Measuring Freshness of Salmon with Handheld Spectroscopic Instruments

While spectroscopic analyses within food science and the adoption of statistical and machine learning (ML) strategies have increased, there is still, in the minds of many, need for more robust models. Data from different modalities are not routinely explored together. In a joint study between the School of Biological Sciences of Queen's University (Belfast, Ireland), the University of North Dakota (Grand Forks, North Dakota), and SafetySpect Incorporated (Grand Forks, North Dakota) suggest that combining data sets may fill the need for a more accurate determination of produce freshness. Their recent paper extends previous experiments conducted by the alliance to include more visible and near-infrared (vis-NIR) absorbance optical data to explore spectral feature selection as a tool to increase classification accuracy. Spectroscopy spoke to Mike Hardy of Queen’s University, corresponding author of that paper, about their work.

Your paper (1) discusses the testing of freshness in salmon by hand-held devices. Why did you concentrate on salmon as opposed to other kinds of fish?

We have looked at other fish species. In our previous work, published in 2023, we also looked at sablefish freshness (2), and we also had a few tests on cod fillet at the Institute for Global Food Security (IGFS), School of Biological Sciences, Queen’s University Belfast (unpublished). Salmon fillet was useful as a test species that allowed us to focus on our methods and the data analysis, which is the real selling point of this latest study with Safety Spect Inc., USA (1). Last year, we also published a study in hyperspectral imaging (HSI) using a set-up from Hinalea Imaging (USA) (3), and again the focus here really was data – we have all this information, what smart things can we do with it to be as useful as possible? These data analysis strategies in spectroscopy (data fusion, feature selection, similarity matching in hyperspectral imaging) are broadly applicable to many different fish types and other foodstuffs.

When considering possible reasons behind this study, health and economic factors immediately come to mind. Are there any others that I’m missing?

No, that’s right. It’s quite difficult to tell the exact freshness state of seafood. A lot of what we did at IGFS was not just food safety or quality but identification of ‘food fraud’. This is where foodstuffs are intentionally meddled with by criminals for economic gain. Fish is a great example for a foodstuff where quality and fraudulent issues can easily arise. If one thinks back to our grandparents or great-grandparents, if they bought fish, it was probably from a local market and one probably knew exactly where it came from and possibly also who caught it, but nowadays it’s a much more complicated supply chain. As a poignant indicator of this, despite the UK being on islands, much of our seafood is imported, as high as 80 or 90%.

Spectroscopy can also be used to determine fish species. One might be surprised at difficult it is to tell that you are eating what you think you are in terms of fish (and not some cheaper species). And this is almost impossible if one is dealing with a processed fish product. There are also other issues too, like detection of freeze-thaw processes (has the fillet been frozen?), and ethical catch (method of slaughter can induce compositional changes to the fillet). Spectroscopy can potentially also offer detection and analysis solutions here.

You utilized fluorescence and absorption spectroscopy across the visible and near-infrared (vis-NIR) range (400–1900 nm) in your research. What was the reasoning behind choosing these spectroscopic techniques and wavelength regions as opposed to other techniques or spectral ranges?

We wanted to cover a wide range of wavelengths, but we may be more selective in the future. 340 nm is a well-known wavelength for fluorescence in biological samples, and it can excite a range of compounds at this wavelength.

Ultraviolet (UV)-visible and NIR absorption are common spectroscopy techniques; fluorescence, even more so. But the combining and/or manipulation of different datasets from the spectroscopic modalities is where the interesting part of our study lies. Looking at Raman spectroscopy also would have been beneficial, but we could not do it on this occasion. We compared the spectroscopic measurements with 1) a nucleotide assay, which looks at relative catabolite percentages as an indicator of time post-mortem that is freshness state, and 2) potentiometry, which is passing a small current through the fillet to judge cell membrane integrity (as an indicator of fillet freshness state).

What were the advantages associated with using handheld spectroscopic equipment in your study over laboratory equipment?

