Electronic Nose Prediction of a Clinical Pneumonia Score: Biosensors and Microbes

Over a period of 4 months, 400 patients from a 52-bed surgical intensive care unit at the Hospital of the University of Pennsylvania were screened for study eligibility as defined below. This number was selected based on the assumption that 10% of the screened patients would be sampled, and a power analysis using an alpha of 0.05 and a P  of 0.8. We screen all mechanically ventilated patients for VAP as defined by Centers for Disease Control criteria.7With approval from the Institutional Review Board (which waived the need for consent), samples of exhaled breath were acquired from 38 randomly selected, mechanically ventilated patients and analyzed using a commercially available electronic nose (Cyrano Sciences, Pasadena, Ca).

Patients undergoing positive pressure ventilation via  endotracheal tube were eligible. Patients with respiratory distress, 10 cm H²O or greater of positive end-expiratory pressure, high levels of pressure support (>20 cm H²O), or respiratory rate >35 breaths/min were excluded. These exclusion criteria were designed to preclude sampling on patients with a significant degree of respiratory insufficiency, as the acquisition of six breath samples takes 30–45 min and patients with this degree of mechanical ventilatory support often require interventions that would interfere with uninterrupted data acquisition. In addition, flow-triggering was discontinued during sampling to maximize signal to noise ratio, and we therefore chose to restrict sampling to patients who were on relatively minimal ventilatory support.

All patients were given a Clinical Pulmonary Infection Score (CPIS) based on the one initially described by Pugin et. al.8and later modified by Luna et al.  9All patients were mechanically ventilated with positive pressure ventilation. For the period of sample acquisition, “flow by” (used either as a triggering adjunct or primary mode) on the patient ventilator was suspended and the heat and moisture exchanger was removed (if present). Breath samples were obtained from an access port in the expiratory limb of the ventilator circuit just distal to the Y piece attached to the patient’s endotracheal tube. The electronic nose was fitted with an Acrodisc (Pall, East Hills, NY) CR 25 mm syringe filter with a polytetrafluorethylene membrane. Approximately 30 ml of expired air were sampled over 30 s. Patients were sampled six times consecutively in a period of 30–45 min using the same procedure. The sample information was downloaded from the electronic nose to a computer for analysis.

The electronic nose used for this study is handheld. It includes a small pump, a battery power source, microcircuitry, and a sensor array. Each sensor in the 32-sensor array consists of a carbon black/polymer composite, in which the carbon black creates conducting pathways through the polymer. As the specific polymer composite interacts with the molecules in the sample volume, it swells, disrupting the carbon black conducting pathways to a variable degree and thereby altering the resistance of the sensor. Carbon black/polymer composite sensors are typically able to detect at a level of <0.1% of saturated vapor pressure (ppm to ppb).

The data were analyzed using linear partial least squares regression and nonlinear partial least squares regression (PLS toolbox for Matlab, Eigenvector Research Inc, Manson, WA) for constructing predictive models relating response variables to explanatory variables.10–12Partial least squares was originally developed for econometrics but is now used in chemometrics and industrial applications in which multivariate data are used to classify or predict outcomes or responses. This method of analysis is directly applicable and frequently used in chemometrics for the analysis of odors using an electronic nose, in which multivariate (from multiple vapor sensors within the e-nose), potentially colinear (because some of the sensors may track together) data are acquired and correlated with outcomes (i.e. , known odors or, in this case, known scores). Multiple linear regression is not appropriate for this analysis because the relationship between the predictive factors and the outcomes is not well understood and because the predictors have the potential to be interrelated (colinear). The nonlinear partial least squares method is a nonlinear classification technique that extends the partial least squares approach and remedies some of its shortcomings.

Partial least squares regression was used to relate CPIS scores to scores predicted by the e-nose and construct a Bland-Altman plot. Nonlinear partial least squares regression was used to determine the degree of correlation between the e-nose and CPIS scores. Cross-validation was used for both the partial least squares and nonlinear partial least squares CPIS prediction models to prevent over-fitting of the data. In cross-validation, the model relating actual outcomes to predicted outcomes is developed for all the pairs (of actual and predicted data) save one and then tested on that holdout case. This is repeated n times, with each case used as the validation case in turn.

Demographic information about the study patients are listed in table 1. The bacterial species cultured (using semiquantitative cultures) from tracheal secretions of patients with CPIS scores greater than 5 are listed in table 2. Culture data were recorded if a sample was acquired within 24 h of breath sampling, and therefore available in 12 of 19 patients with a CPIS score greater than 5. The actual CPIS scores in the patient population were normally distributed. The data generated using nonlinear partial least squares regression is plotted in figure 2. Each point represents the mean score for the six samples (±SEM) acquired during an individual sampling session. The plot shows the correlation between the pneumonia score and the score predicted by the electronic nose. Figure 3shows a Bland-Altman plot of the data and indicates a proportional deviation between the two measurements, which is also evident in figure 2.

