Currently, many different methods are being used for pre-processing, statistical analysis and validation of data obtained by electronic nose technology from exhaled air. These various methods, however, have never been thoroughly compared. We aimed to empirically evaluate and compare the influence of different dimension reduction, classification and validation methods found in published studies on the diagnostic performance in several datasets. Our objective was to facilitate the selection of appropriate statistical methods and to support reviewers in this research area. We reviewed the literature by searching Pubmed up to the end of 2014 for all human studies using an electronic nose and methodological quality was assessed using the QUADAS-2 tool tailored to our review. Forty-six studies were evaluated regarding the range of different approaches to dimension reduction, classification and validation. From forty-six reviewed articles only seven applied external validation in an independent dataset, mostly with a case-control design. We asked their authors to share the original datasets with us. Four of the seven datasets were available for re-analysis. Published statistical methods for eNose signal analysis found in the literature review were applied to the training set of each dataset. The performance (area under the receiver operating characteristics curve (ROC-AUC)) was calculated for the training cohort (in-set) and after internal validation (leave-one-out cross validation). The methods were also applied to the external validation set to assess the external validity of the performance. Risk of bias was high in most studies due to non-random selection of patients. Internal validation resulted in a decrease in ROC-AUCs compared to in-set performance: −0.15,−0.14,−0.1,−0.11 in dataset 1 through 4, respectively. External validation resulted in lower ROC-AUC compared to internal validation in dataset 1 (−0.23) and 3 (−0.09). ROC-AUCs did not decrease in dataset 2 (+0.07) and 4 (+0.04). No single combination of dimension reduction and classification methods gave consistent results between internal and external validation sets in this sample of four datasets. This empirical evaluation showed that it is not meaningful to estimate the diagnostic performance on a training set alone, even after internal validation. Therefore, we recommend the inclusion of an external validation set in all future eNose projects in medicine.

In critically ill patients, glucose control with insulin mandates time– and blood–consuming glucose monitoring. Blood glucose level fluctuations are accompanied by metabolomic changes that alter the composition of volatile organic compounds (VOC), which are detectable in exhaled breath. This review systematically summarizes the available data on the ability of changes in VOC composition to predict blood glucose levels and changes in blood glucose levels.

Introduction: There is a need for continuous glucose monitoring in critically ill patients. The objective of this trial was to determine the point accuracy and reliability of a device designed for continuous monitoring of interstitial glucose levels in intensive care unit patients. Methods: We evaluated point accuracy by comparing device readings with glucose measurements in arterial blood by using blood gas analyzers. Analytical and clinical accuracy was expressed in Bland-Altman plots, glucose prediction errors, and Clarke error grids. We used a linear mixed model to determine which factors affect the point accuracy. In addition, we determined the reliability, including duration of device start-up and calibration, skips in data acquisition, and premature disconnections of sensors. Results: We included 50 patients in whom we used 105 sensors. Five patients from whom we could not collect the predefined minimum number of four consecutive comparative blood draws were excluded from the point accuracy analysis. Therefore, we had 929 comparative samples from 100 sensors in 45 patients (11 (7 to 28) samples per patient) during 4,639 hours (46 (27 to 134) hours per patient and 46 (21 to 69) hours per sensor) for the accuracy analysis. Point accuracy did not meet the International Organization for Standardization (ISO) 14971 standard for insulin dosing accuracy but did improve with increasing numbers of calibrations and was better in patients who did not have a history of diabetes. Out of 105 sensors, 60 were removed prematurely for a variety of reasons. The device start-up time was 49 (43 to 58) minutes. The number of skips in data acquisition was low, resulting in availability of real-time data during 95% (89% to 98%) of the connection time per sensor. Conclusions: The point accuracy of a device designed for continuous real-time monitoring of interstitial glucose levels was relatively low in critically ill patients. The device had few downtimes, but one third of the sensors were removed prematurely because of unresolved sensor- or device-related problems. Trial registration: Netherlands Trial Registry number: NTR3827. Registered 30 January 2013.

