Use of the Feature Scaling and Machine Learning Techniques on Optical Fiber Biosensors for the Detection of Neuroprotector IL-10 in Serum of a Murine Model with Cerebral Ischemia

Abstract

Typically, response analysis of optical fiber biosensors focuses on changes in amplitude and wavelength shifts in the biosensor spectrum; therefore, not all of the spectral range is used for this analysis. On the other hand, if the entire spectrum is used, it is possible to leverage the current data in the spectrum and thus improve the performance of the biosensor. To do this, it is necessary to analyze a large amount of data present in each measured spectrum. This task can be made easier by using dimensionality reduction techniques. In addition, it is necessary to establish which spectral regions provide relevant information. Scaling techniques are mathematical data preprocessing tools used in machine learning to adjust the numerical scale of variables so that they have comparable weight and even highlight those characteristics that provide more information. To our knowledge, the use of these techniques in the development of optical fiber biosensors is not very common, which is why we believe they represent an attractive topic of study in this area. With the help of scaling techniques, we can modify the scale of the data so that all the information contained in the spectrum is used, regardless of its magnitude. In this work, two biosensors based on a chirped long period fiber grating (CLPFG) and a chirped Mach–Zehnder interferometer (CMZI) were developed for the detection of interleukin-10 (IL-10). Principal component analysis (PCA) was used as a dimensionality reduction technique together with a support vector machine (SVM) classifier with four different scaling techniques, standardization, minimum–maximum scaling, robust scaling, and a custom transformer, to compare the IL-10 detection performance of the biosensors. The results showed that robust scaling in CMZI performed best in detecting IL-10, with an F1-score equal to 1, as well as better reliability in detecting the protein.

1. Introduction

The protein interleukin-10 (IL-10), an immunosuppressive cytokine, is considered an important modulator anti-inflammatory that prevents neuronal degeneration in pathological conditions. The treatment of neurodegeneration associated with neuroinflammation in diseases such as multiple sclerosis, cerebral ischemia, Alzheimer’s, and Parkinson is focused on the role of this cytokine as a protective action [1,2,3,4]. The role of IL-10 could help the development of useful targeted therapies to limit the neurodegenerative processes. The most common techniques used to measure IL-10 include enzyme-linked immunosorbent assays (ELISA), which quantify the protein in samples such as serum [5], and flow cytometry to evaluate its expression in specific cellular populations in blood [6]. IL-10 expression may also be measured by techniques like polymerase chain reaction (PCR). Nonetheless, many of these methods are expensive and require a long time to determine IL-10 expression.

As was mentioned above, IL-10 performs a crucial role in the regulation of immune responses by inhibiting the production of proinflammatory cytokines. Therefore, if a murine model is induced with ischemia and treated with estradiol benzoate (EB), it is expected that there might be an expression of IL-10. In this way, the serum samples analysis of a healthy murine model and its comparison with serum samples of murine models with ischemia would allow us to train a classifier for IL-10 detection because the serum samples of a murine model with ischemia and ischemia treated with EB must be a contrary situation to the ischemia process without treatment.

On the other hand, during the last decade, biosensors based on long period fiber grating (LPFG) have attracted a lot of attention due to their potential use in environmental applications and clinical diagnostic research [7,8]. Particularly, the chirped long period fiber gratings (CLPFG) can be used to develop these kinds of biosensors. CLPFG is characterized by a non-uniform period determined by the so-called chirp parameter. It has been demonstrated that the chirp parameter can be easily modified by choosing different periods and steps, and through this method, a special spectral profile can be obtained [9,10]. As far as we know, few studies have been reported on the use of this kind of grating in the development of biosensors.

