Online Supplement
VAP and headspace analysis
Volatile Metabolic FingerprintofVentilator−associated Pneumonia:
A case-control study
Lieuwe D.J. Bos, Ignacio Martin−Loeches, Janine B. Kastelijn, Gisela Gili,
Mateu Espasa, Pedro Povoa, Arend H.J. Kolk, Hans–Gerd Janssen,
Peter J. Sterk, Antonio Artigas, Marcus J. Schultz
Supplemental Methods
Design and setting
This is a study within an international multi–center prospective observational cohort that evaluated the predictive value of biological markers for development of VAP. The study protocol was reviewed and approved by the Medical Ethical Committee of Parc Tauli, Sabadell, Spain (IRB: 2008/524).
Inclusion and exclusion criteria
Inclusion criteria were: (1) recruited to one of the participating ICUs, (2) intubated and ventilated for another reason than pneumonia, and (3) expectation that mechanical ventilation was needed for longer than 48 hours. Exclusion criteria were: (1) age less than 18 years, (2) expectation that withdrawal of treatment could happen within 72 hours, and (3) pregnancy or lactation. Furthermore we excluded patients who received antimicrobial therapy within the last 5 days before ICU admission (prophylactic antimicrobial therapy, e.g., for surgical procedures, was allowed), and patients with fulminant hepatic failure, pancreatitis, or disseminated cancer, since these all could effect levels of biomarker of interest in the original study. Finally, for the purpose of the present analysis we also excluded patients who developed pneumonia within the first two days in the ICU.
Diagnostic definitions and patient selection
Patients were classified into three groups (table 1).VAP was diagnosed using consensus criteria [1] (a new and persistent radiographic infiltrateplus at least 2 of the following criteria: a) temperature >380C or <360C; b) leucocytes >10 or <4 x 103/mm3; c) purulent tracheal aspirate [2, 3]) but always needed microbiological confirmation to fulfill the diagnosis of ‘VAP’ (> 103 or ≥ 106 colony forming units (CFU)/ml in mini-bronchoalveolar lavage (BAL) fluid or tracheal aspirate (TA), respectively). Patients not fulfilling the abovementioned criteria for VAP but of whom microbiological culture revealed presence of bacteria in mini-BAL or TA were classified as ‘colonized patients without VAP’. Patients not fulfilling the abovementioned criteria for VAP with negative cultures were classified as ‘non–colonized patients without VAP’.
Data and sample collection
Patient demographics, primary (and admission) diagnosis, SAPSII[4], APACHE II score[5] and ICU mortality were recorded for all patients. Tracheal aspirates (TA) were collected on Mondays, Wednesdays and Fridays and were sent for quantitative culture. Mini-BAL was performed on the day of clinical suspicion of VAP. Isolates were characterized by colony morphology and Gram stains. The remaining portions of TA samples were saved at −800C (figure S1, bullet 1).
Headspace analysis
TA fluid samples were defrosted and subsequently analysed using an eNose (figure S1, bullet 2). For this, 0.5 ml supernatant (1500 rpm for 15 minutes at 40C) of TA fluid was transferred into a 5ml headspace vial (MN–net N 20–5 DIN, clear with crimp top, Fisher Scientific, Landsmeer, the Netherlands), topped with a 20mm headspace cap (Fisher Scientific, Landsmeer, the Netherlands) and warmed to room temperature. Two needles were inserted into the headspace cap before start of each measurement. One needle was placed into the sampleand was purged with pure nitrogen gas (99.9999%, Linde Gas, Dieren, the Netherlands). The second needle was placed in the headspace for collection of gas (figure S2).
We used the Cyranose 320 eNose (Smith Detections, Pasadena, CA) containing a nano–composite sensor array with 32 polymer sensors (figure S1, bullet 3)[6]. This electronic nose relies on semi-selective recognition of VOCs as each sensor is cross-reactive to a variety of functional chemical groups. The combined response of the sensor array can be used for pattern-recognition and disease classification [7]. A full description of the chemical and statistical rationale for this approach is outside the scope of the present paper and has recently been discussed in an excellent review by Konvalina and Haick in Accounts of chemical research and is summarized in the online supplement [8]. Nitrogen gas was sampled for 30 seconds as a baseline-measurement, whereafter gas from the headspace was analysed for one minute using a low flow (40ml/min). Afterwards, the eNosewas purged to let the sensors recover to baseline. This was done in duplicate.
Electronic nose
Electronic noses (eNose), named after their similarities with mammalian olfactory system[6], integratively capture complex VOC mixtures using an array of different sensors [6]. Sensors can be utilized through two mutually exclusive routes: (1) very specific sensors that follow a “lock-and-key” principle that have a very high sensitivity and specificity or (2) semi-selective sensors that are less sensitive and less selective but an array of which can be used to characterize unknown complex samples [8]. Since the VOCs that should be detected to classify disease most accurately are unknown and most biological samples are highly complex, the latter approach is currently most frequently used.
Metal oxides, conducting polymers, optical and infra–red spectroscopy have been used as sensors. Peaks and intensities obtained by mass-spectrometry can also be presented to pattern–recognition algorithms, hereby virtually converting every detected mass into a “sensor”[9]. In this study, we used the Cyranose 320 eNose (Smith Detections, Pasadena, CA) containing a nano–composite sensor array with 32 polymer sensors [6]. This electronic nose relies on semi-selective recognition of VOCs as each sensor is cross-reactive to a variety of functional chemical groups.
The composite signal of all sensors in an array can be analyzed using pattern–recognition algorithms. The composite signal of eNose analysis results in a unique fingerprint. Subsequently, these fingerprints can be used for diagnostic and monitoring purposes[9].Several types of algorithms have been used in the literature, but most involve a form of dimension reduction followed by a classification algorithm. Unsupervised dimension reduction methods such as principal component analysis are frequently used [10], but are easily disturbed by, for example, co-morbidities to capture other variation in the data than the disease of interest [11]. Therefore, other groups have relied on supervised methods for dimension reduction and classification [12], which can be combined in partial least square regression [13]. The major advantages of both principal component analysis and partial least square regression is that they perform well when the predictor matrix has more variables than observation (hence the dimension reduction) and when there is multicollinearity (as with semi-selective sensors). In this study, we used an adapted version of partial least square regression where sparsity is added, removing sensors with limited added value from the algorithm (putting the regression coefficients to 0). Sparse partial least square regression has been shown to outperform other methods when there are less observations than predictors (in this case sensors) [14].
Statistical analysis
Differences between the groups were compared using the Mann–Whitney U test for continuous variables and chi–square for categorical variables. Data were summarized using the median and inter-quartile range for continuous variables and with count and percentage for categorical variables. All analyses were performed in R statistics using R studio [15]. P–values below 0.05 were considered significant.
Samples from patients with VAP were to be compared in time to those from patients who did not develop VAP. For this, we first determined the day the diagnostic criteria for VAP were met in patients who developed VAP. This was day 7 after ICU admission. We then selected patients who did not develop VAP, but who were still intubated and ventilated on day 7 after ICU admission to create the control group.To train the diagnostic algorithm ‘VAP’ patients were considered cases and both ‘colonized patients without VAP’ and ‘non–colonized patients without VAP’ were taken together as controls.
Sparse partial least square (SPLS) logistic regression was used to produce a diagnostic model (figure S1, bullet 4). SPLS analysis is a suitable form of regression that can select predictive variables and limit false discovery in situations were large number of independent variables are investigated in low numbers of individuals [14]. Two parameters are set in SPLS: K and eta. These were tuned using 10-times cross–validation [14]. SPLS logistic regression resulted in a predicted probability of VAP, which is called the “eNose signal” in this manuscript and was used for further analyses. Receiver operator characteristics (ROC) analysis was performed and the area under the curve (AUC) was reported (figure S1, bullet 5).
A sensitivity analysis was performed for ‘colonized patients without VAP’ and ‘non–colonized patients without VAP’ (figure S1, bullet 6). Secondly, we investigated the correlation between the eNose signal and to the number of colonizing forming units using Spearman’s correlation. Thirdly, the diagnostic performance of the eNose logistic regression function was compared to the CPIS and the net classification improvement of the combination of the two diagnostic tests was assessed [16]. Finally, the development of the eNose signal was displayed over time using a LOESS smoother and analyzed by means of mixed-model analysis. The period preceding VAP (7 days before diagnosis) was investigated (figure S1, bullet 7). The slope of the eNose signal was fitted using a linear mixed model with a random intercept and a random slope for the eNose signal per patient, as described previously [17]. The slopes were compared between patients with and without VAP and the area under the ROC-curve was calculated.
