MV only)	2.59 (1.39–4.95)	0.004	2.61 (1.37–4.99)	0.004
Prone positioning	1.58 (1.09–2.30)	0.017	1.59 (1.06–2.39)	0.025
ECMO	1.68 (1.06–2.65)	0.027	2.01 (1.14–3.53)	0.016
Tracheostomy	0.99 (0.67–1.48)	0.979	0.85 (0.53–1.35)	0.493
CRRT	1.04 (0.63–1.71)	0.889	0.67 (0.38–1.20)	0.177

BSI, bloodstream infection; OR, odds ratio; CI, confidence interval; BMI, body mass index; ICU, intensive care unit; SOFA, Sequential Organ Failure Assessment; MV, mechanical ventilation; ECMO, extracorporeal membrane oxygenation; CRRT, continuous renal replacement therapy.

To further validate these findings using a time-to-event approach, a supplementary Cox PH analysis was performed with the identical covariate set (Supplementary Table 3, only online). Chronic neurological disease (adjusted HR, 1.81; 95% CI, 1.18–2.77; p=0.006) and higher SOFA score (adjusted HR, 1.10; 95% CI, 1.05–1.15; p<0.001) remained independent predictors of time-to-BSI, fully consistent with the primary logistic regression results. Notably, MV (adjusted HR, 0.82; 95% CI, 0.57–1.16; p=0.258), ECMO (adjusted HR, 0.96; 95% CI, 0.59–1.56; p=0.871), and prone positioning (adjusted HR, 0.71; 95% CI, 0.49–1.04; p=0.078) did not retain independent significance in the Cox model. This attenuation likely reflects the fact that patients requiring these advanced interventions had inherently longer hospitalizations, during which BSI risk accumulates, and the Cox model appropriately accounts for this differential time-at-risk. Chronic neurological disease and SOFA score thus represent robust, time-independent predictors of BSI regardless of the analytical approach.

Risk factors for BSI within 2 and 7 days are presented in Supplementary Tables 4 and 5 (only online). After strict adjustment for the timing of interventions, within 2 days of admission, independent risk factors were higher SOFA score (adjusted OR, 1.44; 95% CI, 1.10–1.87; p=0.007), prone positioning (adjusted OR, 5.92; 95% CI, 2.44–13.97; p<0.001), and a lower Clinical Frailty Scale score (adjusted OR, 0.76; 95% CI, 0.59–0.97; p=0.028), while a history of chronic neurological disease showed a borderline trend (adjusted OR, 2.61; 95% CI, 0.99–6.95; p=0.056).

Risk factors for in-hospital mortality

Table 4 shows the results of the analysis of clinical characteristics by in-hospital mortality. Compared to the survivor group, the non-survivor group had a significantly higher mean age (73.2±11.9 years vs. 67.9±13.1 years, p<0.001) and higher proportions of patients with chronic neurological disease (20.8% vs. 12.4%, p=0.008) and chronic kidney disease (15.3% vs. 7.0%, p=0.001). In the non-survivor group, almost all patients received dexamethasone (99.0% vs. 94.9%; p=0.012), although the duration of use was significantly shorter (16.0 [10.0–29.0] days vs. 19.0 [10.0–33.0] days, p=0.008). The duration of HFNC use was also significantly shorter in non-survivors than in survivors (2.0 [0.0–6.0] days vs. 4.0 [1.0–9.8] days, p=0.016). However, more than 90% of patients in the non-survivor group required escalation to advanced respiratory support beyond HFNC (HFNC only: 9.9% vs. 33.2%; p<0.001). The use of prone positioning was comparable between groups; however, its duration was significantly longer in the non-survivor group (4.5 [3.0–9.0] days vs. 3.0 [1.0–6.0] days, p=0.001). The proportion of patients who underwent ECMO was also higher in non-survivors (32.7% vs. 11.3%; p<0.001). The use of inhaled nitric oxide was significantly higher in the non-survivor group compared to the survivor group (9.9% vs. 3.0%, p<0.001). In addition, tracheostomy (43.8% vs. 27.0%, p<0.001), CRRT (41.1% vs. 6.7%, p<0.001), and BSI (42.6% vs. 32.6%, p=0.018) were all significantly more common in non-survivors.