CLPFG biosensors are based on the selective interaction of the analyte of interest with a functionalized coating. Such interaction generates small changes in the refractive index of the medium surrounding the grating, resulting in changes in the transmission spectrum. These spectra are inherently composed by multiple variables; therefore, the changes in them are not always easy to associate with the presence and magnitude of the analyte of interest. Furthermore, there may be a complex relationship between changes in the spectrum (in magnitude or shift) and the spectral region where the biosensor response is analyzed, which ultimately makes it difficult to determine the spectral regions with the best response. This situation is a problem that has been widely observed in the development of machine learning algorithms that usually deal with multidimensional signals. Particularly in multivariable signal analysis, signal scaling is an important aspect, since it allows each feature present in the signal (transmission at a particular wavelength in this case) to contribute its information to the chosen model, despite the different scales of the changes present in the entire spectrum. Thus, the main objective of preprocessing performed by scaling is to modify the magnitude of the transmittance at each wavelength so that, if it contains relevant information, it can be used in the chosen model for analyzing the biosensor responses. The right scaling can radically change the performance of the entire system.

Once the appropriate scaling has been performed, a dimensional reduction technique can be used to reduce the number of characteristics of the biosensor response and process these signals more efficiently. Usually, principal component analysis (PCA) is employed because, besides reducing the number of dimensions, it has an additional effect of avoiding the potential multicollinearity present in the original spectrum. This is achieved because the principal components are orthogonal to each other, which means that the information contained in each of them cannot be expressed as a linear combination of some other principal components. Another interesting feature is that, when plotting the different components of each spectrum, it is possible to observe patterns or trends that help to visualize the overall behavior of the sensor.

If the biosensor PCA plot presents clear patterns, it is then possible to perform a classification task, i.e., apply a classification algorithm to establish, for example, the presence/absence of the analyte.

In the present work we report the development of biosensors using a CLPFG and a chirped Mach–Zehnder interferometer (CMZI), constructed by two cascaded CLPFGs. Both biosensors were coated with a polyclonal antibody IL-10 as a biological recognition element. These allow the detection of interleukin IL-10 in serum samples with cerebral ischemia from a murine model with high specificity. Furthermore, the use of different machine learning techniques has enabled the analysis of biosensors responses for detection of IL-10, improving their reliability. We compared the performance of the biosensor response for IL-10 detection using four types of scaling: standardization, min–max, robust, and customized. PCA was used as a dimensional reduction technique, and the support vector machine (SVM) was used to determine IL-10 presence since this algorithm has demonstrated robustness and easy implementation. The SMV classifier distinguished between cases where IL-10 concentration is higher due to the applied treatment (EB) and, in contrast to an ischemic event without it, in which Il-10 expression must be low or null. Finally, we can say that the best performance was for robust scaling for the case of the CMZI biosensor.

2. Materials and Methods

2.1. Chirped Long Period Fiber Grating (CLPFG)

A CLPFG is a non-uniform modulation of the refractive index in an optical fiber core in which optical modes are coupled from the core to the cladding, resulting in various attenuation bands in the fiber transmission centered at different wavelengths. The evanescent wave on cladding modes can interact with the external environment, and, as a result, it is possible to detect refractive index changes which can be related to a coating material on the surface and its interaction with the analyte to be detected. Therefore, small variations in the analyte concentrations will produce changes in the spectral response by modifying the refractive index of the coating/analyte.

The CLPFGS period can be defined as a function of length $z$ along the optical fiber (OF) such that it is as follows [11]:

$$\mathsf{\Lambda }=A_0+Cz,$$ (1)

where $A_0$ is the initial period and $C$ is a constant parameter on how the period varies, and its sign will change if it is increasing $C$ (+) or decreasing $C$ (−).

Two configurations of the biosensors were performed as is shown in Figure 1. The first configuration had two CLPFG, both with increasing chirp parameter $C$(+) along the OF (L1 $C$ (+), L2 $C$ (+)), as is shown in Figure 1a. The second configuration was a Mach–Zehnder interferometer (CMZI) consisting of two cascaded CLPFGs, the first one with chirp parameter $C$ (+) and the second CLPFG with chirp parameter $C$ (−), both with the same period chirp parameter value, the same points number, and separated by 1 cm (see Figure 1b). These configurations were selected according to the information reported by Ming Yan et al. [11] in order to obtain multiple attenuation bands and to look for better sensitivity of the biosensors.

Figure 1.

