Risk-adjusted comparisons of bloodstream infection rates in neonatal intensive-care units


Corresponding author: P. Leighton (née Phillips), MRC Centre of Epidemiology for Child Health, UCL Institute of Child Health, 30 Guilford Street, London WC1N 1EH, UK
E-mail: ph.leighton@gmail.com


Clin Microbiol Infect 2012; 18: 1206–1211


Comparisons of bloodstream infection (BSI) rates between neonatal intensive-care units (NICUs) should take into account differences in babies’ vulnerability and invasive procedures that can introduce infection. Our aim was to investigate which risk factors recorded in routine records should be adjusted for when NICUs are compared. This was a retrospective cohort study using routine records for two London NICUs. We analysed rates of BSI with Poisson regression models. The level of neonatal care used by the National Health Service was the strongest predictor of BSI incidence. The rate ratios for BSI, adjusted for birthweight, inborn/outborn status, and postnatal age, were 3.15 (95% CI 2.01–4.94) for intensive care and 6.58 (95% CI 4.18–10.36) for high-dependency care, relative to special care. Total parenteral nutrition was significantly associated with BSI incidence, but explained less of the variance among babies than level of care. A case–control study with the same dataset gave similar results. Further multicentre studies are required to confirm our predictive model. Until then, we recommend that comparisons of BSI rates between NICUs should include adjustments for level of care, birthweight, inborn/outborn status, and postnatal age, with the use of routinely recorded standardized measures in hospital administrative data.


Improving hospital-acquired infection control is a priority for neonatal intensive-care units (NICUs). Between 2% and 10% of babies admitted to NICUs experience at least one episode of bloodstream infection (BSI) [1], which can lead to death, neurodevelopmental impairment, and other serious outcomes [2]. Clinicians need tools with which to monitor infection over time and to make comparisons between hospitals, in order to identify potentially better or harmful practices. Monitoring and sharing of improved practices may reduce infection rates [3,4].

NICUs differ in case mix, length of stay, and the invasive medical procedures carried out, all of which can influence BSI rates [5]. Comparisons between hospitals and over time should use BSI rates that are stratified, or adjusted, for factors associated with infection. Any residual variation may be explained, at least in part, by factors amenable to change, such as hygiene practices. A systematic literature review of methods for comparing the incidence of BSI among NICUs found substantial variation [6].

Our aim was to determine which factors should be incorporated into risk adjustment models by comparing BSI rates in two large tertiary NICUs. In contrast to dedicated data collection, the use of electronic routine hospital records would accelerate and minimize staff workload in infection monitoring. We therefore limited our analyses to factors available in routinely collected National Health Service (NHS) data.

Materials and Methods


The study population comprised all babies admitted to two inner-London NICUs with c. 260 (NICU 1) and 430 (NICU 2) admissions each year. Both units admit inborn babies and referrals. Data were analysed for babies admitted on or after 1 May 2001 and discharged up to and including 28 February 2005.

Outcome definition

NICU blood culture results were extracted from routine microbiology laboratory records at each hospital. An episode of BSI was defined as one or more blood cultures in which the same bacterial organism was isolated within a 7-day period. This case definition included potential contaminants such as coagulase-negative staphylococcus (CONS). Previous studies have used case definitions including clinical observations [7,8]; ours was based on blood culture results recorded in routine records, to provide a pragmatic and sustainable approach to infection monitoring.

We wished to compare rates of hospital-acquired infection between NICUs, as this could inform infection control practices. Previous studies have excluded BSI occurring within the first 2 or 3 days of life as maternally transmitted [1,9]. We derived an age threshold by exploring age at first BSI episode using finite mixture models [10–12] (Fig. S1), and we excluded BSI occurring during the first 48 h of life. This threshold also reduced sampling bias, as it removed routine blood samples taken shortly after birth. The more blood samples taken, the greater the risk of detecting asymptomatic BSI or a contaminated blood culture.

Potential risk factors

Clinical data were extracted from the Patient Administration System, and linked with microbiology laboratory results. Some potential risk factors were recorded at birth (gestational age at birth, birthweight, birthweight standardized for gestational age expressed as a standard deviation score [13,14], inborn/outborn status, NICU, sex, and delivery method) and some were recorded daily (level of care defined in Table 1, postnatal age, number of blood samples taken, surgery, total parenteral nutrition, and ventilation) [15]. Total parenteral nutrition and ventilation information were available only for NICU 2.