Table 4. Clinical Characteristics According to In-Hospital Mortality.

Variables		Survivors (n=371)	Non-survivors (n=202)	p
Age (yr)		67.9±13.1	73.2±11.9	<0.001
Sex				0.283
	Male	205 (55.3)	121 (59.9)
	Female	166 (44.7)	81 (40.1)
HTN		213 (57.4)	127 (62.9)	0.204
DM		133 (35.8)	79 (39.1)	0.440
Cardiovascular disease		41 (11.1)	34 (16.8)	0.050
Chronic lung disease		23 (6.2)	20 (9.9)	0.108
Chronic neurological disease		46 (12.4)	42 (20.8)	0.008
Chronic kidney disease		26 (7.0)	31 (15.3)	0.001
Chronic liver disease		9 (2.4)	7 (3.5)	0.596
Immunocompromised*		8 (2.2)	7 (3.5)	0.348
Connective tissue disease		5 (1.3)	2 (1.0)	0.710
Hematological malignancies		4 (1.1)	5 (2.5)	0.290
Solid malignant tumors		20 (5.4)	22 (10.9)	0.019
Length of hospital stay before ICU admission (days)		0.0 [0.0–0.0]	0.0 [0.0–0.0]	0.239
Clinical Frailty Scale score (before admission)		3.1±1.7	3.5±1.9	0.014
SOFA score		3.6±2.0	4.6±2.7	<0.001
Bloodstream infections		121 (32.6)	86 (42.6)	0.018
Outcomes
	ICU length of stay (days)	21.0 [13.0–38.0]	26.0 [15.0–45.0]	0.046
	ICU mortality	0	180 (90.0)	<0.001
	In-hospital length of stay (days)	29.5 [20.0–58.8]	28.5 [17.0–47.0]	0.014
Treatment
	Remdesivir	273 (73.6)	140 (69.3)	0.275
	Duration of remdesivir use (days)	5.0 [5.0–5.0]	5.0 [5.0–5.0]	0.188
	Dexamethasone	352 (94.9)	200 (99.0)	0.012
	Duration of dexamethasone use (days)	19.0 [10.0–33.0]	16.0 [10.0–29.0]	0.008
	Tocilizumab	39 (10.5)	15 (7.4)	0.227
Life supporting interventions
	HFNC therapy	123 (33.2)	20 (9.9)	<0.001
	Duration of HFNC (days)	4.0 [1.0–9.8]	2.0 [0.0–6.0]	0.016
	MV	206 (55.5)	116 (57.4)	0.661
	Duration of MV (days)	15.0 [8.0–37.0]	26.0 [13.0–45.0]	0.014
	Prone positioning	95 (25.6)	62 (30.7)	0.192
	Duration of prone positioning (days)	3.0 [1.0–6.0]	4.5 [3.0–9.0]	0.001
	ECMO	42 (11.3)	66 (32.7)	<0.001
	Duration of ECMO (days)	19.0 [10.0–33.0]	27.5 [12.2–47.0]	0.032
	NIV	34 (9.2)	17 (8.4)	0.878
	Inhaled NO	11 (3.0)	20 (9.9)	<0.001
	Hemoperfusion	3 (0.8)	3 (1.5)	0.430
	CRRT	25 (6.7)	83 (41.1)	<0.001
Tracheostomy		100 (27.0)	88 (43.8)	<0.001

HTN, hypertension; DM, diabetes mellitus; ICU, intensive care unit; SOFA, Sequential Organ Failure Assessment; HFNC, high-flow nasal cannula; MV, mechanical ventilation; ECMO, extracorporeal membrane oxygenation; NIV, non-invasive ventilation; CRRT, continuous renal replacement therapy.

Continuous variables are presented as mean±standard deviation or median [interquartile range]. Categorical variables are presented as numbers (%).

^*Immunocompromised was defined as: HIV infection (CD4 <200 cells/mm³), neutropenia (neutrophils <1000 cells/mm³), or ongoing immunosuppressive therapy following solid organ transplantation.