Biosensor structure. (a) The CLPFG biosensor is composed of two cascaded CLPFGs with $C$ (+). (b) CMZI biosensor is composed of two cascaded CLPFG with $C$ (+) and $C$ (−) separated by 1 cm.

In this case, the first CLPFG $C$ (+) couples some light from the core mode to the cladding mode, allowing a better interaction between the analyte and the evanescent wave. Both modes travel the distance between both gratings, and when they reach the second CLPFG $C$ (−), it couples light back to the core, causing the light to interfere. The latter is due to differences in the optical path in the MZI, described by Equation (2).

$$\varphi =2\pi /\lambda (n_{core}-n_{clad}^i)L$$ (2)

The MZI increases resolution and sensitivity due to the factor $(2\pi /\lambda )L$ to measure small changes in the refractive index.

2.2. Feature Scaling

Feature scaling is a mapping technique in a preprocessing stage by which the user scales or transforms all attributes to make an equal contribution of each feature. In some applications, this data transformation can improve the performance of machine learning models [12,13,14,15,16,17].

Consequently, applying a scaling method without careful evaluation of its suitability for the specific problem and model may be inadvisable and could negatively impact results. This practice risks undermining the validity of the claims regarding model performance and may lead to a feedback loop of overconfidence in the results, known as overfitting.

On another hand, with few exceptions, machine learning algorithms do not perform well when the numerical input attributes have very different scales. Scaling target values are generally not required. There are common ways to get all attributes to have the same scale [16].

Standardization: In this case, first subtract the mean from the original data and the result is divided by the standard deviation, thus obtaining a unit variance. With this, the standardized values will always have a zero mean.

$$X_{std}={\displaystyle \frac{X-\mu }{\sigma }}.$$ (3)

Instead, it does not limit the values to a specific range, which could be a problem for some algorithms; however, it is affected by outliers in the data.

Min–max scaling: In min–max scaling, values are shifted and rescaled to range from 0 to 1. This is done by subtracting the minimum value and dividing it by the difference between the maximum and minimum values.

$$X_{scaled}={\displaystyle \frac{X-X_{min}}{X_{max}-X_{min}}}.$$ (4)

Since min–max scaling limits values in a range between 0 and 1, it is more accessible to algorithms; however, under abnormal values, it suffers value shifts.

Custom transformer: It offers the possibility of applying this transformation to our dataset using what is called a Function Transformer. This allows us to apply any custom function to the $X$ data, for example, log(), sqrt(), or any user-defined transformation.

$$X_{scaled}=Function transformer\left(data\right).$$ (5)

Robust scaling: The robust scaler uses the median and interquartile range (IQR) to scale the data. This method is robust to outliers

$$X_{robust}={\displaystyle \frac{X-X_{median}}{IQR}}.$$ (6)

2.3. Principal Component Analysis (PCA)

The purpose of PCA is to derive a small number of linear combinations (principal components) from a set of variables that retain as much information as possible from the original variables. Often, a small number of principal components can be used instead of the original variables for plotting, regression, clustering, and so on. Principal component analysis can also be considered a technique for eliminating multicollinearity in the data. In this technique, the original set of variables is transformed into a new set of uncorrelated variables, called principal components. These new variables are linear combinations of the original ones and are derived in decreasing order of importance, such that the first principal component explains as much of the variation in the original data as possible. It is often expected that the first few components will explain most of the variation in the original data, allowing the effective dimensionality of the data to be reduced [17].

PCA often reveals relationships that were not previously suspected, enabling interpretations that would not normally be obtained. The new axes represent the directions of maximum variability and provide a simpler and more accurate description of the covariance structure. In this case, PCA was applied to the entire dataset containing all the spectra from the biosensor responses. Two principal components were used to train a classifier to detect IL-10.

2.4. Support Vector Machine (SVM)

The general idea behind SVM is to construct a hyperplane to split a set of data [18,19,20,21]. The hyperplane equation is written as follows:

$$\overrightarrow{w}\cdot \overrightarrow{x}+b=0,$$ (7)

where $\overrightarrow{w}$ and constant b are the hyperplane parameters. Figure 2 shows two different hyperplanes to separate two different kinds of data. As can be seen, width margin determines performance of SVM to correctly classify each class. The main target of SVM is to optimize $\overrightarrow{w}$ and b values to maximize margin width.

