The rapid development of the corona crisis requires new methods and approaches that could help flatten the curve. For this reason, a possible alternative for a detection method is investigated to diagnose in a faster and more reliable way the disease and help prevent the spread. Raman spectroscopy on blood serum is a potential candidate for this issue and thus, research was done towards this direction. The data obtained from the spectrometer was further analyzed through linear discriminant analysis and a predictive model was achieved with an accuracy of 93.5%.
In December 2019, a novel virus has emerged and in a brief period it has reached
all the corners of the world. This virus called the corona virus has afected all
people. Due to the continuous rise in the number of infected and deceased persons,
a new approach must be taken in order to deal with this situation. Since the
measures taken to isolate infected individuals had no significant result with the
current testing results, perhaps the development of a new detection method that
Copyright © 2021 for this paper by its authors. Use permitted under Creative Commons License
Attribution 4.0 International (CC BY 4.0).
is more precise and rapid could prevent the further spread of the virus.[
Raman spectroscopy is a method that relies on a non ionizing laser that can excite
a molecule if the energy of the incident photon matches the energy gap between the
ground state and the excited state of said molecule. The phenomena of fluorescence
will occur when the molecule relaxes and generates emission of photons in both the
visible and near-infrared spectral ranges. The emission can happen either by means
of a elastic scattering which has no relevant information, or plastic scattering which
means that a part of the energy was absorbed and only a fraction of that was
released back into the medium. That diference in terms of energy is studied in order
to obtain relevant information that may be used for bio-medical applications but
not only.[
In many bio-medical applications that include the use of Raman spectroscopy,
the probe being analyzed is blood serum. Many important features can be observed
that prove to be relevant in diagnosis. More precisely, in the blood serum, diferent
organic components are present such as proteins that can indicate the presence of
a virus, or one can analyze the serum in order to diagnose diefrent types of
carcinoma. However, in order to be able to analyse blood serum, a series of preparatory
steps must be taken to obtain a clear and reliable sample.[
Knowingly, the use of Raman spectroscopy on blood serum is proposed as a
detection method of infection with the SARS-COV-2 virus. Further knowledge
regarding the proteins that mark the presence of the virus in the human system was
acquired in order to obtain an estimated location in the spectroscopic data of those
elements for better detection by means of machine learning. The most important
protein that characterizes the virus is the spike glicoprotein also called S-protein.
Because of this particular protein the virus is able to enter our system by attaching
the so called spikes to a receptor called angiotensin converting enzyme 2 (ACE2).[
For this application, a linear classification technique called LDA was used. This
method is straightforward and it consists of analysis of statistical properties of
data that are calculated for each class. The statistical properties of interest are the
mean and covariance matrix over the multiple variables. The assumption made by
the algorithm is that all data is Gaussian and each attribute of the data has equal
variance and each value varies around the mean with the same amount overall.
After these are taken into consideration, the algorithm estimates the mean and
the variance for each class. Having these parameters, the next step can be made:
classification.[
The classification technique discussed above can yield accurate results on its own without the need of reassuring techniques such as leave-one-out cross-validation. However, being a delicate biomedical application, the maximum achievable accuracy is of interest. For this particular reason the sequential training and testing of diferent unique values is made. Precisely, each data point from the data-set will serve as test sample once. By doing this, all data points are thoroughly analyzed and assimilated with one of the classes with greater preciseness.
The data processing was done in MATLAB software by means of machine learning and statistical toolboxes. The data-set consisted of Raman spectra on blood serum of both healthy and infected subjects. The first step was to whiten the data by removing the DC component and then the data was visually inspected in order to observe any outstanding features that might impede the classification process. The whitened data can be observed in Figure 1.
Healthy vs SARS-COV2 6 5 4 ] .
U . [A3 y it s n te2 n I 1 0
The displayed data has subtle diferences which cannot be picked up by the naked eye in order to make a precise classification. For this reason, after the algorithm successfully detects those features that are relevant for the diferentiation between the two classes, the raman shift bands will be displayed for a better visual analysis.
After the normalization process, the data was randomized along with the labels
and in the healthy and infected individuals were mixed together in order to prepare
for the training process of said classifier. The division in terms of samples was done
in the following manner: 70% of the data was used as training set and 30% of the
data was used as test set. However, as it has been mentioned before, the training
was done through LDA and through LDA with leave-one-out cross-validation
technique in order to compare the results and to take into consideration the possibility
of over-fitting of the model.[
The data-set was composed of a total of 309 individuals of which 150 were healthy
and 159 were infected with the corona virus[
From this table it can be seen that from the total of 48 healthy individuals used for the testing process, 47 were correctly identified as being healthy (which accounts for 47 true negatives) and one was erroneously identified as infected (false positive). The infected individuals had a lower accuracy when it comes to detection as from the total of 45 infected, only 40 were correctly identified as indeed infected (true positive) and 5 were assumed to be healthy (false negative). Other parameters were taken into consideration for performance assessment of the classifier such as the accuracy, sensitivity and specificity of the model. All these parameters can be seen in Table 1.
As discussed in 2.1, certain proteins yield Raman shift ranges that are useful for the classification process. The predominant features in terms of Raman shift in this particular data-set were in the following ranges: [400–591]cm −1, [647–673]cm −1, [721–798]cm −1, [820–896]cm −1 and [1003–1241]cm −1. From all the ranges discussed in 2.1 the one that overlaps significantly with the predominant features in this dataset is the one in the 600–800 cm −1 range. The conclusion that can be drawn from this is that the spike glicoprotein that is responsible for the infection with the SARS-COV-2 virus has a high vibrational state in this particular range which can be used to detect infected individuals. Now, additional information has come to light and better approaches can be made in order to increase the accuracy of the model. The predominant features used to achieve the classification process are now reduced to one feature generally speaking. This single feature is represented by a range of Raman shifts in the 600–800 cm −1 that can be seen in Figure 4.
From a visual inspection it can be seen that the data from the two classes show little diferences in terms of intensities. However, there are certain location in the specified ranges of raman shifts in which the diferences in intensities go up to 0.5 [A.U.]. These values are high from the perspective of Raman spectroscopy accounting for a diference of 12.5% in terms of intensity. When analyzing these diferences the most important characteristic that was considered at all times was the quantity of virus present in the sample which could have been very low for some of the subjects. The quantity is highly dependent on one parameter: days after infection (Figure 3). Because of this reason, and because of the fact that there was no information regarding this parameter, possible errors in classification might have occurred. 600 650
700 750 Raman shift [cm-1]
800
The proposed technique for detection of SARS-COV-2 infection proved itself to be
useful and reliable. The potential of Raman spectroscopy in the detection of the
virus is considerable especially due to its ability of observing even subtle diferences
that can be missed by the naked eye and which are key elements for the classification
process. The LDA algorithm combined with leave-one-out cross-validation created
a statistical model that is well suited for the task at hand. Rapid tests are
timeeficient, however the accuracy of detection is only ∼80%. These tests can be done
by the patient itself and they require no medical expertise.[