Recent studies using gas chromatography and mass spectrometry have characterized some of the volatiles excreted from the body. Maple syrup urine disease, for example, is identifiable by a sweet aroma in the urine, which has been analyzed and determined to be attributable to sotolone.13The breath of schizophrenics has been found to contain volatile organic compounds including pentane and carbon disulfide.14,15Breath analysis has also been used as a noninvasive diagnostic test to detect Campylobacter pylori  in the stomach.16Pathologic processes in the distal gut have been diagnosed by breath analysis.17–19Various breath tests have been used to evaluate hepatic function.20–22Vapor pressure has been used as a method of measuring sweat osmolality in screening for cystic fibrosis.23There is a substantial literature regarding breath analysis in the occupational health field.24–32Finally, previous research has also shown alterations in the pentane/isoprene ratio in the expired gas of critically ill patients with adult respiratory distress syndrome and pulmonary infection.33,34 

Although exhaled or expressed volatiles contain information pertaining to pathophysiologic processes,35–37gas chromatography and mass spectroscopy are difficult to bring to the bedside. There are a variety of other sensor technologies in current generation electronic noses.38–40Some of them are commercially available and deployed, whereas others are being developed. They vary in their sensitivity and response to different odors.

The initial application of an electronic nose to the analysis of exhaled gases in intensive care patients was done using an e-nose consisting of an array of 32 intrinsically conducting polymer sensors. In this study, we used an e-nose with a sensor array consisting of the carbon black/polymer composites described above. There are several advantages to the use of this sensor technology. The first is that it is handheld, whereas that of the conducting polymer array e-nose is table-mounted (as are most of the other commercially available e-noses). The smaller footprint of the handheld device makes direct aspiration of exhaled gas into the e-nose possible, thereby obviating problems associated with transport of a gas sample from the bedside to the device. This type of sensor is stable and performs better in the presence of water vapor than the intrinsically conducting polymer e-nose array, permitting more reliable performance in the analysis of humid samples, such as expired gas from mechanically ventilated patients.

VAP occurs in 10–25% of patients ventilated for longer than 48 h, and it is the most commonly acquired infection in intensive care patients according to some studies.41Most studies show an association between VAP and increased mortality. One commonly employed approach to the clinical diagnosis of VAP is based on a “pneumonia score,” with or without invasive procedures such as bronchoscopy.42,43 

Chemometrics is the field of extracting information from multivariate chemical data as in odor analysis. Patterns of association can exist in data sets, but relationships among samples can be difficult to discover when the data matrix exceeds three or more features. Exploratory data analysis can reveal hidden patterns in complex data by reducing the information to a more comprehensible form. It can thereby reveal outliers and indicate whether there are patterns or trends in the data. Algorithms such as principal component analysis are designed to reduce large complex data sets into a series of optimized and interpretable views. These views emphasize the natural groupings in the data and show which variables most strongly influence those patterns.

The goal of chemometric regression analysis, the statistical approach used in this work, is to develop a calibration model that correlates information in a set of known measurements (the e-nose sensor responses) to a standard (the CPIS score). Commonly used algorithms for performing chemometric calibration include partial least squares and nonlinear partial least squares regression, and they are designed to eliminate problems associated with noise and colinear sensor responses. Because these regression algorithms are based in factor analysis, the entire group of known measurements is considered simultaneously, and information about correlations among the variables is automatically built into the calibration model.

There are a number of weaknesses in this study that preclude the immediate adoption of e-nose analysis as a diagnostic approach for VAP. The e-nose provides an aggregate response of the sensors to the chemical analytes, but we have not determined what those analytes are or whether they bear any relationship to the disease under study. The e-nose may smell bacterial metabolites, gases emitted during the bacteriocidal immune response, or a combination thereof. The specific molecules responsible for the sensor response pattern were not analyzed. It is clearly desirable to better characterize the molecular signals using gas chromatography and mass spectroscopy, both to enhance our understanding of the exhaled volatiles in lung infection and to develop a sensor array tuned to those volatiles. It is also unclear whether the “smell” of nonpathologic bacterial colonization of the trachea is the same as that of true VAP.

The design of this study only permits the conclusion that the e-nose sensor responses correlate with the CPIS score. Further studies, in which a “trained” or calibrated e-nose is used to prospectively assign a CPIS score (which is then compared to the actual CPIS), will be necessary to determine whether the e-nose can be used in a clinical setting as a diagnostic adjunct. Additional studies will be necessary to determine whether the e-nose can track bacterial load (i.e. , the quantity of bacteria derived from a culture) or type.

The breath of mechanically ventilated patients is well suited to diagnostic analysis, and this data suggests that an electronic nose may be capable of distinguishing between the infected and uninfected states. The demonstration of a correlation between a clinical pneumonia score and a pneumonia score based on the e-nose’s sensor response provides the possibility that a sufficiently sensitive and specific e-nose could be used as a univariate, noninvasive, concurrent measurement for the diagnosis of VAP. Although the CPIS score has not been shown to be diagnostic of VAP, it has been shown to be sensitive,44and it is conceivable that a high predicted CPIS score from the e-nose could be used as a trigger for noninvasive (tracheal aspirate) or invasive (bronchoscopic) diagnostic studies. One can imagine, for example, the use of a probability of infection score derived from e-nose monitoring in the management of critically ill mechanically ventilated patients. Further study will be needed to determine whether the electronic nose can be used to detect the onset of VAP, follow its response to treatment, or distinguish among infectious etiologies.