Figure 2.

Two different hyperplanes to separate different class data. (a) Best hyperplane, (b) not optimized hyperplane.

Classification could be done by just evaluating the sign of the hyperplane:

$$\begin{gathered} \overrightarrow{w}\cdot \overrightarrow{x}+b>0,\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}:+1 \left(\mathrm{squares}\right), \\ \overrightarrow{w}·\overrightarrow{x}+b<0,\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{s}\mathrm{s}:-1 \left(\mathrm{ball}\right). \end{gathered}$$ (8)

The vectors that determine boundaries of the margin are called the support vectors.

Figure 3 shows the workflow of data processing to train the classifier to detect IL-10 protein. First, all spectra were stored on a matrix, where each row was one spectrum with 1001 measurements in the wavelength range of 1375–1575 nm. Then, feature scaling was performed according to the above-mentioned definitions. After that a PCA was performed, and the first two principal components were selected. Finaly, an SVM classifier was trained to classify healthy control class and ISC class. ISCEB was classified by the SVM, trained to determine the expression of IL-10 protein. All processing data was performed with Python 3.10.10 and Scikit-learn libraries.

Figure 3.

Workflow of data processing to detect IL10 protein from biosensor response.

2.5. Methodology

2.5.1. Development of CLPFG and CMZI Biosensors

The CLPFGs were recorded by the point-to-point method using the electric arc technique with a fiber fusion splicer. The method consisted of taking an optical fiber without coating (4 cm long) and placing it on the splicer. One end of the fiber was fixed to a linear actuator (TLA28A, Zaber, Newton, NJ, USA), which has a resolution of 0.1 µm. An electric arc was applied at each displacement until the desired number of points was reached. It consists of 21 points, with an initial period of 400 μm until to a maximum period of 1092 μm ($L1C=L2C\approx 1.34 \mathrm{c}\mathrm{m}$), with a period function described by Equation (9)

$$\mathsf{\Lambda }\left(z\right)=400+4.6\times {10}^{-2}z$$ (9)

where $z$ is the fiber length (in $\mathsf{\mu }\mathrm{m})$, and the chirp parameter is $4.6\times {10}^{-2}$. The CLPFGs were characterized using a superluminescent diode (SLD), with a wavelength range from 1400 nm to 1550 nm, and the spectra were measured using an optical spectrum analyzer (OSA, Ando AQ6315A model 9057 F/8, Tokio, Japan) and stored in a personal computer (PC). The CMZIs formed by two cascaded CLPFGs, with the same length as mentioned above, were fabricated with an electric arc method and a separation of 1 cm. The CLPFG recording and CMZI fabrication setup is shown in Figure 4.

Figure 4.

Experimental setup of CLPFG and CMZI recording.

Once the CLPFG and CMZI were recorded, the self-assembly of the CLPFG and CMZI were performed similarly to our previous works [22]. Figure 5 shows the relevant stages of self-assembly. First, the surface of the optical fiber was cleaned with potassium hydroxide (KOH), which also added -OH groups and removed organic impurities and metals. After that, the hydroxylation stage was made with a solution prepared by dissolving 841 mg of KOH in 7.5 mL of deionized water and then adding 7.5 mL of methanol. After being exposed to this solution for three and a half hours, the optical fiber was dried with nitrogen gas, rinsed with methanol, and active sites of Si-O, Si-Si, and Si-O2-H were produced. On the other hand, the functionalization stage was made with a 2% solution of (3-Aminopropyl) trimethoxysilane (APTMS) dissolved in toluene. The optical fiber was then rinsed with copious amounts of toluene to remove any residue that could interfere with antibody immobilization. This process bonded the APTMS to the Si-O-Si, Si-Si, and Si-O-H bonds, exposing the active NH sites (functional groups). In the activation stage, the OF surface was then activated by combining 5 mg of NHS with 1 mL of phosphate-buffered saline (PBS) solution, followed by 5 mg of EDC with 1 mL of PBS solution. Subsequently, both solutions were mixed, and the final solution was placed on the surface of the functionalized OF, which was already placed on a shaker at 1.66 Hz and made to interact for 1 h. After that, immobilization stage was made with a polyclonal antibody IL-10 (Anti-IL-10) to detect IL-10 protein in the serum samples. The antibody was preserved in PBS solution. Using a plastic Petri dish, 20 μL of Anti-IL-10 and 380 μL of PBS solution were placed, and the functionalized OF was submerged in it. The Petri dish was placed on an orbital shaker at a frequency of 1.66 Hz for 1 h to ensure uniform distribution of the antibody on the surface. Once all active sites were occupied, the amino terminal group (-NH2) of the free antibody was ready to detect IL-10 protein.

