Data for research purposes (data collection) were accessed between 18 March 2019 and 12 December 2019 with ethical approval from the Hasanuddin University Research Ethics Committee (Reference Number: UH18120948). The authors had no access to information that could identify individual participants during or after data collection. All data were fully anonymized before authors accessed them, and the Education and Research Department of Dr. Wahidin Sudirohusodo Hospital waived the requirement for informed consent.

For the data analysis conducted for this manuscript, the authors extended the local ethical approval from the Hasanuddin University Research Ethics Committee (Reference Number: UH20120948) and obtained additional ethical approval from the London School of Hygiene & Tropical Medicine MSc Research Ethics Committee (Reference Number: 29175).

This analysis used clinical, pharmacy and laboratory data from Dr Wahidin Sudirohusodo Hospital, Makassar, South Sulawesi, Indonesia from 2015 to 2018. Wahidin Hospital is a third-level public health facilities with 935 bed capacity. It receives referral patients from second-level health facilities for specialist services such as oncology and cardiology.

Data used were routinely collected hospital electronic records which were extracted from the hospital information system. Further details of data collection are described elsewhere [9]. This study used data from all hospitalized adult patients with the first BC taken after two calendar days of admission over the study period. One patient could have multiple hospital admissions, and one admission could result in multiple BC tests. The first BC per patient for the full study period was selected for the main analysis. All BCs were included for a sensitivity analysis.

Patient’s CCI was the exposure of interest and derived from ICD10 codes of patients’ clinical diagnoses based on Claims-based, Disease-specific refinements, Matching translation to ICD-10 and Flexibility to use (CDMF) CCI Coding Schemes [8]. We applied the relevant weight (1, 2, 3 or 6 points) for each comorbidity whilst considering the hierarchy rule [8], and used age-adjustment (1 point for 50–59 years, 2 points for 60–69 years, 3 points for 70–79 years and 4 points for ≥ 80 years) [7]. Hierarchy rules refer to the principle that when multiple related conditions are present in a patient’s medical record, only the most severe condition contributes to the total CCI score. Example, if a patient has both cerebrovascular disease (1 point) and hemiplegia or paraplegia (2 points), only the higher score of 2 points for hemiplegia/paraplegia is included in the CCI total, since hemiplegia/paraplegia is a more severe consequence of cerebrovascular disease [8]. A unique CCI score was calculated per patient for each admission.

Positivity of BC taken two calendar days after admission (hospital-acquired BSI) was the primary outcome and was categorized as a binary variable (negative or positive for common contaminants [Bacillus spp. (not B. anthracis), coagulase-negative staphylococci, viridans group streptococci, Aerococcus spp. and Micrococcus spp [10]] vs. positive for any pathogen). For a secondary analysis, we examined positivity of BC taken two days after admission with priority pathogens. We excluded BC positive for other pathogens. We defined priority pathogens as the five most common pathogens in Wahidin Hospital that were included in the 2015 WHO Global AMR priority pathogen list. These were Escherichia coli, Klebsiella pneumoniae, Pseudomonas aeruginosa, Acinetobacter spp. and, Staphylococcus aureus [11]. Since our tertiary care center has a relatively high level of referrals from other healthcare centers, identification of community-onset organisms by the two-calendar day threshold would be biased, and these were therefore excluded.

Blood culture utilization rate (per 1000 patient-days) is the total number of blood cultures collected from 2015 to 2018 divided by the sum of all patient-days spent in the hospital during the same period and multiply by 1000. Patient-days refer to the total number of days that all patients spend in the hospital from 2015 to 2018.

Variables that are considered important confounders, based on a Directed Acyclic Graph (DAG), included age, sex, admission type, direct admission to intensive care unit (ICU) or high care unit (HCU), length of hospitalization (LOS) prior to BC test, ward of BC test and intravenous antibiotic prior to BC test (Fig 1). Admission type (emergency versus elective visit) and direct admission to ICU/HCU were considered as proxy variables of acute severity of illness.

Admission type (binary) was defined as the department (emergency or outpatient clinic) that referred the patient to be hospitalized. LOS prior to BC test was handled as a continuous variable. Ward of BC test was categorized based on type (emergency, general, progressive or HCU and intensive or ICU). For multivariable analysis, wards were grouped into binary variable (emergency/general unit and progressive/intensive unit) due to small numbers in progressive wards. Intravenous antibiotic prior to BC test was a continuous variable and defined as the duration from the first date of any intravenous antibiotic prescription to the BC sampling date, regardless of whether the administration was continuous or intermittent and the number of antibiotics used. As an effect modifier, intravenous antibiotics prior to BC test were grouped into binary variables based on presence/absence of prior antibiotic treatment.

Descriptive analysis was applied, using proportions, median and interquartile range, or mean and 95% confidence interval, whenever appropriate. Multivariable logistic regression analysis was performed to identify the association between CCI score (continuous independent variable) and BC positivity (binary dependent variable) for both the primary (first BC positive for all pathogens) and secondary (first BC positive for five priority pathogens) outcomes. For sensitivity analysis, all repeated BC tests were used as the outcome while considering random effects to take account of the within-cluster correlation variation as one patient could have several admissions or several BC tests during one admission. For patients with missing data of antibiotic dates, we assumed that no antibiotics were given when the BC was taken.

