Conceived and designed the experiments: RH YR DH NS. Performed the experiments: RH AM NS. Analyzed the data: RH NS. Wrote the paper: RH YR DH NS.
The Weizmann Institute has filed a patent on “predicting odorant pleasantness with an electronic nose” as described in this manuscript.
A primary goal for artificial nose (eNose) technology is to report perceptual qualities of novel odors. Currently, however, eNoses primarily detect and discriminate between odorants they previously “learned”. We tuned an eNose to human odor pleasantness estimates. We then used the eNose to predict the pleasantness of novel odorants, and tested these predictions in naïve subjects who had not participated in the tuning procedure. We found that our apparatus generated odorant pleasantness ratings with above 80% similarity to average human ratings, and with above 90% accuracy at discriminating between categorically pleasant or unpleasant odorants. Similar results were obtained in two cultures, native Israeli and native Ethiopian, without retuning of the apparatus. These findings suggest that unlike in vision and audition, in olfaction there is a systematic predictable link between stimulus structure and stimulus pleasantness. This goes in contrast to the popular notion that odorant pleasantness is completely subjective, and may provide a new method for odor screening and environmental monitoring, as well as a critical building block for digital transmission of smell.
Electronic noses (eNoses) are devices aimed at mimicking animal noses. Typically, these devices contain a set of sensors that generate a pattern representing an odor. Application of eNoses entails first “training” the eNose to a particular odor, and once the eNose has “learned”, it can then be used to detect and identify this odor. Using this approach, eNoses have been tested in applications ranging from disease diagnosis to space-ship interior environmental monitoring. However, in contrast to animal noses, eNoses have not been used to generate information on novel odors they hadn't learned. Here, rather than train an eNose on particular odorants, we trained an eNose to the perceptual axis of
Dravnieks envisioned an artificial (or electronic) nose as “an instrument that would inspect samples of odorous air and report the intensity and quality of an odor without the intervention of a human nose”
The main component of an eNose is an array of non-specific chemical sensors. An odor analyte stimulates many of the sensors in the array and elicits a characteristic response pattern. The sensors inside eNoses can be made of a variety of technologies, but in all cases a certain physical property is measured and a set of signals is generated. The stages of the recognition process are similar to those of biological olfaction, where a sensor type responds to more than one odorant and one odorant type activates more than one sensor. Together, the set of activated sensors and their signals characterize the odor (sometimes refered as an odor fingerprint). Thus, an important difference between eNoses and analyte detectors such as gas chromatographs, is that whereas the latter are aimed at identifying the components that contribute to an odor, eNoses can be used to identify, as a whole, the mixture of components that together form an odor.
Despite the promise of an artificial system that may substitute for olfaction, few efforts have been made to use eNoses in tasks that go beyond detection and discrimination. A notable exception are the efforts to develop eNoses for medical diagnosis (reviewed in
eNose output has been linked to some aspects of perception such as odor intensity
We first measured 76 odorants (Supporting
Encouraged by our ability to use an eNose to predict the pleasantness of odorants within the training set (P<0.05 in 100% of the 20 runs), we set out to test its performance with novel odorants, i.e., odorants that were not available during the algorithm development. We used the eNose to measure 22 essential oil odorant mixtures made of unknown components (Supporting
Although these odorants were novel, some of the participants in this study had participated in the original model-building study as well. To address the possibility of any bias introduced by this, we repeated the study again with 17 new participants, and obtained a similar correlation of r = 0.59±0.03, P<0.0001), i.e., a machine-human correlation that was 82% of the human to human correlation.
To further test the robustness of our findings, we conducted a third test of our apparatus, using yet another set of 21 novel neat odorants (Supporting
Up to this point, we considered a continuous scale of odorant pleasantness. Naturally, the correlation between individual human subjects, as well as between human subjects and machine, was lower for ambiguous or intermediately rated odorants. Therefore, we now set out to ask how the eNose would perform if we restricted our analysis to the categorically pleasant and unpleasant odors.
We conducted a classification analysis after removing odorants with intermediate pleasantness scores (odorants with pleasantness rating ranging from 10 to 20 on the 30 point scale). We classified odorants as pleasant if their predicted pleasantness value was above zero, and unpleasant otherwise. Strikingly, the eNose discriminated between the two odor groups with 99% accuracy (
Four typical odorant eNose signals of both the QMB sensor module (upper panels) and MOX sensor modules (lower panels). Each line shows the dynamic response of one sensor. Note that both pleasant and unpleasant odorants generated both strong and weak responses.
A portion of human olfactory perception is modified through culture
Interestingly, despite co-author AM's fluent Amharic, we encountered difficulty in conveying the notion of a visual-analogue rating scale to the native Ethiopian participants. That is, the native Ethiopian participants tended to rate odors at the extremes of the scale, and made lesser use of the middle range. This was made evident in the standard deviation of the VAS scale values. Whereas the average standard deviation of the mean across the same odorants in the native Israeli participants was 6.1±1.5, the average standard deviation of the mean in the native Ethiopian participants was 8±1.5 (T(21) = 5.4, p<0.00002).