Microdialysis is a well-established technology that can be used for continuous blood glucose monitoring. We determined point and trend accuracy, and reliability of a microdialysis-based continuous blood glucose-monitoring device (EIRUS®) in critically ill patients. Prospective study involving patients with an expected intensive care unit stay of ≥48 h. Every 15 min, device readings were compared with blood glucose values measured in arterial blood during blocks of 8 h per day for a maximum of 3 days. The Clarke error grid, Bland–Altman plot, mean absolute relative difference and glucose prediction error analysis were used to express point accuracy and the rate error grid to express trend accuracy. Reliability testing included aspects of the device and the external sensor, and the special central venous catheter (CVC) with a semipermeable membrane for use with this device. We collected 594 paired values in 12 patients (65 [26–80; 8–97] (median [IQR; total range]) paired values per patient). Point accuracy: 93.6 % of paired values were in zone A of the Clarke error grid, 6.4 % were in zone B; bias was 4.1 mg/dL with an upper limit of agreement of 28.6 mg/dL and a lower level of agreement of −20.5 mg/dL in the Bland–Altman analysis; 93.6 % of the values ≥75 mg/dL were within 20 % of the reference values in the glucose prediction error analysis; the mean absolute relative difference was 7.5 %. Trend accuracy: 96.4 % of the paired values were in zone A, and 3.3 and 0.3 % were in zone B and zone C of the rate error grid. Reliability: out of 16 sensors, 4 had to be replaced prematurely; out of 12 CVCs, two malfunctioned (one after unintentional flushing by unsupervised nurses of the ports connected to the internal microdialysis chamber, causing rupture of the semipermeable membrane; one for an unknown reason). Device start-up time was 58 [56–67] min; availability of real-time data was 100 % of the connection time. In this study in critically ill patients who had no hypoglycemic episodes and a limited number of hyperglycemic excursions, point accuracy of the device was moderate to good. Trend accuracy was very good. The device had no downtimes, but 4 out of 16 external sensors and 2 out of 12 CVCs had practical problems.

Introduction: Continuous breath analysis by electronic nose (eNose) technology in the intensive care unit (ICU) may be useful in monitoring (patho) physiological changes. However, the application of breath monitoring in a non-controlled clinical setting introduces noise into the data. We hypothesized that the sensor signal is influenced by: (1) humidity in the side-stream; (2) patient-ventilator disconnections and the nebulization of medication; and (3) changes in ventilator settings and the amount of exhaled CO₂. We aimed to explore whether the aforementioned factors introduce noise into the signal, and discuss several approaches to reduce this noise. Methods: Study in mechanically-ventilated ICU patients. Exhaled breath was monitored using a continuous eNose with metal oxide sensors. Linear (mixed) models were used to study hypothesized associations. Results: In total, 1251 h of eNose data were collected. First, the initial 15 min of the signal was discarded. There was a negative association between humidity and Sensor 1 (Fixed-effect β: −0.05 ± 0.002) and a positive association with Sensors 2–4 (Fixed-effect β: 0.12 ± 0.001); the signal was corrected for this noise. Outliers were most likely due to noise and therefore removed. Sensor values were positively associated with end-tidal CO₂, tidal volume and the pressure variables. The signal was corrected for changes in these ventilator variables after which the associations disappeared. Conclusion: Variations in humidity, ventilator disconnections, nebulization of medication and changes of ventilator settings indeed influenced exhaled breath signals measured in ventilated patients by continuous eNose analysis. We discussed several approaches to reduce the effects of these noise inducing variables.

Continuous glucose monitoring (CGM) can be beneficial in critically ill patients. Current CGM devices rely on subcutaneous or blood plasma glucose measurements and consequently there is an increased risk of infections and the possibility of loss of blood with each measurement. A potential method to continuously and non-invasively measure blood glucose levels is using exhaled breath. A correlation between blood glucose levels and volatile organic compounds (VOCs) in the exhaled breath was already reported. VOCs can be analyzed continuously using a so-called electronic nose (eNose). We hypothesize that continuous exhaled breath analysis using an eNose can be used to accurately predict blood glucose levels in intubated, mechanically ventilated ICU-patients. Mechanically ventilated patients whose blood glucose concentration was monitored with a CGM device were eligible. An eNose with four metal oxide sensors was used to continuously measure changes in exhaled breath. After pre-processing the data, several regression models were trained, consisting of: (1) only eNose sensor values; (2) only the 1st and 2nd principal components (PC) of eNose values; (3) eNose sensor values and last known blood glucose value as random effect; (4) 1st and 2nd PC of eNose sensor values and CGM value of one minute ago as fixed effect; (5) CGM value of one minute ago as fixed effect. Model performance was measured using the R ² value, the akaike information criterion and the Clarke error grid. Twenty-three patients were included in the study and 1165 hours of measurements were collected. Performance was low in models 1, 2 and 3 with a mean R ² of 0.07 [95%-CI: 0.00–0.28], 0.10 [95%-CI: 0.00–0.40] and 0.30 [0.02–0.79], respectively. Performance in models 4 and 5 was better with a mean R ² of 0.77 [0.02–1.00]. Subsequently, eNose data in model 4 had no added value over using CGM only in model 5. Continuous exhaled breath analysis using this eNose cannot be used to accurately predict blood glucose levels in intubated, mechanically ventilated ICU-patients.