Likelihood ratio tests (LRT) were performed to compare nested models with CCI as categorical variable (CCI score of 0, 1, > 1) with the complex model including CCI as a continuous variable. “Use of antibiotics prior to BC test” was hypothesized as a potential effect modifier because the impact of CCI on BC could be different depending on the presence/absence of prior antibiotic treatment, which was tested with LRT as well. Effect modifier was evaluated for any pathogens using the first BC taken after two calendar days of admission.

There were 61,284 adult patients (total 98,102 admissions) hospitalized during four years of study with 4,562 (7.4%) patients having at least one BC test performed (total 5,704 BC). The total patient-days were 892,149, giving a BC utilization rate of 6.4 BC per 1,000 patients-days. Of those 4,562 patients, there were 3,875 (84.9%) patients with their first BC taken two calendar days after admission (total 4,797 BC) (Fig 2). These 3,875 patients were included in the study analysis.

Of the 3,875 patients, 460 (11.9%) had BC taken prior to antibiotic administration. Patients’ characteristics are presented in Table 1. Minimal missing data was found for hospital discharge date (1/3,875; 0.03%) and antibiotic date (132/3,875; 3.4%). For the single patient with a missing discharge date, length of stay was imputed as one day based on identical admission, antibiotic start dates and antibiotic end dates. For 132 cases with missing data of antibiotic dates, we assumed that no antibiotics were given when the BC was taken.

Out of 3,875 patients with their first BC taken two calendar days after hospital admission, 786 (20.3%) had their first BC positive for any pathogen, 2,705 (69.8%) negative or no growth, and 384 (9.9%) positives for common contaminants (Table 1). Within these 786 patients, 371 patients (47.2%) had their first BC positive for priority pathogens; S. aureus (n = 133; 35.9%), Acinetobacter spp. (n = 84; 22.6%), K. pneumoniae (n = 58; 15.6%), E. coli (n = 63; 17.0%) and P. aeruginosa (n = 33; 8.9%).

Overall, the median (IQR, range) of CCI score in all patients, patients with positive BCs, and patients with negative BCs were 2 (1–3, 0–11), 2 (0–4, 0–11), 2 (1–3, 0–11), respectively. Total CCI score and proportion of total CCI score, and BC positivity was showed in Figs 3 and 4.

The most common comorbidities observed in 3,875 patients were any malignancy (n = 494; 12.8%), severe renal disease (372; 9.6%), congestive heart failure (361; 9.3%), diabetes with complications (324; 8.4%), diabetes without chronic complications (247; 6.4%), and cerebrovascular disease (229; 5.9%). The most common comorbidities observed amongst patients who had positive BC tests were severe renal disease (122; 15.5%), any malignancy (n = 87; 11.1%), diabetes with complications (66; 8.4%), congestive heart failure (65; 8.3%), cerebrovascular disease (59; 7.5%) and diabetes without chronic complications (46; 5.9%).

In our main analysis using any positive BC as the outcome, we found that, there was no evidence that CCI score was associated with positive BC (adjusted odds ratio [aOR] 1.01, 95%CI: 0.96-1.06 p = 0.69) (Table 2). Result of sensitivity analysis also support this finding.

Using CCI score as a categorical variable (CCI score of 0, 1, > 1) gave a better model fit (LRT, p = 0.069), and showed some indication that higher scores of CCI were less likely to have positive BC (aOR 0.84, 95%CI:0.63-1.11 and aOR 0.79, 95%CI: 0.62-1.01, p = 0.18) compared to CCI of 0, although confidence intervals included the null value.

We considered an alternative outcome of a positive BC result with any one of five priority pathogens (Table 2). Using this outcome, there was evidence that increasing CCI was associated with BC positivity (aOR 1.08, 95%CI: 1.01-1.15, p = 0.02), which was supported by sensitivity analysis aOR 1.09 95%CI: 1.01-1.19 p = 0.04). Using CCI score as categorical variable (CCI scores of 0, 1, > 1) only showed an increased risk for CCI > 1 with wide confidence intervals (aOR 0.98, 95%CI: 0.66 -1.46 and aOR 1.16, 95%CI: 0.83 - 1.61, p = 0.51).

There was some indication that antibiotic prior to BC test acted as an effect modifier between CCI score (CCI score of 0, 1, > 1) and positive BC for any pathogens (Table 3) but this was not significant (p = 0.17). This non-significance may have been due to small sample size – there were relatively few patients who had BC performed without prior antibiotic treatment. After controlling for other factors, patients who had no antibiotic prior to BC test were likely to have positive BC at higher comorbidity score (aOR 1.16, 95%CI:0.54-2.49 for CCI score 1 and aOR 1.28, 95%CI: 0.27-2.49 for BC for CCI score >1) whereas patients who had antibiotic prior to BC test were less likely to have positive BC with rising score of comorbidities (aOR 0.84, 95%CI: 0.49-1.43 for CCI score 0, aOR 0.66, 95%CI: 0.38-1.16 for CCI score 1 and aOR 0.61, 95%CI: 0.36-1.04 for CCI score >1).