Application of Portable Near-Infrared Instrument for Analysis of Spirulina platensis Aqueous Extracts

Abstract

Spirulina platensis microalga has become recognized as a promising source of highly nutritious food components to feed the growing global population. Because of its high protein content, abundance of essential amino acids, and excellent digestion, it is employed in human nutrition and there is growing interest in analyzing bioactive compound present in Spirulina platensis microalga. In this work, a portable near-infrared (NIR) spectrometer was used for the monitoring of physicochemical properties (extraction yield (EY), total polyphenols concentration (TPC), total proteins concentration (TP), antioxidant activity measured by (i) the DPPH method (DPPH) and (ii) FRAP method (FRAP)) of Spirulina platensis aqueous extracts. The ultrasound-assisted aqueous extraction (ultrasonic bath with an ultrasound frequency of 35 kHz) of bioactive molecules from Spirulina platensis was performed throughout 30 independent experiments. NIR spectra were recorded in the wavelength range of 900–1700 nm. Raw NIR spectra were subjected to the physicochemical properties applying Principal Component Analysis (PCA), partial least square (PLS), and artificial neural network (ANN) modeling. Results show that ANN models developed for the prediction of TPC and DPPH can be utilized for extraction process control (RER > 10), while the other three models can be employed for screening (RER > 4). Generally, the obtained results indicate significant potential for using portable NIR spectroscopy for the analysis of Spirulina platensis aqueous extracts.

1. Introduction

Spirulina, a blue-green microalga categorized within the cyanobacteria family (Cyanophyta), consists of photosynthesizing prokaryotes characterized by a Gram-negative cell wall structure [1]. In their natural habitat, they exist in both single-celled and filamentous forms, with filaments resembling springs. Spirulina encompasses species such as Spirulina platensis and Spirulina maxima, collectively known under the brand name Arthrospira [2]. Notably, Spirulina stands out for its non-toxic nature and easy digestibility attributed to the absence of cellulose in its cell wall, a trait not shared by other microalgae like Chlorella, Ankistrodesmus, Selenestrum, and Sandesmus [3]. Its prominence in the food industry soared when it was adopted as a vitamin and protein supplement, following its successful use by aerospace organizations to nourish astronauts during space missions [4]. According to Doan et al. [5], in human nutrition, Spirulina finds favor not only for its remarkable protein content, which can reach up to 70% of its dry matter, but also for its composition of essential amino acids (AAs) and its high digestibility, positioning it as a promising alternative protein source. Moreover, proteins derived from microalgae exhibit favorable technological properties, serving as effective foaming, gelling, and emulsifying agents [6,7,8]. Numerous studies have demonstrated the competitive edge of microalgae proteins against commercially used emulsifiers like sodium caseinate, whey protein, and soy protein [9,10]. As described by Fernadez da Silva [11] in addition to proteins, Spirulina platensis is rich in polyunsaturated fatty acids, vitamins (A, B₁, B₂, B₆, B₁₂, E, and D), minerals (especially iron), and other high-added-value phytochemicals including pigments (β-carotene, chlorophyll a, and phycocyanin) and polyphenols.

The intracellular nature of bioactive compounds in microalgae poses a challenge for extraction, as it requires the lysis of cells to access these compounds [12]. Traditional extraction methods like maceration, Soxhlet, solid–liquid extraction (SLE), and liquid–liquid extraction (LLE) are known for their use of large quantities of toxic solvents and extended extraction durations. In response to the drawbacks associated with these conventional techniques, non-conventional methods such as microwave-assisted extraction (MAE) [13] and ultrasound-assisted extraction (UAE) [14] have been investigated for the extraction of bioactive pigments from various microalgae. According to Martinis et al. [15], during ultrasound-assisted extraction (UAE), ultrasonic waves initiate a phenomenon called cavitation. This process entails rapid cycles of compression and expansion waves near the surface of a solid matrix. As the waves decompress, large air bubbles form and expand, ultimately collapsing and imploding, releasing stored energy in the form of waves. These micro-explosions generate microscopic channels within tissues, resembling a sponge-like effect, facilitating the penetration of the solvent and the release of the desired compounds more efficiently.