Table 1.   UK National Health Service neonatal intensive-care unit levels of care [ 15 ] Thumbnail image of

It would be useful for clinicians to know which factors precede or predict BSI, in order to identify high-risk groups who could benefit from preventive action. To identify factors that may predict BSI, we analysed care variables recorded in the 3 days preceding infection. For this 3-day period, the most intensive level of care was recorded, and total parenteral nutrition and ventilation were labelled as present or absent.

Poisson regression models for BSI rates with potential risk factors

The analysis of exposures in the 3 days preceding BSI allowed comparisons with findings reported for another tertiary-referral NICU (1367 babies; 124 BSI episodes) in London by Holmes et al. [16]. Following their methods, we calculated rates of BSI as days with onset of a first BSI episode divided by total days of NICU stay. To differentiate days of stay that may contribute to BSI from days of stay that may be the consequence of BSI, baby-days were counted up until the first BSI episode for infected babies and up until discharge from the NICU for uninfected babies. Crude Poisson regression models were fitted to estimate rate ratios for BSI, for each potential risk factor in turn. Generalized estimating equations with an exchangeable correlation structure were used to take account of the fact that days pertaining to the same baby are not independent [17]. Potential risk factors with significant associations with BSI (p <0.01) were examined in combination, by use of Poisson regression and forward selection of risk factors. The combination of risk factors giving the lowest quasi-likelihood information criterion (QIC) [18] was included in the final adjusted model.

The analyses were repeated for each hospital separately. As gestational age at birth and birthweight were correlated, separate adjusted models were built for each of these variables. A similar approach was taken for level of care, total parenteral nutrition and ventilation in NICU 2.

As an alternative to the Poisson regression models, we analysed the odds of BSI by using logistic regression models, assuming a matched case–control design (see Supporting Information, Tables S1 and S2).

Previous studies have used survival analyses to determine risk factors for BSI [7,19]. Babies can have a complicated course through the NICU, with intermittent exposures to different levels of care, ventilation, and/or total parenteral nutrition. For such sporadic exposures, the analysis of risk factors in the 3 days preceding a BSI episode is more practical than a simple survival analysis. Analyses employed R 2.7.0 [20] and Stata 10.0 [21].

Ethics approval was received from the National Hospital for Neurology and Neurosurgery and the UCL Institute of Neurology Joint Research Ethics Committee.


Two-hundred and thirty-six first episodes of BSI were included, of which 176 were caused by CONS, two by group B streptococcus, 27 by Gram-positive organisms other than group B streptococcus, 17 by Gram-negative organisms, and four by yeasts. Ten episodes were mixed cultures, of which six contained CONS. The Poisson regression models were fitted for 2269 babies (940 from NICU 1 and 1329 from NICU 2). The median birthweight was lower in NICU 1 than in NICU 2: 2000 g (interquartile range (IQR) 1320–2955 g) vs. 2536 g (IQR 1740–3240 g). The median gestational age at birth was similar for the NICUs: 35 weeks (IQR 30–39 weeks) in NICU 1 and 36 weeks (IQR 33–39 weeks) in NICU 2. As the two NICUs had similar rates of BSI, c. 6/1000 baby-days, and similar findings for all analyses, we present their aggregate results.

Level of care was the single strongest risk factor for BSI, in terms of optimizing the QIC. Intensive care accounted for 36% (14 443/40 218) of total NICU days and 58% (138/236) of BSIs, and high-dependency care accounted for 9% (3603/40 218) of NICU days and 20% (47/236) of BSIs (Table 2). When both hospitals were combined, the optimal adjusted model consisted of level of care, birthweight, inborn/outborn status, and postnatal age (Table 3).

Table 2.   Associations between potential risk factors and bloodstream infection (BSI) rates: crude Poisson regression models for neonatal intensive-care unit (NICU) 1 and NICU 2 combined
Potential risk factorBabies (n)aDays with onset of BSI/total baby-days (rate/1000 baby-days)Crude rate ratio (95% CI), p-value
  1. CS, caesarean section.

  2. aThe number of babies is given for factors reflecting susceptibility to BSI at birth. Factors that change during the NICU stay include variable numbers of babies.

  3. bIn the previous 3 days.

  4. cOther’ indicates that, for the previous 3 days, the baby was outside the NICU, e.g. at another hospital or undergoing surgery.

  5. dDays/babies with missing variables were few in number and represented few episodes of BSI. For this reason, we considered it acceptable to remove them from the analyses.