The correlation in pleasantness ratings between native Ethiopians and native Israelis was r = 0.75 (p = 0.00004). Although across all odors the median pleasantness assigned by native Ethiopians (14.9±6.5) was not significantly different from the native Israelis (16.7±6.6) (t(21) = 1.8, p = 0.08), when looking at each odorant separately, this group was significantly different from the native Israelis in its pleasantness rating of 7 odorants, 2 of which were rated as significantly more pleasant by native Ethiopians, and 5 of which were rated as significantly less pleasant (
A. Odorant-specific pleasantness ratings for native Ethiopians (blue), native Israelis (brown), and eNose (pink). The blue stars on the upper x axis denote the 7 odorants where native Ethiopians and native Israelis significantly differed in their pleasantness ratings. Note that for odors #6 #18 and #19 the pink line (eNose) is in fact closer to the native Ethiopians than to the native Israelis even though the eNose was tuned on a separate group of native Israelis. B. The correlation between the eNose pleasantness prediction values of 22 odorant mixtures (essential oils) and the values obtained from native Ethiopians. Each dot represents an eNose measurement (many dots overlay). Comparing
The average correlation between the machine prediction ratings and the native Ethiopian's median ratings was r = 0.52±0.01 (P<0.001) (
Finally, because of the standard deviation in VAS scale usage by the native Ethiopian participants, a classification analysis of extremely pleasant versus extremely unpleasant odorants similar to that conducted in the native Israelis is less informative in this case. Put simply, these participants rated nearly all odorants as extremely pleasant or extremely unpleasant, rendering a classification analysis similar to a simple correlation analysis. Nevertheless, we conducted a classification analysis as well, and the eNose discriminated between the two odor groups with 69% accuracy (p<0.0001).
Because the native Ethiopians and native Israelis significantly differed in their pleasantness ratings for only 7 odorants, this is too small a subgroup for independent statistical analysis. However, a descriptive observation of this subset of odorants remains informative in that for several of the odorants with significant differences, the eNose prediction was in fact closer to the estimates of the native Ethiopians than to the estimates of the native Israelis (e.g., odorants #6,18 and #19 in
To test the dependence of our algorithm on the size of the training set, we repeated the leave-group-out test while augmenting the training set with the essential oils data (
To farther test the dependence of our algorithm on the identity of odorants in the training set, we repeated the tests for each of the two novel odorant experiments while augmenting the training data with the other odorant set. The results remained similar: r = 0.56 (P<0.0001) and 100% classification rate in the essential oils experiment and r = 0.49 (P<0.0001) and 88% classification rate in the neat odorants experiment (when removing odorants ranging from 10 to 20 pleasantness ratings). In other words, the prediction was not a result of using a specific training set under specific training parameters.
To further probe the statistical robustness of the results, we scrambled our pleasantness data in a pseudorandom fashion 100 times and repeated our prediction analysis. The average prediction rates dropped to r = 0.08, P>0.23. In other words, the predictions obtained were not due to some internal structure of the data but rather reflected the ability of the algorithm to predict odor pleasantness.
Finally, to ask whether our results were significantly impacted by our outlier removal criteria for eNose measurements, we repeated the correlational analysis using all the data with no exclusions. This resulted in a minimal reduction in correlation between eNose and human pleasantness rating from r = .64 to r = .62, and this correlation remained highly significant (p<0.0004). We also repeated the classification analysis with inclusion of outliers, and classification accuracy remained the same (99%). We conclude that our results were not significantly influenced by outlier removal.
A face can be photographed, digitized and transmitted. Whereas software at the receiving end may be able to rate its beauty in the eyes of previously characterized observers
Similarly, a musical peace can be recorded, digitized and transmitted. Whereas software at the receiving end may be able to rate the appeal of previously characterized music for novel listeners
Here, we eNosed, digitized, and transmitted to receiving software, the smell-print of novel odorants, and in contrast to vision and audition, could predict their pleasantness with accuracy similar to that of a novel smeller. In other words, we could predict whether a person who we never tested before would like the odorant, and this prediction was consistent across Israeli and Ethiopian cultural backgrounds.
We argue that this difference was not a reflection of better hardware (in fact, an eNose is less precise than a modern camera or sound recorder), or better algorithms, but rather a reflection of a fundamental biological property of the sense of smell. These findings imply that unlike in vision and audition, in olfaction pleasantness is written into the molecular properties of the stimulus
It is tempting to speculate as to the specific molecular aspects that our apparatus was most sensitive to in its determination of pleasantness. For example, careful review of Supporting
This finding of hard-wired odorant pleasantness is in contrast to the popular notion that odorant pleasantness is both subjective and learned. We argue that in this respect olfactory pleasantness can be likened to visual color. Most would agree that color is hard-wired to wavelength within a predictable framework. That said, color perception can be influenced by culture
Although our results supported our hypothesis, we would like to clearly state their limitations. First, this manuscript used a rather basic commercially available eNose, and more modern eNose technologies may have performed even better. Thus, here we provided proof-of-concept that an eNose can be tuned to a perceptual axis. Beyond proof-of-concept, we do not claim that this iteration represents the best possible implementation of this concept. Second, odorant pleasantness is related to odorant concentration
Finally, these results may be considered a building-block for digital communication of smell
All subjects participated after signing informed consent to methods approved by the Institutional Review (Helsinki) Committee.