Hand-held devices offer portability meaning that they can be easily transported to where they need to be rather than being big and bulky, and confined to a lab. This is important because the food supply chain is now much more complicated. Foodstuffs can come from the other side of the world and this can raise concerns over food safety/quality. It’s potentially very useful to be able to do spot checks at a factory, say. From a food security point of view, there are now lots of new nodes in the chain where ill-meaning actors can interfere, and again having portable devices, that are easy to operate, could be incredibly useful to check food authenticity. Not doing so can be critical, most people in the UK will remember the ‘Horsemeat Scandal’ from about 10 years ago (4), where traces of horse meat showed up in beefburgers. Since then, there have been many more food crime incidents.

All this does not necessarily mean that lab-based techniques, like mass spectrometry, will be redundant – it’s likely that confirmatory analyses by more precise analytical approaches will still be needed. But spectroscopic solutions will find a place in the field to monitor and red flag any potential problems.

It’s worth pointing out that small devices are poised to make inroads elsewhere too. Previously, I worked with Prof. Pola Goldberg Oppenheimer at the University of Birmingham. Her team there are doing some fantastic work within point-of-care medical device development, particularly with regards to traumatic brain injury for roadside/pitchside analysis. I was interested to read her 2024 Spectroscopy Online interview (5). We recently published a review together on characterization methods for different kinds of surface enhanced Raman spectroscopy (SERS) substrates (6), and these can work well with portable spectroscopic instruments for the low-concentration detection regime outside of the lab.

Briefly summarize your key findings that you discuss in your article, and the conclusions you came to after reviewing these findings.

Spectroscopy offers a much faster, portable and cheaper solution than complicated lab-based procedures like the nucleotide extraction assay or mass spectrometry. And potentially a more accurate assessment than potentiometric analysis. Specifically, combining spectroscopic datasets can allow more accurate determination of fillet freshness state. Data fusion is a fascinating area and it’ll be useful to try to do it more effectively, perhaps with very different types of datasets, in future work.

Similarly, careful feature selection in spectra can also be useful to this end. We noted that while fluorescence in isolation provided a modest classification accuracy, when we isolated a mere four variables (wavelengths) in the spectra–corresponding to a small redshift as the fillet aged–we were able to classify fresh and spoilt fillets with statistical significance (confidence interval (CI) of 95%). Interestingly, extending the analysis region to 13 variables, covering more of the fluorescence peak, meant we could not differentiate between the different fillet freshness states–there was too much redundant spectral information included.

Do you anticipate similar results in using your technique to measure freshness in different fish, or other types of foods?

Yes, this should be possible for different kinds of fish, although I imagine the accuracy of determination for different species will differ. Our colleagues in the US have recently focused more on species determination (7). The usefulness of combining datasets from different modalities that is data fusion is arguably still underappreciated. Feature selection is better appreciated but perhaps could be optimized better. We have looked at a wide range of foodstuffs at IGFS, which is reflective of how wide-ranging food quality and security issues can be.

For example, I was also part of a short study on oregano adulteration. This is interesting because it’s one of the UK’s most adulterated foods. My colleague Dr Yicong Li at IGFS is hoping to publish our study in 2025. Like the fish work, there’s an awful lot of data, and thought is needed about what one should be doing with it. There was also a piece of work on lead chromate detection via x-ray fluorescence in turmeric, which can be used to make the turmeric look more yellow (and hence more profitable), but of course with health concerns. This is also a really pressing issue in food safety so watch this space (or spice!).

What difficulties did you encounter in your work, specifically sampling and analytical challenges?

Well, first off on the sampling, we divided the fillet into different regions (UK study), but these were still large areas on the fillet – it might be nice to be a bit finer with our spatial demarcation. Secondly, we would have liked to study more fillets, but it would have taken a much larger research team and more time and was not possible this time around. Third, having the same experimental conditions at the US and UK-based labs would have been great but this was tricky. As one of our manuscript reviewers pointed out though, worrying about these kinds of details in the first place too much can ‘kill good ideas’. We hope others can look at our methods and find them useful in their own studies.