Multivariable logistic regression analysis to identify risk factors for in-hospital mortality is presented in Table 5. The model included clinically relevant variables such as age, sex, BMI, presence of chronic kidney disease, presence of chronic neurological disease, presence of solid malignant tumors, BSIs, length of hospital stay prior to ICU admission, Clinical Frailty Scale score (prior to admission), prolonged dexamethasone use, SOFA score, tocilizumab use, use of MV, prone positioning, ECMO, CRRT, and tracheostomy. Independent risk factors for in-hospital mortality were age (adjusted OR, 1.06; 95% CI, 1.04–1.08; p<0.001), history of solid malignant tumors (adjusted OR, 3.14; 95% CI, 1.18–8.37; p=0.022), use of MV (adjusted OR, 3.69; 95% CI, 1.95–6.97; p<0.001), ECMO (adjusted OR, 4.34; 95% CI, 2.35–8.01; p<0.001), and CRRT (adjusted OR, 6.55; 95% CI, 3.59–11.98; p<0.001).

Table 5. Multivariable Logistic Regression Analysis of In-Hospital Mortality.

Variables	Univariable		Multivariable
	OR (95% CI)	p	OR (95% CI)	p
Age	1.04 (1.02–1.05)	<0.001	1.06 (1.04–1.08)	<0.001
Sex, male	1.21 (0.85–1.71)	0.284	1.19 (0.77–1.84)	0.428
BMI	1.01 (0.97–1.05)	0.729	1.01 (0.96–1.06)	0.682
Chronic kidney disease	2.41 (1.38–4.18)	0.002	0.89 (0.42–1.89)	0.761
Chronic neurological disease	1.85 (1.17–2.94)	0.008	1.64 (0.87–3.10)	0.125
Solid malignant tumors	1.00 (0.97–1.04)	0.821	3.14 (1.18–8.37)	0.022
Bloodstream infections	1.53 (1.08–2.18)	0.018	0.84 (0.53–1.33)	0.459
Length of hospital stay before ICU admission (days)	1.10 (1.04–1.15)	<0.001	1.03 (0.99–1.07)	0.156
Clinical Frailty Scale score (before admission)	1.13 (1.03–1.24)	0.012	1.07 (0.92–1.24)	0.376
Dexamethasone (extended)	1.58 (1.09–2.30)	0.017	0.84 (0.47–1.49)	0.554
Tocilizumab	0.68 (0.37–1.27)	0.229	0.99 (0.45–2.15)	0.979
SOFA	1.13 (1.07–1.19)	<0.001	1.02 (0.95–1.09)	0.624
MV (MV only)	4.51 (2.71–7.51)	<0.001	3.69 (1.95–6.97)	<0.001
Prone positioning	1.29 (0.88 – 1.88)	0.193	1.23 (0.76–2.00)	0.394
ECMO	3.80 (2.46–5.88)	<0.001	4.34 (2.35–8.01)	<0.001
Tracheostomy	2.09 (1.46–3.00)	<0.001	0.93 (0.58–1.50)	0.765
CRRT	9.65 (5.89–15.81)	<0.001	6.55 (3.59–11.98)	<0.001

OR, odds ratio; CI, confidence interval; BMI, body mass index; ICU, intensive care unit; SOFA, Sequential Organ Failure Assessment; MV, mechanical ventilation; ECMO, extracorporeal membrane oxygenation; CRRT, continuous renal replacement therapy.

To further evaluate the impact of BSI on mortality while accounting for its time-dependent occurrence, a supplementary Cox PH analysis was performed incorporating BSI as a time-varying covariate using the counting process formulation (Supplementary Table 6, only online). In this model, which additionally included prone positioning in the covariate set, BSI was not an independent predictor of in-hospital mortality (adjusted HR, 1.13; 95% CI, 0.82–1.56; p=0.452), consistent with the primary logistic regression result. Age (adjusted HR, 1.02; 95% CI, 1.01–1.04; p=0.007) and CRRT (adjusted HR, 2.58; 95% CI, 1.76–3.79; p<0.001) remained independent predictors of mortality, while extended dexamethasone use was associated with reduced mortality risk (adjusted HR, 0.60; 95% CI, 0.43–0.83; p=0.002). Prone positioning was not independently associated with mortality in this model (adjusted HR, 0.89; 95% CI, 0.64–1.25; p=0.513). The consistency between the logistic regression and Cox model results strengthens the conclusion that BSI serves as a marker of illness severity rather than an independent causal determinant of in-hospital death.

DISCUSSION

This nationwide multicenter retrospective cohort study demonstrated that BSIs are a frequent and clinically significant complication among critically ill adults hospitalized for COVID-19 who require advanced respiratory or organ support.

BSIs occurred in more than one-third (35.9%) of the cohort. While unadjusted analyses showed a strong association with increased ICU/hospital LOS and higher crude mortality rates, multivariable analysis revealed that BSI was not an independent predictor of in-hospital death. Independent risk factors for BSI in critically ill patients with COVID-19 included a history of chronic neurological disease, higher SOFA score, prone positioning, and ECMO in the primary logistic regression analysis; chronic neurological disease and SOFA score were further confirmed as robust predictors in a supplementary Cox PH model. In-hospital mortality was also associated with older age and the use of advanced life-sustaining therapies including CRRT; these findings were consistent across both logistic regression and Cox PH analyses.

Severe illness, reliance on invasive life-sustaining treatments, and prolonged hospital stays contribute to immunosuppression in critically ill patients, thereby increasing vulnerability to secondary infections.¹⁴,¹⁵,¹⁶ Secondary infections are commonly reported in critically ill patients with COVID-19.¹⁷,¹⁸ The incidence of BSIs in patients with COVID-19 admitted to ICUs ranges from 10% to 50%, nearly four times higher than in ICU patients without COVID-19.⁸,¹⁹ In this study, the incidence of BSIs was 35.9%, aligning with the upper range of previously reported rates of 10%–50%.

Previous studies have evaluated BSI risk factors in COVID-19. Several have reported an association between MV and increased BSI risk.²⁰,²¹ Tracheal intubation, required for MV, can introduce pathogens or cause airway injury, predisposing patients to pulmonary infections.²² In our study, while the unadjusted comparison showed a higher rate of MV use in the BSI group, it remained independently associated with BSI in the primary logistic regression after strictly adjusting for the timing of the intervention. However, in the supplementary Cox PH model—which additionally accounts for differences in time-at-risk across patients—MV did not retain independent significance. This attenuation suggests that a portion of the logistic association may reflect the longer hospitalization periods inherent to ventilated patients, during which BSI risk accumulates passively. After accounting for this differential time-at-risk, MV does not appear to confer independent BSI risk beyond what is captured by illness severity. In this context, the need for MV may act more as a proxy for overall illness severity in this highly selected cohort. Although the VIF analysis indicated no severe statistical multicollinearity (all VIFs <1.4), sequential modeling revealed that the baseline SOFA score was a highly significant independent predictor of both BSI and mortality prior to the inclusion of advanced interventions. Upon adding MV, ECMO, and CRRT to the model, the predictive weight of the SOFA score was substantially absorbed. This suggests that these interventions serve as direct clinical manifestations of severe organ dysfunction, effectively capturing the prognostic information initially provided by the SOFA score.

In our multivariable analysis, tocilizumab was not significantly associated with an increased risk of BSI. Unlike some reports suggesting higher infection risks with interleukin-6 inhibitors, our findings indicate that tocilizumab use is not necessarily linked to increased BSI rates in strictly monitored ICU settings. The observed trend (OR 0.64) may reflect tocilizumab’s role in mitigating hyper-inflammation, which could prevent clinical deterioration and reduce cumulative exposure to invasive ICU interventions. However, these findings should be interpreted with caution due to potential institutional variations in protocols.

A particularly noteworthy finding is the robust association between chronic neurological diseases and both early and overall BSI risk. In this study, chronic neurological disease—encompassing stroke, neuromuscular disease, epilepsy, and movement disorders—was identified as a significant independent predictor of BSI. Notably, while a recent 2025 meta-analysis⁸ provided a comprehensive overview of general risk factors for secondary infections in COVID-19, it did not explicitly identify chronic neurological conditions as a distinct risk factor. The association observed in the present study may reflect an increased vulnerability due to impaired airway protection, immobility, or altered immune responses unique to this subpopulation. Particularly, these specific neurological diseases may predispose patients to secondary infections, potentially due to a reduced cough reflex, impaired swallowing function (dysphagia), and an increased need for endotracheal intubation, which could contribute to the observed outcomes. Our findings thus extend the current literature by highlighting a vulnerable group that has been largely overlooked in pooled meta-analytical data. Importantly, the association between chronic neurological disease and BSI risk was consistent across both the primary logistic regression and the supplementary Cox model, reinforcing its status as a robust, time-independent predictor.