  6. eBirthweight categories included in a sensitivity analysis.

  7. fStandard deviation scores were calculated with the LMS (λ–μ−σ) method [13,14].

  8. ..

Highest level of careb
 Intensive care138/14 443 (9.55)5.42 (3.78–7.77), <0.001
 High-dependency care47/3603 (13.04)7.30 (4.76–11.19), <0.001
 Special care36/20 919 (1.72)1
Total236/40 218
Gestational age (weeks)
 <2612672/5619 (12.81)2.67 (1.80–3.94), <0.001
 26 to <2810635/4398 (7.96)1.58 (1.00–2.48), 0.050
 28 to <3232353/11 319 (4.68)0.92 (0.61–1.37), 0.668
 32 to <3765030/10 620 (2.82)0.51 (0.32–0.81), 0.005
 ≥37105246/8233 (5.59)1
Birthweight (g)
 <70010262/3634 (17.06)4.82 (3.39–6.85), <0.001
 700 to <120027578/12 400 (6.29)1.76 (1.29–2.40), <0.001
 ≥1200187796/24 146 (3.98)1
Birthweight (g)e
 <70010262/3634 (17.06)2.28 (1.25–4.17), 0.007
 700 to <120027578/12 400 (6.29)0.83 (0.47–1.49), 0.537
 1200 to <250084558/16 404 (3.54)0.42 (0.24–0.77), 0.004
 2500 to <350073324/6108 (3.93)0.46 (0.24–0.90), 0.022
 ≥350029914/1634 (8.57)1
Birthweight standardized for gestational age: standard deviation scoref
 <−227538/5660 (6.71)1.04 (0.68–1.59), 0.849
 −2 to <−144649/7811 (6.27)0.97 (0.66–1.42), 0.855
 −1 to <066662/12 529 (4.95)0.77 (0.54–1.10), 0.147
 0 to <153961/9446 (6.46)1
 1 to <222218/3052 (5.90)0.91 (0.53–1.55), 0.714
 ≥2996/1365 (4.40)0.66 (0.30–1.45), 0.296
 Missingd: gestational age <23  weeks72/317
 Missingd: birthweight missing or birthweight and gestational age missing150/38
Postnatal age (days)
 3 to <1082/11 317 (7.25)2.09 (1.32–3.31), 0.002
 10 to <2071/9674 (7.34)2.12 (1.33–3.38), 0.002
 20 to <3028/6096 (4.59)1.32 (0.76–2.29), 0.319
 30 to <4022/3934 (5.59)1.61 (0.90–2.88), 0.108
 40 to <5010/2601 (3.84)1.11 (0.53–2.32), 0.792
 ≥5023/6596 (3.49)1
Inborn/outborn status
 Outborn31280/7571 (10.57)2.21 (1.68–2.89), <0.001
 Inborn1932154/32 476 (4.74)1
 NICU 21329129/21 281 (6.06)1.08 (0.84–1.41), 0.544
 NICU 1940107/18 937 (5.65)1
 Male1268130/21 015 (6.19)1.12 (0.86–1.45), 0.413
 Female1001106/19 203 (5.52)1
Delivery method
 Emergency CS73592/15 531 (5.92)0.97 (0.73–1.29), 0.824
 Elective CS36033/6789 (4.86)0.79 (0.53–1.17), 0.237
 Vaginal1159110/17 856 (6.16)1
Number of blood samples takenb
 ≥22/388 (5.15)0.78 (0.19–3.24), 0.734
 127/6169 (4.38)0.69 (0.46–1.02), 0.060
 0207/33 661 (6.15)1
 Yes5/935 (5.35)0.89 (0.38–2.10), 0.796
 No231/39 283 (5.88)1
For NICU 2 only
 Total parenteral nutritionb
  Yes72/3375 (21.33)6.50 (4.53–9.33), <0.001
  No57/17 906 (3.18)1
  Yes75/7093 (10.57)2.80 (1.99–3.93), <0.001
  No54/14 188 (3.81)1
Table 3.   Associations between risk factors and bloodstream infection (BSI) rates: adjusted Poisson regression models
Potential risk factorFor NICU 1 and NICU 2 combined, optimal risk adjustment model
Adjusted rate ratio (95% CI), p-value
For NICU 2 only, risk adjustment model incorporating total parenteral nutrition
Adjusted rate ratio (95% CI), p-value
  1. aIn the previous 3 days.

