Heterogeneity of Benefit from Earlier Time-to-Antibiotics for Sepsis

Abstract

Rationale

Shorter time-to-antibiotics improves survival from sepsis, particularly among patients in shock. There may be other subgroups for whom faster antibiotics are particularly beneficial.

Objectives

Identify patient characteristics associated with greater benefit from shorter time-to-antibiotics.

Methods

Observational cohort study of patients hospitalized with community-onset sepsis at 173 hospitals and treated with antimicrobials within 12 hours. We used three approaches to evaluate heterogeneity of benefit from shorter time-to-antibiotics: 1) conditional average treatment effects of shorter (⩽3 h) versus longer (>3–12 h) time-to-antibiotics on 30-day mortality using multivariable Poisson regression; 2) causal forest to identify characteristics associated with greatest benefit from shorter time-to-antibiotics; and 3) logistic regression with time-to-antibiotics modeled as a spline.

Measurements and Main Results

Among 273,255 patients with community-onset sepsis, 131,094 (48.0%) received antibiotics within 3 hours. In Poisson models, shorter time-to-antibiotics was associated with greater absolute mortality reduction among patients with metastatic cancer (5.0% [95% confidence interval; CI: 4.3–5.7] vs. 0.4% [95% CI: 0.2–0.6] for patients without cancer, P < 0.001); patients with shock (7.0% [95% CI: 5.8–8.2%] vs. 2.8% [95% CI: 2.7–3.5%] for patients without shock, P = 0.005); and patients with more acute organ dysfunctions (4.8% [95% CI: 3.9–5.6%] for three or more dysfunctions vs. 0.5% [95% CI: 0.3–0.8] for one dysfunction, P < 0.001). In causal forest, metastatic cancer and shock were associated with greatest benefit from shorter time-to-antibiotics. Spline analysis confirmed differential nonlinear associations of time-to-antibiotics with mortality in patients with metastatic cancer and shock.

Conclusions

In patients with community-onset sepsis, the mortality benefit of shorter time-to-antibiotics varied by patient characteristics. These findings suggest that shorter time-to-antibiotics for sepsis is particularly important among patients with cancer and/or shock.

At a Glance Commentary

Scientific Knowledge on the Subject

Shorter time-to-antibiotics improves survival from sepsis, particularly among patients in shock. There may be other subgroups for whom faster antibiotics are particularly beneficial.

What This Study Adds to the Field

This is an observational cohort study of 273,255 patients with community-onset sepsis. Three approaches were used to identify and evaluate patient characteristics associated with greater benefit from shorter time-to-antibiotics.

Prompt antimicrobial therapy is the foundation of treatment for bacterial sepsis. Shorter time to antimicrobial therapy (hereinafter termed “time-to-antibiotics”) improves survival from sepsis (1–4), although the estimated magnitude of benefit has varied across studies. This variable benefit may be explained, in part, by the heterogeneity of clinical presentations of sepsis, as well as our imperfect ability to diagnose sepsis, such that some patients treated for bacterial sepsis may have a noninfectious or nonbacterial illness (5–7). The association between time-to-antibiotics and mortality among patients treated for sepsis is consistently strongest among patients with shock (1–3), and recent studies have suggested that the impact of shorter time-to-antibiotics also varies by age (8, 9) and by clinical sepsis phenotype (10).

Given the stronger evidence for the need for prompt treatment in patients with shock, the Surviving Sepsis Campaign guidelines stratify their recommendations for antibiotic timing by the presence of shock (11, 12). This dichotomization may be overly simplistic, however, as there may be other groups of patients for whom shorter time-to-antibiotics is particularly important.

In this exploratory study, we sought to identify patient factors that are associated with greater mortality benefit from shorter time-to-antibiotics in sepsis. We evaluated how patient demographics, comorbidities, and clinical presentation modify the effect of time-to-antibiotics on 30-day mortality.