In extraction processes, it is necessary to select an analytical method that allows for simple, nondestructive, and on-line measurement of the compounds of interest. In the past three decades, spectroscopic methods have emerged for predicting the content of chemical compounds in a non-destructive, rapid, effective, and efficient way, with high sensitivity, minimal sample requirements, and versatility in analyzing samples from various matrices in an environmentally friendly manner [16,17]. Near-infrared (NIR) spectroscopy, a technique based on the interaction of NIR wave electromagnetic radiation at specific wavelengths with constituent compounds in organic materials such as carbohydrates, proteins, fats, and water, has gained prominence [18]. The NIR spectrum detects vibrations within the near-infrared region ranging from 800 to 2500 nm wavelengths, where absorption and reflectance occur at varying intensities due to a combination of vibrations resulting from the presence of molecular bonds such as O–H, C–O, C–H, and N–H in the near-infrared region. In addition to its numerous advantages, NIR spectroscopy has one limitation: it necessitates the development of a multivariate chemometrics calibration model to predict the content of chemical compounds [19]. The objective of chemometrics is to enhance the accuracy of obtained analytical results. Practical aspects of chemometrics or multivariate analysis encompass spectral pre-processing, wavelength (variable) selection, data dimension reduction, quantitative calibration, pattern recognition, calibration transfer, calibration maintenance, and multispectral data fusion [20]. While most multivariate calibration methods, like partial least square (PLS) modeling, assume a linear relationship between independent and dependent variables, this assumption may not always hold true, especially when dealing with nonlinear data. Traditional methods often rely on specific wavelengths to quantify compounds, which can overlook nonlinearities in the data [21]. To address this challenge, nonlinear models such as artificial neural networks (ANNs) can be employed. ANNs are self-adaptive and massively parallel machine learning systems, modeled after biological network structures (neurons) [22]. They possess the ability to store experiential knowledge and recognize cause–effect relationships through training for multiple input–output systems. This makes them capable of representing even the most complex systems. ANNs offer advantages such as universal approximation capability to model almost all types of nonlinear functions, including quadratic functions. However, it is important to note that the use of ANNs comes with the limitation that no global optimal solution can be guaranteed. There are several examples of efficient applications of NIR coupled with ANN modeling for the prediction of physicochemical properties of plant extracts. For example, Valinger et al. [23] developed reliable ANN models for rapid quantification of physical and chemical properties of industrial hemp extracts based on combined UV-VIS-NIR spectra, Aghdamifar et al. [24] applied ANN modeling for classification and quantification of coffee beans on VIS-NIR spectra, Sharabiani et al. [25] predicted wheat leaf chlorophyll content based on VIS-NIR spectra using ANN, while Pissard et al. [26] used NIR for the prediction of phenolic content and dry matter in peel and flesh of fresh apples.

Based on the aforementioned, the aim of this work was to assess the potential of a portable NIR spectrometer with PLS modeling and ANN modeling for the analysis of Spirulina platensis aqueous extracts specifically, extraction yield, total polyphenols, and antioxidant activity measured by DPPH and FRAP method and protein concentration. To our best knowledge, this is the first time a portable NIR spectrometer was used for the analysis of Spirulina platensis aqueous extracts.

2. Materials and Methods

2.1. Algae and Chemicals

Dried Spirulina platensis powder was procured from Nutrigold (Zagreb, Croatia). TPTZ (2,4,6-tris(2-pyridyl)-s-triazine), gallic acid (98%), iron (II) sulfate heptahydrate, DPPH (1,1-diphenyl-2-picrylhydrazyl), Trolox (6-hydroxy-2,5,7,8-tetramethylchromane-2-carboxylic acid), and sodium chloride were obtained from Sigma-Aldrich Chemie (Steinheim, Germany). Hydrochloric acid (30%), iron (III) chloride hexahydrate, sodium carbonate, and sodium chloride were purchased from Gram-Mol d.o.o. (Zagreb, Croatia). Sodium acetate trihydrate was obtained from J.T. Baker (Deventer, The Netherlands). Sodium hydrogen carbonate was obtained from Franck (Zagreb, Croatia). Folin–Ciocalteu reagent, disodium hydrogen phosphate, and sodium dihydrogen phosphate dihydrate were obtained from Kemika d.d. (Zagreb, Croatia), acetic acid was purchased from T.T.T. d.o.o. (Sveta Nedjelja, Croatia), while methanol was obtained from Carlo Erba Reagents S.A.S. (Val de Reuil, France). All chemicals were of analytical reagent grade.

2.2. