Glucose prediction by analysis of exhaled metabolites – a systematic review

The literature search in Pubmed yielded 1041 titles (Figure 1). After reading titles and abstracts, 1012 articles were excluded because the topic was outside of the scope of this review and 29 articles were retained for full–text assessment. After the exclusion of 20 papers (5 reviews/non-original studies, 13 on unrelated topics, 2 index test not compliant with inclusion criteria), 9 articles were included in the analysis. Characteristics of selected articles are presented in Table 1. Five studies included healthy non-diabetic subjects, two studies included Type 1 Diabetes Mellitus (T1DM) subjects, one study included Type 2 Diabetes Mellitus (T2DM) subjects and one study included both healthy and T1DM subjects.

Results of the quality assessment using the QUADAS–2-tool are presented in Table 2. The risk of bias was considered high for all studies; none of the studies used a random sample of patients, all using a pre-specified target group such as T1DM or T2DM patients. The use of blood gas measurements or central lab measurements was considered to be the correct reference standard [16]. The adequate reference standard was used in four studies. Four studies used finger prick measurement, which increases the possibility of incorrect insulin titration in clinical practice [16]. Comparing these measurements to exhaled breath could lead to biased results. However, none of these studies were excluded from our review.

Authors of seven out of nine studies found a strong correlation between one or more metabolites in exhaled breath and blood glucose levels, with a mean linear regression coefficient of 0.82 [range: 0.08-0.98] (Table 1). However, none of the included studies validated their results internally (e.g. cross-validation) or externally (e.g. in an separate validation cohort). A total of ten metabolites have been reported to correlate with blood glucose levels, including exhaled acetone, VOCs produced by gut flora (ethanol, methanol, and propane), exogenous compounds (Ethyl benzene, o-xylene, and m/p-xylene) and VOCs that reflect oxidative status (methyl nitrate, 2-pentyl nitrate, and carbon monoxide (CO)) (Table 3). Authors of two studies did not observe a strong correlation. The first one of those did not find a significant correlation between a single measurement of breath acetone and blood glucose in T2DM subjects. Authors of the second study were unable to demonstrate a strong correlation between glucose levels and exhaled CO in healthy subjects. Researchers in one of the studies that did show a strong correlation between breath metabolites and glucose levels, only observed this after overnight fast, showing a weak correlation after consuming a meal [17].

Researchers in seven out of nine studies performed multiple measurements with an interval ranging from 2.5 to 40 minutes. Two studies had a cross-sectional design and only performed a single measurement, or two unpaired measurements. None of the authors of the included studies reported on the possibility of predicting glucose trend.

A significant correlation between metabolites in exhaled breath and blood glucose levels was found using GC-MS, SIFT-MS, PTR-TOF-MS, a nano-sensing film-based sensor, and an electrochemical analyzer as analytical method. GC-MS is considered to be the gold standard for VOC detection and has shown to have a high sensitivity to identify single VOCs [23]. Therefore, GC-MS is suitable to accurately quantify a number of different VOCs in a cross-sectional study. However, the time-consuming nature of the technique limits use of the device for real-time and continuous measurements, which hampers clinical application. Other analytical techniques such as SIFT-MS [22, 23] and PTR-TOF-MS [23‐25] can also identify single VOCs and can be used for real-time continuous measurements. Disadvantages include possible selection bias [26] and the limitation to the concentration range that can be detected [25].

The electrochemical analyzers used in selected studies are two different Smokerlyzer Micro (Bedfont, UK) devices. These devices measure the amount of CO in exhaled breath. However, there is a cross sensitivity to hydrogen [19]. While a correlation between exhaled CO and glucose levels was found by the researchers of one study [27], researchers of another study [19] could not reproduce these findings. Contrasting results may be due to the high cross sensitivity to hydrogen in the electrochemical analyzer used previously [27], which was less apparent using a newer device [19]. This exemplifies the importance of an adequate analytical technique that suits the aim of the study.

An important limitation of the techniques used in included studies is that none of them was used to continuously monitor exhaled breath. Continuous analysis of the exhaled breath was previously described by means of IMR-MS [28], PTR-MS [24], and electronic nose [29, 30]. After a training phase, electronic noses learn to recognize specific disease states and can therefore be used for classification. The devices cannot identify and quantify single VOCs, but they do give a rapid, bedside diagnosis, which, from a clinical perspective, renders this device attractive. The electronic nose has been used to discriminate between patients with and without diabetes [31]. One could postulate that the ability to diagnose diabetes is partly due to the metabolomic alterations because of higher blood glucose levels in diabetic patients. Therefore, electronic nose analysis may complement mass-spectrometry based techniques for the monitoring of blood glucose levels in clinical practice, providing signals based on probabilistic training and validation. Alternatively to semi-selective recognition, nanosensors also rely on specific recognition of certain VOCs [30]. In one study, an acetone-selective nanomaterial-based sensor was used alongside PTR-TOF-MS and showed a strong correlation between acetone, glucose and the sensor [17]. Small size of devices using nanomaterial-based sensors as compared to spectrometry-based methods facilitates clinical application.