Methods

Study Design and Cohort

This is a multicenter observational cohort study of adult patients hospitalized through the emergency department (ED) with community-onset sepsis at 152 U.S. Veterans Affairs (VA) and 21 Kaiser Permanente Northern California (KPNC) hospitals (2013–2018). Patient and hospitalization characteristics were extracted from VA and KPNC electronic health records, as in prior work (13–16). Data on 30-day mortality, including deaths postdischarge, were captured through linkages to the national death index. Community-onset sepsis was identified by the following criteria: meeting two or more systemic inflammatory response syndrome (SIRS) criteria (17); administration of systemic antimicrobial therapy within 12 hours of ED arrival and continuation for 4 or more consecutive days (or died while receiving therapy); and acute organ dysfunction present within 48 hours of ED arrival, akin to the U.S. Centers for Disease Control and Prevention’s Adult Sepsis Event criteria (18). Time-to-antibiotics was defined as time from ED presentation to time of first systemic antimicrobial administration determined by barcode medication administration and physician order entry records (15). We used barcode medication administration data when available and otherwise considered administration time to be within 45 minutes of the antimicrobial order, as in prior work (15). In analyses using a dichotomous time-to-antibiotics variable, we considered shorter time-to-antibiotics as 3 hours or less, because the Surviving Sepsis Campaign Guidelines recommends treatment within 1–3 hours (11, 12). For consistency with prior work, we excluded hospitalizations at hospitals with less than 15 sepsis cases per year. (For further details on study definitions and assumptions, see Appendixes E1 and E2 in the online supplement.)

Statistical Analysis

The goal of this exploratory analysis was to assess patient characteristics associated with greater benefit from shorter time-to-antibiotics in sepsis. The 36 patient characteristics of interest were age, sex, six acute organ dysfunctions, four criteria for SIRS, and 24 chronic conditions (see Appendix E3). We initially focused on acute organ dysfunctions, age, and temperature. Post hoc, on the basis of results of the causal forest analysis, we also focused on cancer. Chronic conditions were identified according to Elixhauser criteria using diagnostic codes from inpatient and outpatient encounters in the 540 days (1.5 yr) preceding sepsis hospitalization (19–21). For unadjusted comparisons, we used a two-sided t test, Wilcoxon rank-sum test, chi-square analysis, or Fisher exact test. P < 0.05 denoted significance.

We used three approaches for assessing heterogeneity of benefit from shorter time-to-antibiotics among patients with sepsis. First, we evaluated conditional average treatment effects across subgroups defined by patient characteristics by fitting a series of robust Poisson regression models (22, 23), with 30-day mortality as the outcome and including an interaction term between the patient characteristic(s) of interest and time-to-antibiotics. Poisson regression was selected to allow for presentation of relative risks, and time-to-antibiotics was dichotomized as shorter (⩽3 h) versus longer (>3–12 h) to harmonize with the causal forest analysis (which requires a dichotomous treatment variable) (24). We calculated the absolute and relative differences in mortality between shorter versus longer time-to-antibiotics using adjusted conditional average treatment effects estimated from the models (25). All models included adjustment for age, sex, SIRS criteria, acute organ dysfunctions, and chronic conditions, but interaction terms differed across the models depending on the characteristic of interest for the particular model. For the Poisson regressions, we also performed subgroup analyses limited to 1) patients who received antibiotics within 6 hours, and 2) patients without shock. We calculated E-values to evaluate the strength of association an unmeasured covariate would need to negate the effect of time-to-antibiotics.

In the second approach, we fit an honest causal forest model (24, 26–28) to simultaneously examine all patient characteristics within a single model. Causal forest is a systematic approach for exploring treatment heterogeneity across many characteristics while accounting for complex, nonlinear interactions between characteristics (26). It overcomes several limitations of traditional approaches to exploring heterogeneity, including limited power, risk of false discovery from multiple comparisons, and difficulty comparing results across models (29, 30). Treatment estimates from honest causal forests were obtained by averaging results of hundreds of causal trees, each created by iteratively partitioning the sample into subgroups to maximize differences in estimated absolute mortality reduction across subgroups (31). The trees are considered “honest,” because the data used for making partitions are distinct from the data used to estimate treatment effects. The sample was then split into quartiles that were based on least (Quartile 1) to greatest (Quartile 4) estimated mortality benefit from shorter versus longer time-to-antibiotics averaged across the causal forest. Estimated treatment effects were compared across quartiles with augmented inverse propensity weighting, which compares a difference in means weighted by the estimated probability of treatment assignment (32, 33). Characteristics were then compared across quartiles by calculating a coefficient of variation, as well as variance, to identify characteristics with most variation across the quartiles. This approach is recommended over measures of variable importance for summarizing heterogeneity (26).

