Cerebrospinal fluid cell count variability is a major confounding factor in external ventricular drain-associated infection surveillance diagnostics: a prospective observational study

This prospective observational single-center study with allocation, included adult (18 or older) patients treated with EVDs at the neurointensive care unit (NICU) at the Karolinska University Hospital in Stockholm. Exclusion criteria were patients admitted with bacterial or viral CNS infections. The data collection process is shown in Fig. 1.

Two CSF samples (paired samples) were collected biweekly or during workup due to suspected infection as per routine. The first eight patients were considered burn-in patients in order to familiarize the staff with the study and to identify any problems related to the sampling routine. All paired samples were allocated individually to two serially drawn samples performed before and after a standard clinical repositioning (repositioned samples, R-samples), or a 10-min wait period between two serially drawn samples (control samples, C-samples). As such, individual patients could have both R-samples and C-samples. For R-samples, patients were positioned in a lateral side position two hours prior to sample 1. After the first sample, patients were repositioned to the contralateral side after which the second sample was collected. For C-samples, patients were stationary 2 h prior to sample 1 and 10 min separated the two samples, corresponding to the average time between R-samples, as identified during the burn-in period. Non-sedated and compliant patients were asked to lie in a lateral side position prior to the first sample in applicable cases. Open drains were closed 5 min prior to the first sample and remained so between samples. 1.5 mL of CSF was drawn and discarded prior to each sample. After the burn-in period of eight patients with only R-samples, all patients were allocated to R-samples or C-samples for each sample session. CSF was analyzed for cell counts, lactate, glucose, and albumin by the Karolinska University Hospital Laboratory. The laboratory utilizes machine cell counting using the Sysmex XN10 system. Sampling was performed by a registered nurse. Information on primary diagnosis, length of stay, sex, Glascow Coma Scale (GCS), GCS motor response, and antibiotic treatment were collected. The intraventricular blood volume was measured from CT scans using the MPR volume tool in the Sectra IDS7 system (version 21.1.1.1997) as well as estimated using the intraventricular hemorrhage score (IVHS) by Hallevi et al. [29]. The shortest two-dimensional distance between the tip of the EVD catheter and intraventricular blood was measured, as well as the depth of blood sediment in the dorsal horns of the lateral ventricles. The measured IVH volume, IVHS, the two-dimensional catheter distance, and the sediment depth were measured or estimated using the earliest CT scans after EVD insertion. For IVHS calculation see Additional file 2.

The cell count components of the Karolinska University Hospital (earlier Swedish national guideline) EVDI definitions were used to denote suspected EVDI status for each individual sample as suspected EVDI (SI) or no EVDI (NI) to identify clinical implications of any observed CSF cell count variability. The cell count component of our local guidelines is one of several metrics used to identify suspected infections and guide antibiotic treatment in wait of culture results. The cell count component of the new current Swedish national guidelines was also tested for the same purpose and is included below [30].

Derived variables are defined as below. \(\Delta\) is used to denote the difference between paired samples (sample 1 and sample 2) for all CSF parameters. The absolute value (nonnegative) is denoted by two vertical lines (|x|).A positive \(\Delta\) indicates an increase from sample 1 to sample 2.The relative cell count \(\Delta\) is presented as percentage (IQR).A positive and negative \(\Delta\) will here have the same number.

Continuous variables were described as median, interquartile range (IQR), and percentage. The LE ratio was calculated for samples with cell count levels above the lower limit of the Sysmex XN10 system (Additional file 1), as measurement imprecision could produce false positive changes in the LE ratio at lower cell counts. The nonparametric Clustered Wilcoxon Signed Rank test was used to compare the median location shift for each CSF parameter between sample 1 and 2 [31, 32], clustered by Patient ID. The nonparametric Fligner-Killeen test was used to compare the variance of the derived variables between R-samples and C-samples. Mixed effects linear regression analysis was used to identify correlations between CT-derived variables and the pair-wise differences in a univariate setting with Patient ID as a random effect with random intercept, but not slope. Mixed effects logistic regression analysis was used to evaluate the effect of repositioning on pair-wise differences and to identify variables that were correlated to changes in suspected EVDI status between samples, also with Patient ID as random intercept but not slope. Outliers were identified based on Cook’s distance and removed. Fifteen R-samples and 5 C-samples had CSF cell counts below the functional sensitivity of the Sysmex XN10 system and were excluded from the mixed effects regression analyses and LE ratio calculations. Thus, 104 R-samples and 38 C-samples had the LE ratio calculated and were included in the mixed effects regression analyses.

All data were analyzed and allocation was performed using computer generated randomization in the statistical program R version 4.0.3. Mixed effects regression analyses were performed using R package lme4 v1.1-21. A p-value of \(p<.05\) is considered significant in regression analyses. In tables with multiple comparisons, p-values of \(p<.01\) are bolded.

The mean age of the cohort was 61 years and 61% were female (Table 1). The dominant NICU admission diagnosis was subarachnoidal hemorrhage (80%). Sample groups and numbers are depicted (Fig. 1). One hundred sixty-two paired samples were included in the final analysis, 119 R-samples and 43 C-samples. The difference between the two groups was in part attributable to the burn-in period of eight patients and differences in length of stay between patients, contributing to large interindividual differences in the number of samples per patient (Fig. 2).