Mechanisms related to the association between VOCs and glucose levels can be found in Table 3. Acetone appeared to be associated with blood glucose levels [17, 20‐22, 32]. As a result of increased synthesis of acetone and degradation of ketone bodies, acetone is expected to be higher in diabetics [33]. On the other hand, healthy humans only have elevated levels of ketone bodies when fasting or exercising [11]. Therefore, it is more likely to find a correlation between exhaled acetone and glucose levels after fasting compared to finding a correlation after lunch [17]. Acetone possibly is a good marker for glucose levels in ICU-patients. However, the large variation in breath acetone levels between subjects [18, 34‐36] may result in low accuracy when using acetone cross-sectionally.

VOCs such as ethanol [17, 20, 22, 32], propane [22] and methanol [17, 22] are likely to reflect gut flora activity, since the metabolism of gut bacteria is responsive to glycemic fluctuations [22, 32]. However, we cannot exclude that other biochemical pathways also contribute to the production of these compounds. In critically ill patients on the ICU, the quantity and composition of the gut microbiome are changing over time and therefore the amount of VOCs they produce may not be stable [37] Therefore, these markers are less likely to predict glucose levels in ICU-patients.

Ethyl benzene [20, 22], o-xylene [20] and m/p-xylene [20] are gasses that are inhaled, partially metabolized by the liver and subsequently exhaled at lower concentrations [20]. Rapid-onset hyperglycemia likely suppresses hepatic metabolism, thus causing peak concentrations of these compounds in exhaled air. Recent evidence suggests that cyclic hydrocarbons such as ethyl benzene and xylene are emitted by the ventilator and tubing [38]. Given that exhaled air is readily accessible for measurements in mechanically ventilated ICU patients, use of exhaled air for the prediction of glucose levels is therefore plausible.

An isomer of methyl nitrate [22, 39] is formed when methanol reacts with nitric oxide, which in turn reacts with superoxide ion (O₂ ^-), a by-product of oxidative reactions [39]. Furthermore, 2-pentyl nitrate [22] is generated through pathways involving organic peroxy radical (RO₂) and NO or NO₂. This could be modulated by acute changes in systematic oxidative status [22]. Changes in CO [27] in exhaled breath are possibly related to oxidative stress. When glucose levels rise, particularly in diabetic patients, this can lead to oxidative stress. As a protective response, heme oxygenase is activated, leading to the positive modulation of CO on insulin secretion [27]. For critically ill patients on the ICU however, markers of oxidative stress will be non-specific for high blood glucose, as they increase with any form of oxidative stress such as sepsis, high inspired-oxygen fraction and acute respiratory distress syndrome [40].

The observed correlation between blood glucose levels and exhaled VOCs may be due to the inclusion of T2DM patients and/or a cross-sectional study design. First, T2DM influences the responsiveness of the body to changes in blood glucose levels [41]. This is typically characterized by insulin resistance but may also influence the formation of ketone bodies and the induction of liver enzymes. Second, breath acetone levels tend to differ between T1DM, T2DM, and healthy subjects [18, 21, 35]. Therefore, a decrease in blood glucose levels may not induce the same rise in breath acetone levels with different baseline values and in the context of different co-morbidities. Finally, in line with the previous point, correction for baseline differences between subjects cannot be accomplished with a cross-sectional study design. This is further acknowledged by the fact that the predictive algorithm requires calibration for every subject in several studies [20, 22]. Since the relation between exhaled breath metabolites and blood glucose levels shows high inter-person variation, a cross-sectional design may not be ideal for predicting glucose levels using breath metabolites. The possibility of using a single breath maneuver to estimate blood glucose levels thus seems implausible. Future studies may therefore focus on longitudinal measurements in the same subject.

Five included studies used a clamp study design and 2 studies used an oral glucose tolerance test (OGTT). Clamp studies and OGTT result in a more or less predictable course of blood glucose levels. Although a clamp design is ideal for research purposes and enables comparability between studies, clinical practice is often very different and less predictable. The transition of the results of these studies to the clinical setting will be a major challenge for the field of blood glucose estimation by exhaled breath analysis.

We used a standardized systematic review approach, combining all evidence available. All VOCs that are linked to changes in glucose levels are discussed and their most likely biochemical pathways are described. In addition, we carefully assessed the quality of the included studies.

This systematic review also has an important limitation. The included studies were highly heterogeneous with respect to patient selection, exhaled breath sampling and analysis and blood glucose measurement, limiting the comparability of the studies. Therefore, we decided to describe the results separately. Most of the included studies had a relatively high risk of bias and we found that included studies did not validate their results. Possibly, this is inevitable in the early stages of biomedical research but it hinders strong conclusions. Furthermore, models can possibly be overfit, yielding overoptimistic results. Our search only identified one negative study. Negative studies are often not published leading to publication bias.

None of the studies investigated ICU-patients, while glucose fluctuations are large and frequent in this population [42]. Therefore, we cannot draw firm conclusions on the use of these methods in ICU-patients. We did try to identify potential pitfalls for the implementation of these methods in ICU patients by reviewing the biochemical pathways for the formation of VOCs.

Finally, the use of exhaled breath to monitor glucose trends was not discussed in any of the articles. Monitoring glucose trends (in ICU patients) however, has several potential advantages over using single values. First, trend has a better predictive value compared to single glucose levels; recent trend can be used to predict future levels. In ICU patients, this can lead to improved insulin titration. Second, because outliers can be filtered out, trend is less susceptible to random noise. Third, possible bias (constantly predicting values too high/low) will be constant throughout the trend, having a smaller effect. Potential disadvantages of using glucose trend are possible lag in the signal, and the potential of amplification of errors.