In the third approach, added post hoc, we fit logistic regression models with time-to-antibiotics modeled as a spline (k = 3 knots at 0.1, 0.5, and 0.9 quantiles) (34) and included interaction terms between time-to-antibiotics and 1) shock and 2) metastatic cancer. We focused this analysis on shock and metastatic cancer, because they were the most important factors explaining heterogeneity of treatment effect in the causal forest analysis. As in the other approaches, we adjusted for sex, age, SIRS criteria, acute organ dysfunctions, and chronic conditions (see Appendix E3).

To facilitate understanding and replication of our analyses, statistical code is publicly available at GitHub (CCMRPulmCritCare/AntibioticTimingHTE; www.github.com). Data management and analysis were performed using SAS Version 9.4 (SAS Institute), Stata 17 (StataCorp LLC), and R (R Core Team and R Foundation for Statistical Computing) (35). This study follows the Strengthening the Reporting of Observational Studies in Epidemiology (or STROBE) reporting guideline (36). The study was reviewed by the University of Michigan, VA Ann Arbor, and KPNC Institutional Review Boards, approved, and deemed exempt from the need for consent under 45 CFR §46, Category 4 (secondary use of identifiable data).

Results

Among 1,560,126 patients with SIRS criteria who were hospitalized through the ED during the study period, 273,255 (17.5%) met criteria for sepsis and received antibiotics within 12 hours of arrival (see Table E1 in the online supplement). Of these 273,255 hospitalizations with community-onset sepsis, 131,094 (48.0%) received antibiotics in 3 hours or less (“shorter time-to-antibiotics”) and 142,161 (52.0%) received antibiotics between more than 3 and 12 hours (“longer time-to-antibiotics”). (For a full distribution of time-to-antibiotics, see Figure E1). Patient age was a median of 69 years (interquartile range [IQR], 61–79), and 78.9% were male. The most common SIRS criteria on presentation were elevated heart rate (89.6%), abnormal white blood cell count (76.2%), and elevated respiratory rate (72.8%). The most common acute organ dysfunction was elevated lactate (57.0%), and 12.5% had shock. A total of 9.0% had metastatic cancer, and 13.3% had nonmetastatic cancer (Table 1).

Table 1.

Characteristics of Patients with Shorter versus Longer Time-to-Antibiotics

Characteristic	Total Cohort	Shorter Time-to-Antibiotics (⩽3 h)	Longer Time-to-Antibiotics (>3–12 h)
Total, n (%)	273,255 (100.0)	131,094 (48.0)	142,161 (52.0)
Time-to-antibiotics, h, median (IQR)	3.1 (1.9–5.3)	1.9 (1.4–2.4)	5.2 (3.9–7.3)
Age, yr, median (IQR)	69 (61–79)	70 (62–80)	69 (61–78)
Male, n (%)	215,598 (78.9)	96,756 (73.8)	118,828 (83.6)
SIRS, n (%)
Elevated heart rate	244,890 (89.6)	119,376 (91.1)	125,514 (88.3)
Abnormal white blood cell count	208,296 (76.2)	102,024 (77.8)	106,272 (74.8)
Elevated respiratory rate	198,992 (72.8)	102,478 (78.2)	96,514 (67.9)
Abnormal temperature	156,945 (57.4)	81,189 (61.9)	75,756 (53.3)
Acute organ dysfunction, n (%)
Lactate	155,858 (57.0)	88,379 (67.4)	67,479 (47.5)
Kidney	140,129 (51.3)	59,610 (45.5)	80,519 (56.6)
Hematologic	35,554 (13.0)	16,318 (12.4)	19,236 (13.5)
Shock	34,215 (12.5)	19,732 (15.1)	14,483 (10.2)
Liver	29,004 (10.6)	10,566 (8.1)	18,438 (13.0)
Respiratory	25,020 (9.2)	13,636 (10.4)	11,384 (8.0)
Comorbidity, n (%)
Chronic pulmonary disease	125,137 (45.8)	61,036 (46.6)	64,101 (45.1)
Renal disease	104,143 (38.1)	50,730 (38.7)	53,413 (37.6)
Heart failure	90,827 (33.2)	44,074 (33.6)	46,753 (32.9)
Liver disease	57,400 (21.0)	26,616 (20.3)	30,784 (21.7)
Nonmetastatic cancer	36,412 (13.3)	15,677 (12.0)	20,735 (14.6)
Metastatic cancer	24,466 (9.0)	11,698 (8.9)	12,768 (9.0)

Definition of abbreviations: IQR = interquartile range; SIRS = systemic inflammatory response criteria. Acute renal dysfunction: creatinine >1.2 mg/dl and a 50% increase from baseline. Patients with preexisting end-stage renal disease, as identified by diagnostic codes, were not eligible to have acute renal dysfunction. Acute liver dysfunction: total bilirubin >2.0 mg/dl and a 100% increase from baseline. Acute hematologic dysfunction: platelet count less than 100 cells per microliter and a 50% decrease from baseline. Abnormal lactate: >2.0 mmol/L. Acute lung/respiratory dysfunction: receipt of invasive mechanical ventilation. Acute cardiovascular dysfunction/shock: receipt of intravenous vasopressor therapy.

The median (IQR) time-to-antibiotics was 3.1 hours (1.9–5.3) overall, 1.9 (1.4–2.4) hours among patients treated within 3 hours, and 5.2 (3.9–7.3) hours among patients treated within 3–12 hours. Table 1 shows that patients who received shorter time-to-antibiotics were older (median age: 70 vs. 69, P < 0.001); had more SIRS criteria (e.g., elevated respiratory rate: 78.2% vs. 67.9%, P < 0.001; and abnormal temperature: 61.9% vs. 53.3%, P < 0.001); had more lactate elevation (67.4% vs. 47.5%, P < 0.001); and had more shock (15.1% vs. 10.2%, P < 0.001). By contrast, patients with longer time-to-antibiotics had acute kidney dysfunction (56.6% vs. 45.5%, P < 0.001) and acute liver dysfunction (13.0% vs. 8.1%, P < 0.001). Overall, in the baseline Poisson model without any interaction terms, shorter time-to-antibiotics was associated with a 1.20% absolute reduction in 30-day mortality (95% CI: 0.98–1.41) and a relative risk of 0.91 (95% CI: 0.89–0.93; P < 0.001) (Table 2).

Table 2.

Effect of Shorter versus Longer Time-to-Antibiotics on Absolute and Relative Risk of 30-day Mortality on the Basis of Patient Characteristics