Substantial pair-wise differences were seen in both R- and C-samples for granulocytes (absolute \(\Delta\) IQR [3, 95.5] and [0, 8.5], respectively) and erythrocytes (absolute \(\Delta\) IQR [900, 20,050] and [250, 4950], respectively), where the variability of drawn samples suggests a potential to confound early identification of EVDIs (Table 2.1). The IQR does not clearly reflect the large variability seen in this study as many patients had low cell counts. While we expected cells to be sedimented and affected in a positive direction, changes were both negative and positive in equal amounts (bidirectional), resulting in no significant differences in mean or medians between paired samples on a cohort level (\(p > 0.05\), clustered signed rank) (Table 2.2), with the exception of albumin. This suggests a heterogeneous sample space.

The variance of granulocytes (range R-samples [− 3505, 619] and C-samples [− 77, 757]) and erythrocytes (range R-samples [− 271,500, 331,700] and C-samples [− 47,500, 121,100]) was significantly greater (\(p<.01\), Fligner Killeen) (Table 2.1, Fig. 3) in the R-samples, suggesting that movement affect variability, again bidirectional. The correlation between repositioning and cell count \(\Delta\)’s and derivatives were evaluated with logistic regression mixed model analyses with Patient ID as a random effect (Table 3). The absolute \(\Delta\), but not the \(\Delta\) or relative \(\Delta\), were significantly related to repositioning for all variables, supporting that the variance is greater in the repositioned group, but that the direction of change is random (\(p<.05\), Table 3). To summarize, the exacerbated CSF cell count variability seen in the R-samples was significantly correlated to repositioning of the patient in a mixed effects setting (Table 3), and not merely an effect of sample group heterogeneity.

Cell counts varied more compared to glucose, lactate, and albumin, which exhibited an intermediate level, suggesting that molecular weight or size may also influence variability (Fig. 4). Here, the paired LE ratio is more highly correlated than erythrocytes and leukocytes individually, suggesting that changes may be in concert between cell types. However, the correlation between paired samples of monocytes and granulocytes is stronger than that of leukocytes and erythrocytes (Additional file 3), suggesting that cell types are differentially distributed in the sample space. Pair-wise variability increased with increasing levels in sample 1 for cells in particular (Fig. 5), but not glucose and lactate, consequently cells showed larger variability at levels used for clinical EVDI cutoffs for suspected infection. In conclusion, CSF cell counts varied significantly and could not be reproduced in serially drawn CSF, where R-samples exhibited greater variability.

When the Stockholm EVDI cell count criteria for suspected infection were applied on our cohort as to evaluate changes of suspected EVDI status (suspected infection, SI, or no infection, NI) between paired samples, 14 R-samples from 13 unique patients changed EVDI diagnostic group, corresponding to 30% of all patients with collected R-samples in our cohort. Eleven R-samples changed from NI to SI, and 3 R-samples from SI to NI. One patient had changes in both directions with one paired sample that changed from NI to SI, and one that changed from SI to NI. No changes in suspected EVDI status were seen for any C-samples. Two patients were initiated on broad-spectrum antibiotics due to significant pleocytosis in the second sample. No patient had a positive CSF culture confirmed EVDI in this study.

Repositioning as a cause of suspected EVDI group change could not computationally be evaluated with logistic regression as all positive events occurred solely for R-samples. Repositioning was, however, significantly correlated with suspected EVDI group change in a chi-squared test (\(p=.042\)) suggesting that cell changes in regions near diagnostic cutoffs may be influenced by patient movement. The cell count criteria from the new national EVDI guidelines applied to the sample cohort resulted in 18 R-samples and 2 C-samples from 16 unique patients changing diagnostic status (\(p=.06\)) in logistic regression. No random effects were seen for patient ID. No other studied parameter was found significantly correlated with changes of suspected EVDI status (Table 4). In summary, 30% of all patients with R-samples changed diagnostic EVDI status at least once during the study period, indicating that EVDI surveillance diagnostics based on cell counts are neither sensitive nor specific in a clinical setting.

Catheter distance, sediment depth, measured blood volume, and IVHS were calculated or estimated based on the first CT scan after EVD insertion. The correlation between all CT-derived variables and the observed \(\Delta\)’s and derivatives for all CSF parameters were evaluated in a univariate mixed effects regression model with Patient ID as a random effect. IVH volume was not seen significantly related to any CSF variable. Other CT parameters appear most highly related to erythrocyte-derived variables (Additional file 4), but not the direction of change. However, no significance here withstands adjustment for multiple testing and the explanatory value is found limited with no pseudo \(R^2\) above 0.0635 for the fixed effects.

There are several limitations to this study. This is a single-center clinical trial and the external validity of this study may need further validation. As different centers will have different treatment regimes, routines, and algorithms concerning EVD care and EVDI surveillance diagnostics, findings may be to some extent be population and location dependent. Moreover, we have not studied the time component of cell change, where the aseptic inflammation component might be expected to increase over time, and heterogeneity might itself be related to the inflammatory process. However, as the risk of true infection also increases over time, late variability would also be expected to impact diagnostics. Additionally, our study cohort is 80% SAH patients, and power will not permit if there is a disease-related component to variability. As, we have no positive cultures we provide no information on if heterogeneity is different in true infection. Finally, our sample groups are unbalanced, mainly due to a burn-in period where all patient samples were allocated around repositioning, with by chance long ICU stays. We have attempted to account for this using mixed model analyses.