  2. bDays/babies with missing variables were few in number, and represented few episodes of BSI. For this reason, we considered it acceptable to remove them from the analyses.

Highest level of carea
 Intensive care3.15 (2.01–4.94), <0.001
 High-dependency care6.58 (4.18–10.36), <0.001
 Special care1
Birthweight (g)
 <7003.69 (2.37–5.74), <0.0012.66 (1.58–4.48), <0.001
 700 to <12001.60 (1.09–2.35), 0.0161.19 (0.71–1.98), 0.508
Postnatal age (days)
 3 to <102.79 (1.64–4.74), <0.0013.13 (1.45–6.74), 0.004
 10 to <202.94 (1.78–4.83), <0.0012.04 (0.90–4.63), 0.086
 20 to <301.93 (1.10–3.39), 0.0231.73 (0.72–4.15), 0.218
 30 to <402.15 (1.19–3.89), 0.0111.98 (0.78–5.04), 0.153
 40 to <501.42 (0.68–2.96), 0.3580.26 (0.03–2.33), 0.229
Inborn/outborn status
 Outborn1.51 (1.12–2.04), 0.0071.58 (1.06–2.35), 0.024
Total parenteral nutritiona
 Yes4.30 (2.63–7.04), <0.001

Total parenteral nutrition was the second strongest risk factor for BSI. In NICU 2, total parenteral nutrition accounted for 16% (3375/21 281) of NICU days and 56% of BSIs (72/129) (Table 2). The separate model including total parenteral nutrition is shown in Table 3. When data for NICU 2 only were used, the multivariable model including total parenteral nutrition (QIC 1450) did not fit the data as well as a multivariable model incorporating level of care (QIC 1323; full results not shown). Ventilation was associated with an increased BSI risk in the crude analysis, but this effect was attenuated by adjustment for birthweight, inborn/outborn status, and postnatal age (adjusted rate ratio 1.30, 95% CI 0.81–2.11, p 0.277).

BSI risk was highest in the most premature and in term babies, with preterm babies born in the third trimester having the lowest risk. Babies with birthweights below 1200 g were at higher risk than heavier babies (Table 2). The optimal adjusted model retained birthweight as an independent risk factor for BSI (Table 3).

We performed a sensitivity analysis with birthweight ≥1200 g split into three categories (1200 to <2500 g, 2500 to <3500 g, and ≥3500 g). In the multivariable analysis, variation in risk in babies with birthweights above 1200 g diminished after adjustment for level of care, inborn/outborn status, and postnatal age. The best-fitting model grouped babies with birthweights above 1200 g into one category (Table 3).

Babies born at another maternity unit and transferred to the NICU (outborn) had a higher BSI risk than inborn babies. BSI risk was highest between days 3 and 20 of life in crude and adjusted analyses (Tables 2 and 3).

No associations were found between BSI and birthweight standardized for gestational age, NICU, sex, delivery method, number of blood samples taken, or surgery.

The same risk factors were found with a case–control method (see Supporting Information, Tables S1 and S2).

We also found that the same factors predicted BSI when only non-CONS BSI episodes were included, albeit with a loss of statistical confidence, owing to the small number of episodes (n = 60). For both hospitals combined, level of care was the strongest single risk factor for non-CONS BSI, in terms of optimizing the QIC (crude rate ratios against special care: high-dependency care, 9.36 (95% CI 3.90–22.45; intensive care, 5.77 (95% CI 2.67–12.44)). For NICU 2, total parenteral nutrition was the second strongest risk factor for non-CONS BSI, after level of care.


The BSI rate in NICU 1 and NICU 2 was 6/1000 baby-days. NHS level of care and total parenteral nutrition in the previous 3 days were the strongest single risk factors for BSI. The adjusted models combined these risk factors with birthweight, inborn/outborn status, and postnatal age. Similar results were found with an alternative, case–control method (see Supporting Information, Tables S1 and S2).

The BSI rate was similar to that reported by Holmes et al. for another tertiary London NICU (5.8/1000 baby-days). In the Poisson regression models, we used similar statistical methods to those of Holmes et al. [16]. However, we included level of care as a potential risk factor, and analysed a larger study population, spanning two NICUs. Level of care is a composite measure reflecting a baby’s vulnerability, invasive procedures, and intensity of care. All of these factors could explain the strong association between level of care and BSI. Level of care has not previously been explored in adjusted analyses. Holmes et al. [16] identified parenteral nutrition and gestational age below 26 weeks as the only significant independent risk factors for BSI, and recommended stratification by these factors for BSI monitoring. We found similar results for total parenteral nutrition, but a similarly strong association between BSI and level of care. Parenteral nutrition is primarily delivered in high-dependency care, which may explain the higher rate of BSI in high-dependency care than in intensive care.