Supplemental Results
Patients
Figure S3 shows the CONSORT diagram of the study; of 154 eligible patients, 10 development of pneumonia within 48 hours, and from 27 patients too much data was missing. Of the remaining 117 patients 14 patients fulfilled the diagnostic criteria for VAP (on median day 7); 103 patients never met the diagnostic criteria for VAP, and of them only 31 patients were intubated and ventilated for longer than 7 days. These patients served as controls in the planned analysis. Of the patients without VAP, 14 had colonized airways. Table 1 shows the patient characteristics and microbiological data. There were no missing physiological data (ventilator setting, hemodynamics, temperature) and follow–up for mortality was available for all patients in the database. None of the patients were treated with inhaled antibiotics or systemic corticosteroids.
Airway samples
175 airway samples were obtained. The median number of samples per patient was 4 [3 – 6] and 2 [2 – 3] in the VAP group and the groups of patients without VAP, respectively. One sample per patient was used for the cross-sectional statistical analysis (for patients with pneumonia: sample on day of diagnosis, for patients without pneumonia: sample obtained closest to day 7). The median day of sample collection was 7 [5 – 9] and 5 [4 – 8] for patients with VAP and patients without VAP, respectively.
Discrimination betweensamples of patients with VAP and without VAP
SPLS analysis of eNose data (K = 4, eta = 0.8) resulted in the selection of 26 sensors (for coefficients, see Table S2). Using the eNose model, VAP (median predicted probability: 0.53 [0.44 – 0.90]) and patients without VAP (colonized and non-colonized) (median: 0.14 [0.10 – 0.23]) could be discriminated with a ROC–AUC of 0.85 [CI: 0.69 – 1.0] (table 2).
Sensitivity analysis for colonization in patients without VAP
The eNose predicted probability was not different for patients without VAP, with and without colonized airways: 0.12 [0.09 – 0.22] and 0.14 [0.12 – 0.24], respectively (p = 0.42). VAP was well distinguished from non–colonized and from colonized patients without VAP: ROC–AUC of 0.84 [CI: 0.68 – 1.0] and 0.85 [CI: 0.68 – 1.0], respectively (table S3).
Correlation with bacterial growth
Bacterial growth was associated with a higher eNose signal (Spearman’s correlation coefficient: 0.56, p<0.001; figure S2). When the analysis was limited to samples with bacterial growth, the amount of CFU was not correlated with the eNose signal (Spearman’s correlation coefficient: 0.22, p=0.36, figure S4)
Comparison and combination with CPIS
The ROC–AUC for the CPIS by itself was 0.89 [CI: 0.80 – 0.99] for ‘VAP’ (table S3). The ROC–AUC was 0.95 [CI: 0.88 – 1.0] after combination of CPIS with the eNose algorithm (table S3). The net classification improvement [16] was 1.4 [CI: 0.80 – 2.07, p < 0.001] for ‘VAP’.
Longitudinal analysis
Figure S5 shows the development of the eNose signal during ICU-stay. The slope of the eNose signal was 0.04 [0.02 – 0.05] for patients with VAP and 0.002 [-0.02 – 0.01] for patients without VAP (p=0.009). Using the slope of the eNose signal VAP and patients without VAP (colonized and non-colonized) could be discriminated with a ROC–AUC of 0.76 [CI: 0.56 – 0.96] (table S3).