In this study, the microbiological spectrum of BSI shifted over time. The predominance of gram-positive cocci, particularly Staphylococcus species, in BSIs occurring within the first 2 days of admission is likely related to central line-associated BSIs and skin colonization. This is in line with previous reports that early-onset BSIs in ICU settings are often device-related.²³,²⁴ Furthermore, a substantial proportion of our cohort was transferred from outside hospitals or long-term care facilities, often arriving with pre-existing invasive devices. For these transferred patients, the “early” phase of admission to our ICUs actually reflects a later phase of their overall healthcare exposure, which strongly contributes to the early isolation of these healthcare-associated pathogens. The marked increase in gram-negative organisms, especially K. pneumoniae, from days 3 to 7 and the sustained predominance of Klebsiella and Acinetobacter species thereafter may reflect the cumulative impact of prolonged MV, hospital-acquired infection, and broad-spectrum antibiotic use. Furthermore, the proportion of fungal BSIs, predominantly Candida species, increased substantially with longer hospital stay, which is consistent with the known risks conferred by immunosuppression, broad-spectrum antimicrobials, and multiple invasive devices.

Recent data from Asia and Europe challenge historical paradigms by demonstrating gram-positive predominance during early ICU course, with Staphylococcus aureus and E. faecium accounting for 42%–58% of BSIs within the first week of admission.²⁵,²⁶ This observed shift appears to correlate with the widespread use of invasive procedures, including MV, tracheostomy, and central venous catheterization, each of which has been associated with a 2.5–9.3-fold increase in BSI risk.⁹,²⁷ Moreover, ICU stays longer than 15 days are associated with heightened exposure to multidrug-resistant organisms, with late-onset infections frequently caused by carbapenem-resistant Acinetobacter baumannii and extended-spectrum β-lactamase-producing Enterobacteriaceae. ⁹,²⁸ Compared with previous studies on critically ill patients with COVID-19 that reported gram-negative pathogens as the predominant causative agents of BSIs,⁹,¹⁰ our findings indicate that gram-positive cocci are more prevalent in the early stages. Clinically, these findings highlight the need for time-sensitive, targeted infection prevention and antimicrobial stewardship. These results may inform clinical decision-making among practicing clinicians.

Patients with BSI experienced significantly worse outcomes, including longer ICU and hospital stays and higher mortality rates. In the multivariable logistic regression analysis, in-hospital mortality was also independently associated with older age, solid malignant tumors, higher clinical frailty scale scores (before admission), and greater use of invasive organ support (MV, ECMO, and CRRT). These findings reinforce the concept that both patient-related factors (age and comorbidities) and treatment-related factors (invasiveness of support) contribute to poor outcomes in patients with severe COVID-19.

Previous studies have frequently suggested that BSI is independently associated with increased mortality in patients with COVID-19. However, in the present study, BSI was not identified as an independent risk factor for in-hospital mortality after adjusting for baseline covariates. This discrepancy may be explained by the specific characteristics of our study population, which consisted exclusively of critically ill COVID-19 patients. In this highly selected cohort presenting with severe baseline illness and a high reliance on invasive organ support (e.g., MV and ECMO), the attributable mortality of the BSI itself is likely overshadowed by the profound severity of the underlying viral pneumonia, acute respiratory distress syndrome, and associated multi-organ failure. Furthermore, in such a critical care setting, BSI may serve more as a surrogate marker of prolonged intensive care, extensive invasive procedures, and extreme severity of illness. We acknowledge, however, that the absence of independent statistical significance in the adjusted model does not exclude a potential causal contribution of BSI to mortality; rather, it suggests that any such contribution may be largely mediated through, or confounded by, the severity of the underlying critical illness in this highly selected population. This interpretation is further strengthened by the supplementary Cox PH analysis, in which BSI was modeled as a time-varying covariate to directly account for its dynamic onset during hospitalization and the potential for immortal time bias. Even under this more stringent analytical framework, BSI remained a non-significant predictor of in-hospital mortality. The convergence of results across both logistic regression and Cox modeling is consistent with the interpretation that the prognostic impact of BSI on mortality, if present, may not be independent of illness severity and life-support intensity, and warrants further investigation in prospective studies with causal inference frameworks.