Characteristic	Total Cohort, N (%)	Absolute Mortality Reduction % (95% CI)	Relative Risk (95% CI)
Overall cohort	273,255	1.20 (0.98, 1.41)	0.91 (0.89, 0.93)
Acute organ dysfunction
Lactate
Lactate elevation	155,858 (57.0)	5.19 (4.37, 6.00)	0.83 (0.80, 0.86)
No lactate elevation	117,397 (43.0)	4.00 (3.18, 4.79)	0.83 (0.79, 0.86)
Kidney
Kidney dysfunction	140,129 (51.3)	5.54 (4.68, 6.37)	0.82 (0.79, 0.85)
No kidney dysfunction	133,126 (48.7)	3.73 (2.96, 4.48)	0.84 (0.81, 0.87)
Hematologic*
Hematologic dysfunction	35,554 (13.0)	6.96 (5.86, 8.03)	0.78 (0.75, 0.81)
No hematologic dysfunction	237,701 (87.0)	2.67 (2.01, 3.32)	0.88 (0.85, 0.91)
Shock*
Shock	34,215 (12.5)	7.03 (5.85, 8.17)	0.80 (0.77, 0.83)
No shock	239,040 (87.5)	2.86 (2.20, 3.50)	0.86 (0.83, 0.89)
Liver*
Liver dysfunction	29,004 (10.6)	7.05 (5.78, 8.27)	0.77 (0.73, 0.81)
No liver dysfunction	244,251 (89.4)	2.54 (1.97, 3.09)	0.89 (0.86, 0.91)
Respiratory
Respiratory dysfunction	25,020 (9.2)	6.01 (4.56, 7.40)	0.84 (0.80, 0.88)
No respiratory dysfunction	248,235 (90.8)	3.44 (2.90, 3.96)	0.82 (0.79, 0.85)
Number of acute organ dysfunctions*
1	177,039 (64.8)	0.54 (0.28, 0.80)	0.95 (0.92, 0.97)
2	62,263 (22.8)	0.95 (0.44, 1.45)	0.94 (0.91, 0.97)
3+	33,953 (12.4)	4.80 (3.89, 5.68)	0.85 (0.83, 0.88)
Comorbidities*
No cancer	212,377 (77.7)	0.41 (0.22, 0.60)	0.96 (0.94, 0.98)
Nonmetastatic cancer	36,412 (13.3)	1.38 (0.92, 1.83)	0.88 (0.84, 0.92)
Metastatic cancer	24,466 (9.0)	5.03 (4.28, 5.76)	0.78 (0.75, 0.81)
SIRS criteria*
Abnormal temperature	156,945 (57.4)	1.50 (1.22, 1.76)	0.88 (0.86, 0.90)
Normal temperature	116,310 (42.6)	0.79 (0.44, 1.12)	0.94 (0.92, 0.97)
Age, yr*
<40	9,200 (3.4)	0.95 (0.16, 1.64)	0.83 (0.70, 0.97)
40–49	11,737 (4.3)	1.21 (0.41, 1.93)	0.84 (0.74, 0.95)
50–59	35,386 (12.9)	1.91 (1.42, 2.37)	0.80 (0.75, 0.85)
60–69	84,058 (30.8)	1.55 (1.19, 1.91)	0.87 (0.84, 0.90)
70–79	66,405 (24.3)	1.22 (0.75, 1.68)	0.91 (0.88, 0.95)
80–89	49,836 (18.2)	0.83 (0.15, 1.49)	0.96 (0.93, 0.99)
90+	16,633 (6.1)	0.92 (0.56, 2.32)	0.97 (0.92, 1.02)

Definition of abbreviations: CI = confidence interval; SIRS = systemic inflammatory response criteria. The impact of individual patient characteristics on the effect of time-to-antibiotics on mortality was estimated by fitting a series of robust Poisson regression models with 30-day mortality as the outcome and including an interaction term between the characteristic of interest and antibiotic timing. The models were then used to estimate the effect of shorter versus longer time-to-antibiotics, which was used to calculate the estimated absolute mortality difference and relative risk with shorter vs longer time-to-antibiotics. All models included adjustment for age, sex, SIRS criteria, acute organ dysfunction, and chronic conditions, but the interaction terms differed across the models.

^*P < 0.05, significant interaction between the covariate and the effect of shorter time-to-antibiotics.

Treatment Heterogeneity in Poisson Regression Models

In the series of Poisson regression models considering interactions between patient characteristic(s) and time-to-antibiotics, the association between shorter time-to-antibiotics and 30-day mortality differed by type and number of acute organ dysfunctions, comorbid cancer, temperature, and age (Table 2; see Figures E2, E3, E4, and E5).

Shorter time-to-antibiotics was associated with greater mortality benefit in patients with certain acute organ dysfunctions and with more acute organ dysfunctions (Table 2; see Figure E2). For example, among patients with versus those without shock, shorter time-to-antibiotics was associated with a 7.0% (95% CI: 5.9–8.2%) versus 2.9% (95% CI: 2.2–3.5%) adjusted absolute mortality reduction and a 0.80 (95% CI: 0.77–0.83) versus 0.86 (95% CI: 0.83–0.89) relative risk; P = 0.005 for the interaction between shock and effect of shorter time-to-antibiotics. Likewise, among patients with versus without acute hematologic dysfunction, shorter time-to-antibiotics was associated with a 7.0% (95% CI: 5.9–8.0) versus 2.7% (95% CI: 2.0–3.3) adjusted absolute mortality reduction and a 0.78 (95% CI: 0.75–0.81) versus 0.88 (95% CI: 0.85–0.91) relative risk; P < 0.001 for the interaction between acute hematologic dysfunction and effect of shorter time-to-antibiotics. Finally, among patients with versus without acute liver dysfunction, shorter time-to-antibiotics was associated with a 7.1% (95% CI: 5.8–8.3) versus 2.5% (95% CI: 2.0–3.1%) absolute mortality reduction and a 0.77 (95% CI: 0.73–0.81) versus 0.89 (95% CI: 0.86–0.91) relative risk; P < 0.001 for interaction between acute liver dysfunction and effect of shorter time-to-antibiotics. By contrast, lactate elevation, acute kidney injury, and acute respiratory dysfunction were not associated with differential relative risk from shorter time-to-antibiotics. However, the absolute mortality reduction with earlier time-to-antibiotics was greater among patients with versus without respiratory dysfunction, despite similar relative risk reductions (Table 2).