Supplemental Tables
Table S1: Patient characteristics
No VAP (n=31) / VAP (n = 14) / P−value*non–colonized
airways
(n = 17) / colonized
airways
(n = 14)
Age, yrs N (%) / 61 (33) / 66 (27) / 59 (31) / 45 (24) / 0.20
Male N (%) / 15 (48) / 7 (41) / 8 (57) / 7 (50) / 0.89
APACHE II median (IQR) / 20 (10) / 20 (10) / 20 (13) / 24 (7) / 0.11
SAPS II median (IQR) / 55 (20) / 50 (15) / 60 (19) / 52 (27) / 0.48
COPD N (%) / 2 (6) / 2 (12) / 0 (0) / 2 (14) / 0.91
Reason for MV / Respiratory failure N (%) / 4 (13) / 3 (21) / 1 (8) / 3 (21) / 0.58
Shock N (%) / 4 (13) / 3 (21) / 1 (8) / 4 (29)
Low consciousness N (%) / 17 (55) / 8 (57) / 9 (75) / 6 (43)
Other N (%) / 6 (19) / 3 (21) / 3 (21) / 1 (7)
Microbiologic findings† / No N (%) / 17 (55) / 17 (100) / 0 (0) / 0 (0) / -
Escherichia coli / 2 (6) / 0 (0) / 2 (14) / 3 (21)
Pseudomonas aer. / 0 (0) / 0 (0) / 0 (0) / 0 (0)
Klepsiella pneu. / 1 (3) / 0 (0) / 1 (7) / 2 (14)
Haemophilus inf. / 0 (0) / 0 (0) / 0 (0) / 3 (21)
Acinetobacter car. / 0 (0) / 0 (0) / 0 (0) / 3 (21)
Other gram negative / 3 (10) / 0 (0) / 3 (21) / 4 (28)
Staphylococcus aureus / 6 (19) / 0 (0) / 6 (43) / 2 (14)
Candida sp. / 4 (13) / 0 (0) / 4 (29) / 0 (0)
Tidal volume (ml) median (IQR) / 446 (145) / 445 (108) / 447 (161) / 438 (89) / 0.40
Plateau pressure (cmH2O) median (IQR) / 20 (8) / 17 (6) / 23 (8) / 24 (8) / 0.40
Positive end expiratory pressure (cmH2O) median (IQR) / 5 (2) / 5 (2) / 5 (0) / 8 (4) / <0.001
Sepsis N (%) / 11 (35) / 7 (41) / 4 (29) / 11 (79) / 0.01
Septic Shock N (%) / 9 (29) / 6 (35) / 3 (21) / 9 (64) / 0.03
SOFA score (day measurement) median (IQR) / 19 (3) / 20 (3) / 19 (2) / 8 (16) / 0.06
CPIS (day measurement) median (IQR) / 3 (4) / 3 (5) / 3 (4) / 11 (2) / <0.001
WBC (x103) median (IQR) / 10 (4) / 10.4 (7.9) / 10.5 (2.5) / 14.1 (8.0) / 0.06
28 day mortality N (%) / 4 (13) / 3 (21) / 1 (8) / 6 (43) / 0.13
*: P-value for all patients without VAP vs. all patients with VAP. †: Multiple organisms could have been cultured per patient. APACHE: Acute Physiology and Chronic Health Evaluation, SAPS: Simplified Acute Physiology Score, MV: Mechanical Ventilation, SOFA: Sequential Organ Failure Assessment, CPIS: Clinical Pulmonary Infection Score
Table S2: Coefficients for eNose sensors
Sensor / CoefficientIntercept / -1.00
Sensor1 / -0.71
Sensor2 / 0.64
Sensor3 / 0.84
Sensor4 / -1.07
Sensor5 / -1.69
Sensor6 / -1.58
Sensor8 / 2.28
Sensor9 / 2.37
Sensor10 / 1.98
Sensor11 / -1.26
Sensor12 / -1.06
Sensor13 / 0.45
Sensor14 / 0.72
Sensor15 / 0.10
Sensor16 / -1.19
Sensor17 / 0.28
Sensor18 / 0.70
Sensor20 / 0.60
Sensor21 / -1.83
Sensor22 / -0.92
Sensor23 / 0.83
Sensor24 / 0.32
Sensor25 / 0.41
Sensor26 / 0.40
Sensor27 / -0.32
Sensor28 / -0.47
Table S3: Discrimination by electronic nose analysis
Measure / Cases / Control / ROC-AUC [95% CI] / Cut-off / Sens / Spec / PPV / NPVeNose / VAP / All controls / 0.85 [0.69 – 1.0] / 0.41 / 94% / 79% / 91% / 85%
VAP / Non-colonized controls / 0.84 [0.68 – 1.0] / 0.38 / 94% / 79% / 84% / 92%
VAP / Colonized-controls / 0.85 [0.68 – 1.0] / 0.41 / 93% / 79% / 81% / 92%
CPIS (logistic regression) / VAP / All controls / 0.89 [0.80 – 0.99] / 0.45 / 87% / 86% / 72% / 92%
CPIS + eNose / VAP / All controls / 0.95 [0.88 – 1.0] / 0.38 / 94% / 86% / 81% / 92%
eNose slope / VAP / All controls / 0.76 [0.56 – 0.96] / 0.034 / 89% / 75% / 89% / 75%