In terms of analytical considerations, we performed extensive testing on the spectrometer devices prior to measurement looking for artefacts, linearity of response with varying integration times and so forth. This is especially important to do with prototype devices. We also performed preliminary measurements with the potentiometric device and had done some test runs with the nucleotide assay. In the latter case, it’s a rather protracted process and easy to make mistakes so the trials were useful–especially for me because I’m not a biologist by trade!

How did you process the spectroscopic data to obtain the results you were looking for? Are you confident in these results and why?

We used R and Python (SciKit Learn) for data clean-up and machine learning (ML). We used the R base functions and the R Kohonen package for self-organizing map analysis. This is an interesting ML technique that we were working with at the University of Birmingham for sugars discrimination (8). Yes, we have confidence in the results because the data strategies should be repeatable. A lot of problems with reproducibility can be avoided just with simple notes on how the data was processed for example scaling, smoothing and so forth. The real question for us is why some algorithms and processing techniques, including order of application, work better than others for different kinds of datasets. This kind of insight is important because it will allow us to design robust predictive models going forward.

Were there any factors that might affect the accuracy of your findings?

The devices used in the study were prototypes, although they worked well. We monitored ambient temperature, time the fillets spent in and out of refrigeration etc. but we could impose even tighter controls – this is likely in a further study. As I say, there are also some challenges with inter-laboratory studies (we are in the UK, our colleagues in the US). However, I view it as a positive that we can achieve similar results in different labs despite being on the other side of the world.

How do you imagine the results of your study can/will be applied?

The global portable spectrometer market is booming with a projected 9% Compound Annual Growth Rate from 2021-2030, passing $4B (9). It will be interesting to see what modes of spectroscopy will come to the fore for different applications. What kind of analytical sensitivity and specificity is really needed? I like to read Richard Crocombe on these kinds of things, and I know Dr Crocombe has discussed the possibility of spectroscopy in consumer goods at home (10), and that’s a fascinating idea. I remember hand-held Raman devices appearing about 15 years ago and I think there’s still a lot of excitement around what portable devices, wherever they appear, will do.

Are there any next steps in this research?

It’s a good question how small/integrated one needs to go with spectroscopic devices, and the inevitable performance trade-off. What’s a suitable cost range, how integrated do they need to be? Can there be one-use disposable chips? Lots of questions here. My background is in surface-enhanced Raman spectroscopy (SERS) and there may be a useful comparison to be made. There was a lot of excitement over the single-molecule SERS studies and gigantic enhancement factors, but really, anything near this level of sensitivity isn’t necessary most of the time. Instead, we need cheaper substrates and better reproducibility.

I am back at Queen’s University Belfast’s School of Maths and Physics these days, and our current project, Smart Nano NI (11), is looking into these kinds of questions. We are a UKRI-funded consortium led by Seagate. Our partner companies, Causeway Sensors, Cirdan, Yelo, are all photonics-based businesses and based in Northern Ireland who can provide expert input on device design. The project is funded under the Strength in Places scheme, and we view this location-based aspect of the project as crucial to achieving the overall aims which involves making sensing devices in our Smart Nano lab, so it is quite exciting. But, as I say, we need to think carefully about device scale and performance levels required.

We’re always happy to hear from anyone interested, so please feel free to reach out!

References

1. Hardy, M.; Kashani Zadeh, H.; Tzouchas, A.; Vasefi, F.; MacKinnon, N.; Bearman, G.; Sokolov, Y.; Haughey, S.A.; Elliott, C.T. Freshness in Salmon by Hand-Held Devices: Methods in Feature Selection and Data Fusion for Spectroscopy. ACS Food Sci. Technol. 2024, 4 (12), 1813. DOI: 10.1021/acsfoodscitech.4c00331

