Chapter 2. Artificial Olfaction
2.3 Analyses and Publications
2.3.1 Scotch Whiskey
To gain intuition for the data collection process, a study was performed in which di↵erent scotch whiskeys were run through the FAIMS, generating a dataset of whiskey odours on which a basic analysis was performed. Five di↵erent brands of
Study Technique Sample type Section Publication Whiskey FAIMS Sample headspace 2.3.1
Irritable Bowel Disease FAIMS Breath 2.3.2 [Arasaradnam et al., 2016b] Tuberculosis uvFAIMS Breath 2.3.3 [Sahota et al., 2016] Hepatic Encephalopathy uvFAIMS Breath 2.3.4 [Arasaradnam et al., 2016a] Diabetes E-nose Urine headspace 2.3.5 [Esfahani et al., 2015] Bacterial Vaginosis E-nose Vaginal swab 2.3.6
Diabetes/Obesity GC-IMS Urine headspace 2.3.7
Table 2.2: The studies performed used a range of artificial olfaction instruments, and used di↵erent biological samples to measure. Studies in which sample separation was successful were published.
scotch whiskey were used (see Table 2.3). Four of the bottles are single malts; i.e., produced by a single distillery. A blended whiskey was also included. For each single malt, four 5cl bottles were purchased to investigate within-distillery variation. The FAIMS used was the Lonestar (Owlstone, Cambridge, UK) [Owlstone Nanotech, 2017a,b, 2006; Wilks et al., 2012; Shvartsburg et al., 2009].
The Ardbeg whiskey is “peated”. This means that peat smoke is used in production of the whiskey, and gives a strong recognisable smell in the final product. This was included to investigate if the FAIMS can distinguish peated from un-peated whiskeys. A range of maturation times were also selected, as maturation may also be a property the FAIMS can detect.
Data Collection
The FAIMS detector chip is highly sensitive to alcohol; the high alcohol content of the neat whiskey would saturate the response from the sensor and would require decontaminating the sensor for long periods of time (days or more). The whiskey must thus be diluted. 10%, 4% and 2% dilutions were attempted, and 4% was determined to give a good signal strength without saturating the sensor, so was used in the following experiment.
For each of the 4 single malts, samples were taken from 4 di↵erent bottles. Only a single bottle of the blended whiskey was used. For each sample, three sequential runs were done, recording three data points. The total number of smell signatures recorded was thus 4⇥ 4 ⇥ 3 + 3 = 51. The methodology used in data collection was as follows:
1. Prepare Sample: pipette 4800ml water and 200ml spirit into sample container. 2. Replace sample in FAIMS with newly prepared sample.
Distillery Single Malt Age ABV Peated
Aberlour Yes 10 40% No
Ardbeg Yes 10 46% Yes
Glenfiddich Yes 15 40% No
Glenlivet Yes 18 43% No
Co-operative Blended Scotch Whiskey No – 40% –
Table 2.3: Whiskey brands analysed with the FAIMS. A range of maturation ages was selected, as was the presence of a peated whiskey and a blend. The omitted information for the blended whiskey is not commercially available.
Figure 2.10: Whiskey data projected onto the first two Principal Components. Each sample was left in the FAIMS long enough to collect three measurements, and each of the three runs of each sample is connected by a dotted line. It can be seen that the first principal component is simply how long the sample has been in the FAIMS. Each type of whiskey has been given a di↵erent symbol, and no separation between whiskeys can be seen.
3. Warm sample for 5 minutes.
4. Start the FAIMS, collecting three data points over 7 minutes 30 seconds. 5. Stop data collection, wait for pressure and temperature to stabilise (⇠ 0.5minL ,
⇠ 0.1bar).
6. Remove sample, replace with empty sample container.
7. Perform blank runs until the sensor is no longer contaminated. 8. Repeat.
Analysis
Performing PCA on the whiskey data and projecting the data onto the first two principal components, we obtain Figure 2.10. Each sample from a single bottle gives three data points due to being left in the FAIMS for thee runs to investigate by how much the runs di↵er. This turns out to be the largest source of variance in the data, and is thus the first Principal Component of the data. The third run is used for the remainder of the analysis, so each individual bottle corresponds to exactly one data point. This was chosen as the third run has the greatest mean
Figure 2.11: AUCs obtained with Random Forest classifier after supervised dimension reduction. The 4 classification tasks are 1-against-rest for each of the 4 single malt whiskeys. It is clearly an easier task to distinguish the Ardbeg from the other whiskeys, which may be related to it being the only peated whiskey.
We attempt 4 binary classification tasks: for each sample we classify whether it is from a single malt of interest, or from the remaining 3 single malts. This is done with a leave-one-out cross-validation to produce a set of predictive probabilities, and uses the previously developed pipeline detailed in Section 2.2.2. AUCs for each task are displayed in Figure 2.11.
Observing Figure 2.11, it can be seen that Ardbeg is readily distinguished from the other whiskeys. This may be due to it being peated, which would be an interesting hypothesis to study further. The remaining three single malts do not separate so easily. Glenlivet looks like it separates well and the 95% confidence interval of the AUC does not include 0.5, but more data would be needed to draw the conclusion that the FAIMS can distinguish it easily.
Linear Discriminant Analysis
Linear Discriminant Analysis (LDA) is performed to linearly project the data onto the two dimensions in which each of the 5 classes of whiskey are maximally separated (Figure 2.12). Since the number of data points and classes are low compared to the dimensionality of the data, it is unsurprising that the classes can be separated well. To test generalisation of separation across types of whiskey, each whiskey is removed in turn when learning the projection. When applying the projection to the data, the held out data continues to form clusters distinct from the whiskeys used to learn the projection. This shows that, under this statistical model, the FAIMS sees samples from our unseen whiskeys as similar to each other and di↵erent to other whiskeys seen.
Figure 2.12: LDA is applied to the whiskey data, reducing the data to two dimensions. Plotting this (top left) shows good separation between groups. To see how the projection gen- eralises to new whiskeys, we learn the projection 5 more times (remaining plots), each with one of the whiskeys omitted during learning. Applying the learned projection to all whiskeys, we see the omitted whiskey is still separated from the others, showing good generalisation.
Conclusions
This small sample study provides evidence that the FAIMS is, with the correct signal processing pipeline, able to distinguish between certain types of whiskey. With more data perhaps the whiskeys could be distinguished with greater accuracy. The easiest whiskey to distinguish from all others in the study was the only peated whiskey investigated. Further investigation would be required to show that the FAIMS can distinguish peated from un-peated whiskeys, but this study provides support for such a hypothesis.
Care must be taken in the experimental process to produce high quality data. The FAIMS must be decontaminated for sufficient time between running samples, particularly if they contain a compound to which the sensor is particularly sensitive, such as alcohol. The data produced are also sensitive to how long the sample has been in the FAIMS chamber. If a machine learning classifier has been trained on just the third run of each training sample, then only the third run of a sample should
be classified.
The study was only of small sample size, and a number of analyses were attempted, such as including and excluding the cooperative blended whiskey. This kind of high-flexibility and low-power study can often appear to produce significant results when there are none [Simmons et al., 2011]. Whilst it may be sensible to believe the FAIMS is able to distinguish Ardbeg from the other un-peated whiskeys, the 100% accuracy achieved here is unlikely to hold on a larger study.
2.3.2 Non-invasive exhaled volatile organic biomarker analysis to