Figure 5.

Diagram of the biosensors self-assembly. This process consists of four stages: hydroxylation, functionalization, activation and immobilization [22].

2.5.2. Serum Samples from the Murine Model

Serum samples were provided by Centro de Investigación en Reproducción Animal de la Universidad Autónoma de Tlaxcala-CINVESTAV. These serum samples were extracted from murine models, using male Sprague Dawley rats weighing approximately 350 g and aged 3.5 months. For the expression of IL-10 in the induction of cerebral ischemia, the following procedures were performed: Electrocautery of the vertebral arteries, in which a permanent occlusion of the blood flow projected by these arteries was performed without inducing an ischemic process. The model then remains in recovery for 15 days before continuing with the second preparatory surgery, which was the exposure of the carotid arteries. This surgery was performed to locate and highlight the common carotid arteries to facilitate localization. After the two surgeries, the ischemic protocol was performed to induce a deficit in the blood supply to the brain due to the occlusion of the main vessels that irrigate it, resembling the strokes that occur in humans. Two and four hours post-ischemia, blood serum samples were immediately obtained and stored in EDTA tubes. They were extracted using micropipettes and centrifuged for 15 min at 3000 rpm to decant the solid part and proceed to extract the purified serum, which was frozen until use [23]. This process was repeated on murine models with a treatment EB before ischemia, and after that two samples of serum were obtained at 2 and 4 h. Finally, serum samples from healthy murine models were obtained.

The detection process was performed after immobilizing the polyclonal antibody; 200 µL of serum sample was added onto the biosensor area and interacted for 1 h. Detection was stopped when there were no more changes in the transmission signal. Five CLPFG and five CMZI biosensors were performed. Each kind of biosensor interacted with one of the following serum samples: healthy control, 2 ISC, 4 ISC, 2 ISCEB, and 4 ISCEB. The transmission spectrum was measured every 5 min. The experiment setup for detection is shown in Figure 6. It consisted of a superluminescent diode (SLD1450, Thorlabs, Newton, NJ, USA) with center wavelength at 1465 nm and bandwidth of 50 nm. The biosensors were placed over an orbital rotator to ensure homogeneity of the sample on the biosensor; the other end of the OF was connected to an optical spectrum analyzer (OSA, AQ6317B, Tokio, Japan) with a bandwidth of 600–1750 nm, which was used to measure the biosensors spectra.

Figure 6.

Experimental setup used for detection of IL-10 in CLPFG and CMZI biosensors.

3. Results

Figure 7 shows the transmission spectra of the CMZI biosensor as a function of exposure time to serum samples with ISC and ISCEB. The wavelength range is between 1375 nm and 1575 nm.

Figure 7.

CMZI transmission spectra for (a) 2ISC; (b) 2ISCEB.

Figure 7a,b shows that the biosensor response was barely perceptible to the naked eye. However, an approach around 1471 nm (Figure 7a) and 1465 nm (Figure 7b) displays that there truly are amplitude variations and spectral shifts, but it is not easy to establish a relation between these changes due to the presence or absence of IL-10. There are similar behaviors in the biosensor CLPFG response.