2. Kashani Zadeh, H.; Hardy, M.; Sueker, M.; Li, Y.; Tzouchas, A.; MacKinnon, N.; Bearman, G.; Haughey, S. A.; Akhbardeh, A.; Baek, I.; Hwang, C.; Qin, J.; Tabb, A. M.; Hellberg, R. S.; Ismail, S.; Reza, H.; Vasefi, F.; Kim, M.; Tavakolian, K.; Elliott, C. T. Rapid Assessment of Fish Freshness for Multiple Supply-Chain Nodes Using Multi-Mode Spectroscopy and Fusion-Based Artificial Intelligence. Sensors 2023, 23 (11), 5149. DOI: 10.3390/s23115149

3. Hardy, M.; Moser, B.; Haughey, S. A.; Elliott, C. T. Does the Fish Rot from the Head? Hyperspectral Imaging and Machine Learning for the Evaluation of Fish Freshness. Chemom. Intell. Lab. Syst. 2024, 245, 105059. DOI: 10.1016/j.chemolab.2023.105059

4. Elliott, C. Elliott Review into the Integrity and Assurance of Food Supply Networks─Final Report [Gov. Uk], 2014.

5. Acevedo, A. Raman Spectroscopy to Detect Traumatic Brain Injuries: An Interview with Pola Goldberg Oppenheimer. Spectroscopy website 2024. https://www.spectroscopyonline.com/view/raman-spectroscopy-to-detect-traumatic-brain-injuries-an-interview-with-pola-goldberg-oppenheimer

6. Hardy, M.; Goldberg Oppenheimer, P. When Is a Hotspot a Good Nanospot’ − Review of Analytical and Hotspot-Dominated Surface Enhanced Raman Spectroscopy Nanoplatforms. Nanoscale 2024, 16, 3293−3323. DOI: 10.1039/D3NR05332F

7. Sueker, M.; Daghighi, A.; Akhbardeh, A.; MacKinnon, N.; Bearman, G.; Baek, I.; Hwang, C.; Qin, J.; Tabb, A. M.; Roungchun, J. B.; Hellberg, R. S.; Vasefi, F.; Kim, M.; Tavakolian, K.; Kashani Zadeh, H. A Novel Machine-Learning Framework Based on a Hierarchy of Dispute Models for the Identification of Fish Species Using Multi-Mode Spectroscopy. Sensors 2023, 23 (22), 9062. DOI: 10.3390/s23229062

8. De Carvalho Gomes, P.; Hardy, M.; Tagger, Y.; Rickard, J. J. S.; Mendes, P.; Oppenheimer, P. G. Optimization of Nanosubstrates toward Molecularly Surface-Functionalized Raman Spectroscopy. J. Phys. Chem. C 2022, 126, 13774−13784. DOI: 10.1021/acs.jpcc.2c03524

9. Portable Spectrometer Market Size, Share, Competitive Landscape and Trend Analysis Report by Product Type, Distribution Channel and Application: Global Opportunity Analysis and Industry Forecast, 2021-2030. Allied Market Research website. https://www.alliedmarketresearch.com/press-release/portable-spectrometers-market.html

10. Crocombe, R. Spectrometers in Wonderland: Shrinking, Shrinking, Shrinking. Spectroscopy 2022, 37 (s11), 6-11. DOI: 10.56530/spectroscopy.lz8466z5

11. Smart Nano NI website. www.smartnanoni.com

Mike Hardy is an interdisciplinary scientist who works in spectroscopy across Physics, Biology, Chemistry and Engineering to solve a variety of problems within healthcare, food security, and defense/homeland security. This is chiefly Raman and surface enhanced Raman spectroscopy, but he has also worked in UV-visible, infrared, fluorescence and x-ray fluorescence (XRF). Mike's background is in optics / photonics, plasmonics, and condensed matter physics. He has recently been working on machine learning techniques and the development of portable devices to complement his spectroscopy work. He received both his BSc and PhD from the School of Maths and Physics at Queen's University Belfast, under the supervision of Dr Paul Dawson, and he is currently working on the Smart Nano NI project in the school. He has also held research positions at Queen's School of Biological Sciences, and the School of Chemical Engineering at the University of Birmingham. Photo courtesy of Hardy.