Shorter time-to-antibiotics was associated with greater mortality benefit in patients with solid organ cancer, particularly metastatic (Table 2; see Figure E3). Among patients with metastatic cancer, patients with nonmetastatic cancer, and patients with no cancer, shorter time-to-antibiotics was associated with a 5.0% (95% CI: 4.3–5.8), 1.4% (95% CI: 0.9–1.8), and 0.4% (95% CI: 0.2–0.6%) adjusted absolute mortality reduction, respectively. Patients with metastatic cancer showed a relative risk of 0.78 (95% CI: 0.75–0.81) versus 0.88 (95% CI 0.84–0.92) for nonmetastatic cancer and 0.96 (95% CI: 0.94–0.98) for patients without cancer; P < 0.001 for the interaction between solid cancer status and shorter time-to-antibiotics.

Finally, shorter time-to-antibiotics was associated with greater mortality benefit in patients with temperature abnormality on admission and in younger versus older patients (Table 2; see Figures E4 and E5). For example, in patients 50–59 years old versus patients 80–89 years old, shorter time-to-antibiotics was associated with a 1.9% (95% CI: 1.4–2.4) versus 0.8% (95% CI: 0.2–1.5) adjusted absolute mortality reduction and a 0.80 (95% CI: 0.75–0.85) versus 0.96 (95% CI: 0.93–0.99) relative risk of 30-day mortality; P < 0.001 for the interaction between age group and shorter time-to-antibiotics.

Results were similar in subgroup analyses limited to patients who received antibiotics within 6 hours (see Table E2) and in patients without shock (see Table E3). E-values for the relative risks are provided elsewhere (see Table E4) and overall suggest that a moderate-to-strong confounder would be required to negate the benefit of shorter time-to-antibiotics for most subgroups.

Treatment Heterogeneity in the Causal Forest Analysis

In the causal forest model, the overall estimated effect of shorter time-to-antibiotics was an absolute 30-day mortality reduction of 1.2% (95% CI: 0.9–1.5%), which is consistent with the Poisson regression estimate. Across the four quartiles, the absolute mortality benefit of shorter time-to-antibiotics was 3.6% (95% CI: 2.9–4.3%) for patients in Quartile 4 (greatest benefit), 1.4% (95% CI: 0.9–1.9%) for Quartile 3, 0.6% (95% CI: 0.1–1.0%) for Quartile 2, and −0.8% (95% CI: −1.4 to -0.3%]) for Quartile 1 (least benefit) (see Figure E6). The patient characteristics that differed most across quartiles (and, therefore, were most strongly associated with assignment to the highest benefit quartile) were metastatic cancer and shock, for which the coefficients of variation were 116.5% and 82.4%, respectively (Figure 2). Metastatic cancer was present in 26.8% of patients assigned to the highest benefit Quartile 4 versus in 1.2% of patients assigned to the lowest benefit Quartile 1. Shock was present in 30.2% of patients assisted to highest benefit Quartile 4 versus 6.0% in lowest benefit Quartile 1.

Figure 2.

Prevalence of patient characteristics by causal forest-generated quartile from least (Quartile 1) to greatest (Quartile 4) estimated mortality benefit with shorter time-to-antibiotics. Quartiles generated by causal forest analysis partitioning sample by estimated least (Quartile 1) to greatest (Quartile 4) mortality benefit with shorter time-to-antibiotics. Heatmap colors indicate lower (green) to higher (red) proportion of the sample with a particular characteristic. *Coefficient of variation (CV) is a measure of relative variability, where larger values indicate more variation between groups. WBC = white blood cells.