In Figure 8, the raw data PCA results for CLPFG and CMZI biosensors are shown. We can observe that CLPFG biosensor cannot distinguish between ISC and ISCEB samples, since both are in same region in the map. On the other hand, for the case of the CMZI biosensor, even though some data could be correctly classified, there are still some misclassifications.

Figure 8.

Raw data PCA for (a) CLPFG biosensor and (b) CMZI biosensor. In both cases, samples with IL-10 are not well detected.

To improve the detection of IL-10, it is necessary to perform preprocessing by data scaling.

Figure 9 shows results of different kinds of scaling of the original CMZI biosensor spectra. There are similar results for the biosensor based on CLPFG, but they are not shown. Figure 9a shows spectra with min–max scaling. As was already mentioned, these values are bound between zero and one; no region particularly stands out. Figure 9b shows spectra with standardized scaling, in this case compared to the min–max scaling; the data are bound between zero and beyond the unit due to the use of each characteristic variance. In this case, it is visible that there are some highlighted regions overall (1400, 1435, 1500 nm), which could mean that these regions contribute with more information and may help to improve the classification process. Figure 9c shows scaled spectra using robust scaling; in this case, there are different regions that stand out overall. For instance, we can observe some characteristic peaks around 1425, 1475, 1500, 1575 nm to mention some of them, which do not correspond with the peaks of the original spectrum. This could be due to special behavior of the biosensor responses. As explained below, this scaling noticeably enhances the classification process, which again could be related to the information in these bands. Finally, Figure 9d shows scaled spectra using a customized scaling based on the log2() scaling function. In this case, although there are regions that stand out, employing this scaling, as explained below, does not make a remarkable improvement in the classification.

Figure 9.

IL-10 spectrum detection of the biosensor CMZI with (a) min–max scaling; (b) standardized; (c) robust scaling; (d) customized scaling.

After the scaling process, a PCA was performed using all the biosensor spectra. Each spectrum contains 1001 measurements in different wavelengths, which implies 1001 predictors for serum sample detection over time. This number was reduced to only two components by PCA. Each spectrum is represented as one point in the bidimensional PCA space. Figure 10 shows the PCA obtained from the transmission spectra scaling, showing the percentage of total variance explained by the two principal components for CLPFG and CMZI biosensors, with both min–max and standardized scaling.

Figure 10.

PCA with different types of scaling and SVM for IL-10 detection in CLPFG and CMZI biosensors. CLPFG: (a) min–max scaling and (c) standardization. CMZI: (b) min–max scaling and (d) standardization.

In the CLPFG biosensor case, as shown in Figure 10a,c, neither of these two scaling methods enhance the classification performance. We think this is due to the low biosensor sensitivity compared with the CMZI (as has been reported in a previous work [24]) and/or a low IL-10 expression. On the other hand, Figure 10b and d show that the CMZI biosensor scaling along with the SVM classifier can identify serum samples responses in the ISCEB subject as well as serum samples of healthy subjects, which initially can be understood as IL-10 detection. Nonetheless, in the min–max scaling case, where all samples are clustered and close to the decision border, the use of this scaling technique in the detection system could be very susceptible to misclassification detection. On the other hand, standardization further affects all classes; however, the ISCEB samples are still at the decision border, which makes it susceptible to detect errors.

The PCA and SVM for the spectra with robust and customized scaling in CLPFG and CMZI biosensors are shown in Figure 11. Figure 11a,c shows no improvement in the CLPFG biosensor response, like the previous results, since ISCEB data are still in the ISC area. On the other hand, Figure 11b shows robust scaling in the CMZI biosensor response; there is a noticeable improvement, given that both ISCEB measures are clearly closer to the healthy control area and further apart from the border decision, as is expected. On the other hand, Figure 11d shows that the customized scaling in the CMZI biosensor response also enhances the classification of 4ISCEB data, as they are getting further apart from the border decision and closer to the healthy control data. However, the classification of 2ISCEB data can cause confusion during the detection due to the misclassification of these measurements.

Figure 11.