The strength of this study includes the identification of risk factors for BSIs in critically ill patients with COVID-19 and the characterization of causative pathogens by timing of onset, using data from a large multicenter cohort.

However, this study has certain limitations. First, the retrospective design may have introduced residual confounding and incomplete data capture. Second, detailed data on antimicrobial use and infection prevention protocols were unavailable, potentially influencing both BSI risk and outcomes. Third, the findings may not be generalizable to non–COVID-19 ICU populations or healthcare systems with different resources. Fourth, excluding 541 patients without blood culture data introduces potential selection bias, as these excluded patients were significantly less severely ill and required fewer life-supporting interventions (Supplementary Table 7, only online). This indicates that blood cultures were preferentially performed in sicker patients with higher clinical suspicion, a practice likely compounded by institutional protocol variations. Consequently, our study may overestimate BSI incidence, meaning the identified risk factors predominantly reflect a more critically ill subpopulation and highlight the need for standardized prospective studies. Fifth, due to the retrospective, multicenter nature of our registry, we lacked rigorously adjudicated clinical data to definitively determine the primary source of the BSI (e.g., catheter-related, respiratory, or urinary sources). The inability to differentiate primary from secondary BSIs restricts a comprehensive pathophysiological understanding of the infections. Sixth, the lack of granular sub-diagnoses within the “chronic neurological disease” category limits our ability to identify which specific neurological condition (e.g., stroke versus epilepsy) drives the greatest BSI risk. Seventh, although logistic regression was selected as the primary analytical approach given the binary framing of our outcomes and the application of strict temporal criteria for time-varying exposures, this method does not inherently account for variability in LOS or time-at-risk. Supplementary Cox PH analyses were therefore performed and confirmed that the primary findings were robust to this methodological consideration. Finally, our study cohort predates the Omicron variant and widespread COVID-19 vaccination. Although the viral landscape and population immunity have evolved, the core mechanisms driving secondary BSIs in the ICU—such as severe baseline organ failure, impaired airway protection, and prolonged use of invasive therapies like ECMO—remain unchanged. Therefore, our findings on risk stratification and temporal pathogen shifts remain highly relevant to contemporary critical care.

In conclusion, the development of BSI in critically ill COVID-19 patients is largely predicted by high baseline severity, chronic neurological comorbidities, and the need for specific intensive interventions such as ECMO and prone positioning prior to infection onset. These associations were robust across both primary logistic regression and supplementary Cox PH analyses. Furthermore, after strictly adjusting for confounding factors, temporal biases, and the dynamic time-dependent onset of BSI, BSI was not an independent risk factor for in-hospital mortality in either analytical framework. This convergent evidence suggests that the prognostic impact of BSI on in-hospital mortality, if any, may be largely attributable to its association with greater illness severity and more intensive life-sustaining interventions, rather than representing an independent causal pathway to death. Prospective studies employing causal inference frameworks will be needed to further elucidate this relationship. The temporal patterns—early gram-positive predominance followed by increasing gram-negative and fungal infections—underscore the need for tailored prevention strategies and empirical treatment approaches to improve outcomes in this vulnerable population.

DATA AVAILABILITY

All of the individual participant data collected during the trial will be available. The study protocol and statistical analysis will also be available immediately following publication. The data can be shared with anyone who wishes to access the data. Proposals should be directed to hihogogo@yuhs.ac to gain access.

SUPPLEMENTARY MATERIALS

Supplementary Table 1

Baseline characteristics (≤2days)

Supplementary Table 2

Baseline characteristics (≤7days)

Supplementary Table 3

Cox proportional hazards analysis of risk factors for BSI (time-to-BSI model)

Supplementary Table 4

Multivariable logistic regression analysis of risk factors for BSI (≤2 days)

Supplementary Table 5

Multivariable logistic regression analysis of risk factors for BSI (≤7 days)

Supplementary Table 6

Cox proportional hazards analysis of in-hospital mortality

Supplementary Table 7

Clinical characteristics, treatments, and outcomes (Excluded vs. Included)