Treatment Heterogeneity in the Spline Analysis

In the analysis treating time-to-antibiotics as a spline, the risk of mortality increased nonlinearly with longer time-to-antibiotics (Figure 1). As expected, predicted probability of mortality was higher among patients with shock and with metastatic cancer. Visual inspection of Figure 1 also shows a steeper increase in mortality with longer time-to-antibiotics in patients with shock and metastatic cancer, particularly among the first 4 hours.

Figure 1.

(A) Adjusted relative risk in 30-day mortality with shorter time-to-antibiotics by presence of shock, with time-to-antibiotics modeled as a continuous variable through spline regression. (B) Adjusted relative risk in 30-day mortality with shorter time-to-antibiotics by presence of metastatic cancer, with time-to-antibiotics modeled as a continuous variable through spline regression. This spline analysis demonstrates a nonlinear association of time-to-antibiotics with outcomes, but cannot be used to recommend specific treatment timing thresholds because the splines knots that inform the bends are set by the underlying distribution of time-to-antibiotics, as is recommended best practice for parameterizing splines (34). The restricted cubic spline included K = 3 knots at 0.1, 0.5, and 0.9 quantiles of time-to-antibiotics. For the figures, we set all covariates at their means, report predicted probabilities of 30-day mortality, and include 95% confidence intervals for predictions.

Discussion

In this multicenter cohort study of over 250,000 sepsis hospitalizations at 173 hospitals with bar code medication administration data, we show that the benefit of prompt antibiotic administration varies among patient subgroups. Shorter time-to-antibiotics was associated with a 1.2% absolute reduction in 30-day mortality among all hospitalizations meeting objective criteria for community-onset sepsis. However, the estimated benefit was greater among patients with cancer, particularly metastatic cancer, patients with shock, patients with acute hematologic dysfunction, and patients with multiple acute organ dysfunctions. Overall, across three sets of analyses, the study findings showed meaningful variation in treatment benefit from shorter time-to-antibiotics.

Several of the findings of our study are consistent with prior work. First, the impact of shorter time-to-antibiotics was greater among patients in shock, consistent with prior studies (1–4). Shock was the second most influential characteristic associated with assignment to the highest benefit quartile in causal forest analysis. Second, the impact of shorter time-to-antibiotics was greater among patients with metastatic cancer, which is consistent with at least one prior study (9). In a retrospective cohort of 60,000 ED admissions with infection that used causal forest analysis, malignancy (localized and/or metastatic) was more common among the subgroup that was likely to be harmed by delayed antibiotic administration (24% in the cluster with most harm vs. 22% in the remaining cohort, P < 0.001) (9). In our causal forest analysis, metastatic cancer was the most influential characteristic associated with assignment to the highest benefit quartile and was more than 20-fold more common among patients in the highest versus lowest benefit quartiles. Our spline analysis also demonstrated that patients with metastatic cancer had a steeper increase in mortality with longer time-to-antibiotics. It is unclear why cancer was more strongly associated with treatment effect heterogeneity in our study compared with the prior study, but may relate to our differentiation of metastatic versus nonmetastatic cancer. We hypothesize that the greater benefit of shorter time-to-antibiotics for patients with metastatic cancer is multifactorial and may reflect both the greater likelihood of bacterial infection among patients treated for suspected sepsis and the greater risk for adverse outcomes because of the state of immune suppression induced by cancer and its treatment. Time-to-antibiotics in patients with cancer may also be associated with other factors that influence outcome independent of antibiotic timing. For example, some patients with metastatic cancer may be treated more rapidly because of their recognized risk of deterioration, whereas others may be treated less rapidly in the setting of treatment limitations.

Third, we found some evidence for heterogeneity of benefit by patient age. In the Poisson regression analysis, patients younger than 60 years experienced more than twice the absolute mortality benefit, with shorter time-to-antibiotics than patients older than 80. This finding is consistent with a recent study of the PHANTASi trial cohort, which found age to be an important factor explaining heterogeneity of benefit of randomization to pre–hospital antibiotics (8). The heterogeneity of treatment effect by age in the PHANTASi trial cohort was sensitive to parameterization of age, however, with statistically significant heterogeneity identified using an age cutoff of 76 years but not 74 years. This may explain why age dichotomized at age 75 was only moderately impactful in our causal forest model.