PCA with different types of scaling and SVM for IL-10 detection in CLPFG and CMZI biosensors. CLPFG: (a) robust scaling and (c) customized scaling. CMZI: (b) robust scaling and (d) customized scaling.

It can be concluded that the use of robust scaling applied in the CMZI biosensor response improves the IL-10 expression detection, with a misclassification reduction, given that the data are apart from the border decision established by the SVM classifier and closer to the healthy control data. We think this is because the technique is effective in equalizing the variances among the different attributes even when one of them presents outliers. This happens because when the interquartile range is used as the scaling factor, the presence of the outliers is neglected, since they lie beyond this interval [17]. It is interesting to highlight that standardized scaling does not play a bad role; nevertheless, it might be susceptible to errors due to the closeness to the border decision classifier, where any outlier can cause a false negative.

4. Conclusions

In this work, two types of biosensors were developed using CLPFG and CMZI. CLPFG was developed with two gratings, with a period from 400 μm to 1092 μm and with a $C$ (+) parameter configuration. CMZI used two gratings separated by 1 cm with the same period as CLPFG with $C$ (+) and $C$ (−) parameters for the first and second grating, respectively. Both were functionalized to immobilize the antibody IL-10 to detect IL-10 protein in serum samples of the murine model. The effect of four scaling techniques was studied over the spectra of both biosensors. The results in this application showed that the CLPFG was not able to detect IL-10 expression; we think this is due to the low sensitivity of this biosensor.

On the other hand, the CMZI biosensor is always able to detect IL-10, due to its high sensitivity, present in these kinds of biosensors. Out of all the scaling used in this work, robust scaling allows us to improve the performance of the detection process by relating the spectra measurements in ischemia samples with EB treatment, along with healthy subject samples in the murine model. The same happens in the distinction of ischemia samples without EB treatment that are associated with ill subjects and, therefore, with IL-10 expression deficit.

Finally, it must be emphasized that IL-10 detection using standardization and min–max scaling also has high efficiency; however, our results show that all ISCEB measures are closer to the border decision, which makes them susceptible to detection errors, and this situation is minimized using robust scaling technique.

Abbreviations

IL-10	Interleukin 10
LPFG	Long period fiber grating
CLPFG	Chirped long period fiber grating
CMZI	Chirped Mach–Zehnder interferometer
PCA	Principal component analysis
SVM	Support vector machine
OSA	Optical spectrum analyzer
SLD	Superluminescent diode
ELISA	Enzyme-linked immunosorbent assays
PCR	Polymerase chain reaction
ISC	Ischemia
ISCEB	Ischemia with estradiol benzoate
EB	Estradiol benzoate
PBS	Phosphate-buffered saline
OF	Optical fiber

Author Contributions

Conceptualization, R.I.B.-D., K.G.-L., J.C.-M., S.M.-A. and G.B.-P.; methodology, R.I.B.-D., L.O.-Z., K.G.-L. and G.B.-P.; validation, R.I.B.-D., K.G.-L., J.C.-M., S.M.-A. and G.B.-P.; formal analysis, J.C.-M., S.M.-A. and G.B.-P.; investigation, G.B.-P. and R.D.-M.; resources, S.M.-A., M.G.-J. and G.B.-P., data Curation, R.I.B.-D., L.O.-Z., J.C.-M., S.M.-A. and G.B.-P.; writing-original draft preparation, R.I.B.-D., L.O.-Z., K.G.-L., J.C.-M. and G.B.-P.; writing-review and editing, S.M.-A., J.C.-M., and G.B.-P.; visualization, K.G.-L., S.M.-A., J.C.-M. and G.B.-P.; project administration, S.M.-A., M.G.-J. and G.B.-P., funding acquisition, S.M.-A., M.G.-J. and G.B.-P. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The data sets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding Statement

This research was funded by SECIHTI for the grant CBF-2025-G-91 and VIEP for the grant ID-CVIEP 00049. SEP-CONACYT, CIENCIA BÁSICA-2016, Grant number 286647, and CONAHCyT for the grant CF-2023-G-289.