The MOSES II eNose we used contains eight metal-oxide (MOX) sensors and eight quartz microbalance (QMB) sensors. MOX and QMB are two very different sensor technologies that together capture many facets of the ligand's nature. The 1ml (without any dilution) samples were put in 20-ml vials in an HP7694 headspace sampler, which heated them to 50°C and injected the headspace content into the MOSESII with a flow rate of 40ml/liter. These parameters maximized the number of chemicals that elicited a strong response. To avoid the problem of conditioning we put a blank vial before every measurement and cleaned the system using steamed air after each run of 22 odors. Each analyte was first introduced into the QMB chamber, whence it flowed through to the 300°C heated MOX chamber. The injection lasted 30 seconds, and was followed by a 20 minute purging stage using clean air. Each chemical was measured five or six times over a period of several days. In total, we performed 424 measurements. Each odorant was measured at the same level of humidity and temperature. Each single measurement consisted of sixteen time-dependent signals, corresponding to the eNose sixteen sensors. All the raw eNose data is available for download as Supporting
From each of the 16 sensor signals we extracted four parameters. These parameters were: the signal max value and latency to max, the time the signal reaches the half max value on the decay part and on the rise part. In many cases the signal max value can change considerably between measurements of the same odorants, however, the relative height of the 8 sensors in each of the two sensor modules was largely maintained. Thus, to capture this behaviour we added to each odorant representation the 28 possible ratios of the 8 MOX signals and 28 ratios of the 8 QMB signals. We thus ended up with 120 features for each odorant. To ask whether this feature extractions method was a good representation of the odorants, we clustered the 424 eNose measurements we had into the 76 odorant classes and tested how many odorants fail to cluster into their odor class. Out of the 424 measurements 85% clustered correctly. We removed the 10% signals that failed to cluster to their class, although this did not change the result signifantly (see text). After this signal removal, we ended up with 3 to 6 repetitions per sample measured. We normalized both the feature values and the odorant signature thus removing bias to specific sensor type and odor concentration respectively.
Fifty six healthy normosmic native Israeli-born subjects (31 females) ranging in age from 23 to 54 years, and 31 healthy normosmic native Ethiopian-born subjects (24 females) ranging in age from 20 to 37 years, participated in the study. The Ethiopian subjects arrived in Israel between 1 and 5.5 years before testing (mean 2.3). All subjects were paid for participation.
The total of 123 odors (the 76 training odors and the 43 test odors) were divided into groups of 20–25 odors each. This grouping reflected the maximal time a human subject will typically consistently rate odors (∼40 minutes, with at least 30 seconds between odorant presentations). All odors were first individually diluted to be perceptually iso-intense. Each group of odors was then rated by 14 to 21 subjects. Each subject ranked the pleasantness and intensity of each odor on a visual analogue scale. The visual scale did not contain any markings or indicators other than the terms “very unpleasant” and “very pleasant” at each end. For purposes of analysis, the VAS was later scored from 0 to 30 as a function of the physical location where the VAS line was crossed (0 = very unpleasant, 30 = very pleasant). Each odor was randomly presented twice to each subject. In total, for each odor we had more than 30 ratings (few subjects did not want to rate for the second time). The pleasantness of an odor was calculated by taking the median of all subject's ratings.
To estimate human to human ratings we calculated the Pearson correlation between all subject pairs and calculated the average correlation value (n>100). To verify that our results were not biased due to the use of visual analogue scale (VAS) we ran an additional experiment using 21 odorants with 6 subjects using a 7 category rating experiments (categories were: The worse odor you ever smelled, very bad, bad, Ok, good, very good, the best odor you have ever smelled). The between human correlation was similar (r = 0.57 in the category rating experiment versus r = 0.6 in the VAS rating experiment; P<0.01 in both). Overall, when considering all our humans ratings, the human to human correlation was 0.45±0.18 (P<0.01) and human to the human group average correlation was 0.67±0.12 (P<0.01). Calculating the average correlation of each subject first rating to his second rating we obtained r = 0.73±0.15 (P<0.01).
We used MATLAB's implementation of a three layered feed-forward back-propagation neural network with 5 internal neurons and 20 epochs. Changing the number of neurons or epochs in the range of 3–10 and 10–30, respectively, did not change the result. The layers' transfer functions were ‘tansig’ and ‘purelin’. The training function was ‘traingd’. To calculate the prediction we ran the algorithm 20 times and used the average value as our best predictor.
To classify odors we used the same algorithm we used for the prediction. Odors with positive predicted value were classified as pleasant and odors with negative predicted value were classified as unpleasant.
List of odorants used
(0.04 MB XLS)
Raw eNose data. The zipped directory contains all the raw eNose data in text files, and a read-me file explaining its structure.
(14.88 MB ZIP)
We thank Eli Goldshmit and all the personnel and new immigrants at the JAFI Ashkelon Absorption Center for their help and cooperation. We also thank A. Elite.