CI: Confidence interval, CPIS: Clinical pulmonary infection score, eNose: Electronic nose, PPV: Positive predictive value, NPV: Negative predictive value, Sens: Sensitivity, Spec: Specificity, ROC-AUC: Area under the receiver operating charactistics curve
Supplemental Figures
Figure S1: Methodology
Figure S2: Analysis of tracheal aspirate by electronic nose
Figure S3: Patient flow
Figure S4: Correlation between eNose signal and bacterial growth
Figure S5: Longitudinal signal
Supplemental References
1.Niederman M, Craven DE, Bonten M, (2005) Guidelines for the management of adults with hospital-acquired, ventilator-associated, and healthcare-associated pneumonia. Am J Respir Crit Care Med 171: 388-416
2.Heyland D, Dodek P, Muscedere J, Day A, (2006) A randomized trial of diagnostic techniques for ventilator-associated pneumonia. N Engl J Med 355: 2619-2630
3.Ewig S, Bauer T, Torres A, (2002) The pulmonary physician in critical care * 4: Nosocomial pneumonia. Thorax 57: 366-371
4.Le Gall JR, Lemeshow S, Saulnier F, (1993) A new Simplified Acute Physiology Score (SAPS II) based on a European/North American multicenter study. JAMA 270: 2957-2963
5.Knaus WA, Draper EA, Wagner DP, Zimmerman JE, (1985) APACHE II: a severity of disease classification system. Crit Care Med 13: 818-829
6.Röck F, Barsan N, Weimar U, (2008) Electronic nose: current status and future trends. Chem Rev 108: 705-725
7.Fens N, van der Schee MP, Brinkman P, Sterk PJ, (2013) Exhaled breath analysis by electronic nose in airways disease. Established issues and key questions. Clinical & Experimental Allergy 43: 705-715
8.Konvalina G, Haick H, (2013) Sensors for Breath Testing: From Nanomaterials to Comprehensive Disease Detection. Acc Chem Res
9.Bos LDJ, Sterk PJ, Schultz MJ (2012) Metabolomics in Critically ill Patients: Focus on Exhaled Air. In: Vincent JL (ed) Annual Update in Intensive Care and Emergency Medicine 2012. Springer
10.Fens N, Zwinderman AH, van der Schee MP, de Nijs SB, Dijkers E, Roldaan AC, Cheung D, Bel EH, Sterk PJ, (2009) Exhaled breath profiling enables discrimination of chronic obstructive pulmonary disease and asthma. Am J Respir Crit Care Med 180: 1076-1082
11.Fens N, Douma RA, Sterk PJ, Kamphuisen PW, (2010) Breathomics as a diagnostic tool for pulmonary embolism. J Thromb Haemost 8: 2831-2833
12.Machado RF, Laskowski D, Deffenderfer O, Burch T, Zheng S, Mazzone PJ, Mekhail T, Jennings C, Stoller JK, Pyle J, Duncan J, Dweik RA, Erzurum SC, (2005) Detection of Lung Cancer by Sensor Array Analyses of Exhaled Breath. Am J Respir Crit Care Med 171: 1286-1291
13.D'Amico A, Pennazza G, Santonico M, Martinelli E, Roscioni C, Galluccio G, Paolesse R, Di Natale C, (2009) An investigation on electronic nose diagnosis of lung cancer. Lung Cancer 68: 170-176
14.Chun H, Keles S, (2010) Sparse partial least squares regression for simultaneous dimension reduction and variable selection. J R Stat Soc Series B Stat Methodol 72: 3-25
15.R-Development-Core-Team (2010) R: A language and environment for statistical computing. In: Editor (ed)^(eds) Book R: A language and environment for statistical computing. City, pp.
16.Pencina MJ, D'Agostino RB, Sr., D'Agostino RB, Jr., Vasan RS, (2008) Evaluating the added predictive ability of a new marker: from area under the ROC curve to reclassification and beyond. Stat Med 27: 157-172; discussion 207-112
17.Povoa P, Teixeira-Pinto AM, Carneiro AH, (2011) C-reactive protein, an early marker of community-acquired sepsis resolution: a multi-center prospective observational study. Crit Care 15: R169