One finding of our study differed from prior research. Lactate elevation (⩾2 mmol/L) was not associated with differential mortality benefit from shorter time-to-antibiotics in Poisson analysis and was only minimally different across quartiles of benefit in causal forest analysis. This finding contrasts with prior studies that have found lactate elevation to be associated with greater harm from delayed antibiotic administration or, alternatively, greater benefit from prompt administration (9, 10). These differences may be explained by differences across study cohorts, as one prior study focused on patients with infection and only 43% had sepsis or septic shock (9). Our study focused exclusively on patients with sepsis and adjusted for shock and other acute organ dysfunctions that may be colinear with lactate elevation.

Our findings may have important implications for sepsis quality improvement and performance measurement efforts. Shorter time-to-antibiotics is a key performance measure that has been incentivized by sepsis quality improvement initiatives and federal performance measures, such as the Center for Medicare and Medicaid Services SEP-1 measure. However, our findings suggest that the benefit of shorter time-to-antibiotics may not be experienced equally across patients. Patients with certain characteristics, such as shock and cancer, appear to experience greater benefit from shorter time-to-antibiotics; as a result, differences in underlying case mix could impact quality measurement and outcomes. Future studies should seek to confirm our findings to improve targeted approaches to sepsis care that maximize benefits and minimize harms. Our findings support the Surviving Sepsis Campaign emphasis on timely antibiotics in patients with shock and suggest that those who provide performance measurement efforts may likewise consider focusing on subgroups of patients who are at highest risk of increased mortality from delayed antibiotics or, at the very least, stratifying the reporting of sepsis management by patient subgroups.

This study has several limitations. First, conditional average treatment effects cannot be compared across Poisson models. Thus, the Poisson models can assess whether a given factor is associated with treatment heterogeneity but cannot compare results across models to determine which factors are most important for explaining treatment heterogeneity. To address this limitation, we performed a complementary analysis using causal forest, which allows for the direct comparison of many factors within a single model. Second, as an observational study, our findings may be biased by unmeasured confounding. We had granular health record data to account for SIRS criteria, acute organ dysfunction, and chronic health conditions, but we did not have data to adjust for all factors, such as patients’ presenting symptoms (37), treatment limitations, or other factors that may contribute to delays in antibiotic administration, such as ED crowding. Third, we were unable to assess appropriateness of antibiotics, but we hypothesize that, if anything, inadequate antibiotics would bias effects of earlier antibiotics toward the null.

Fourth, we identified community-onset sepsis through objective electronic health record criteria. This approach is generally considered superior to approaches using diagnostic codes with greater accuracy (38) but may still result in misclassification (39). We hypothesize that factors associated with greater benefit from earlier time-to-antibiotics may reflect both greater accuracy of sepsis diagnosis as well as greater sensitivity to the timing of antibiotic treatment. Fifth, the data collected are older (2013–2018), although more recent data (2020–2021) are impacted by significant disruptions to usual care because of the coronavirus disease (COVID-19) pandemic. Sixth, most of the sample is male, which reflects the large proportion of patients from VA hospitals and which may impact the generalizability of the findings. However, we are not aware of any data to suggest that the impact of time-to-treatment varies by sex; indeed, male sex was among the least important factors explaining heterogeneity in the causal forest analysis. Finally, although our spline analysis showed nonlinear associations of time-to-antibiotics with outcomes, they cannot be used to recommend specific treatment timing thresholds, because the splines’ knots, which inform the bends, are set by the underlying distribution of time-to-antibiotics, as is recommended best practice for parameterizing splines (34).

In conclusion, in this large, multicenter cohort of sepsis hospitalizations, we show meaningful heterogeneity of benefit from shorter time-to-antibiotics. Factors that were most associated with benefit from earlier antibiotics included shock, metastatic cancer, and number of acute organ dysfunctions. Our study suggests that, although prompt administration of antibiotics is important for all patients with bacterial sepsis, those who provide bedside care, quality improvement efforts, and performance measurement may consider focusing on patients who are likely to receive the greatest benefit from shorter time-to-antibiotics.