The factors that should be used for risk adjustment depend on the clinical questions being addressed. For comparison of overall quality of care between NICUs, adjustment for daily level of care would be preferable, because it includes all babies across the full spectrum of risk. In contrast, only a minority (24%) [22] of NICU patients receive parenteral nutrition, and those who do not have widely differing risks of BSI. We therefore recommend that comparisons between NICUs should be based on overall BSI rates adjusted for level of care, birthweight, inborn/outborn status, and postnatal age. Additional stratification of BSI rates by parenteral nutrition may be useful for monitoring infection control interventions focused on parenteral nutrition.

One limitation of our analytical method is potential bias associated with the length of follow-up. To determine factors predicting infection, analyses are often restricted to the days preceding infection. However, this practice is prone to bias; if follow-up time is truncated at infection, exposures to NICU care differ systematically between infected and uninfected individuals [23]. A novel, alternative, case–control method with which to overcome this bias is presented in the Supporting Information, Tables S1 and S2. It identified the same predictive factors for BSI as the Poisson regression models. Generalizability of these results should be confirmed with a greater number of units, and with more recent data. Risk factors for BSI may have changed since our data were collected, because of staffing or administrative changes, for example.

Birthweight standardized for gestational age had no significant effect on BSI. This may be the result of a selection effect, as the babies most at risk of BSI, i.e. those who were very preterm and small for gestational age, were also those least likely to survive and be represented in the unit.

The disadvantage of our case definition is that some BSI episodes may have represented blood sample contamination or subclinical infection, rather than clinically important BSI. In paediatric intensive care units and NICUs, about 45% of blood cultures positive for CONS may reflect contamination [24]. However, we found no association between BSI and blood-sampling frequency, which would have been expected if a large proportion of BSI ‘episodes’ represented sample contamination. The advantage of our case definition is that the use of routine records would accelerate data collection and minimize staff workload to provide a sustainable approach for long-term monitoring. In contrast, case definitions including clinical observations require skilled data collection and stand-alone data systems, which can be time-consuming and expensive. Case definitions incorporating clinical symptoms may themselves differ in sensitivity and specificity, as the diagnosis of BSI is not clear-cut [6]. Case definitions based on routine data could exploit the current growth of routine datasets in medical care (http://www.neonatal.org.uk/SEND). If NICUs wish to differentiate between infections that are more or less likely to represent contamination whilst using routine data, risk adjustment and monitoring can be performed separately for CONS and non-CONS BSIs. We found that the same risk factors predicted non-CONS BSIs and all BSIs. Reporting the rates of CONS may help to address contamination itself, as false-positive blood cultures can lead to increased antibiotic use and longer durations of hospital stay.

Conclusion and Clinical Implications

We restricted our analyses to factors available in routine data, so the appropriate method of risk adjustment may vary and require more research if further potential risk factors, such as nurse/infant ratio, are routinely recorded in the future. On the basis of the current evidence, BSI rates should be adjusted for level of care, birthweight, inborn/outborn status, and postnatal age, to provide comparisons between NICUs. This approach is implementable in the NHS, as the required data are standardized and routinely recorded. Since our study, electronic patient records have become more widespread and standardized for NICUs across the NHS, and linkage to microbiology laboratory records is becoming easier [25]. With a modest investment in analytical support, our approach could be used to provide ongoing, risk-adjusted comparisons of infection rates in NICUs, without burdening clinicians with additional data entry.

Transparency Declaration

This work was undertaken at GOSH/UCL Institute of Child Health, which received a proportion of funding from the Department of Health’s NIHR Biomedical Research Centres funding scheme. The Centre for Paediatric Epidemiology and Biostatistics also benefits from funding support from the Medical Research Council in its capacity as the MRC Centre of Epidemiology for Child Health. P. Leighton was funded by the Medical Research Council. The study sponsors played no part in the study design, in the collection, analysis and interpretation of data, in the writing of the manuscript, or in the decision to submit the manuscript for publication. The authors have no relevant conflicts of interest to declare.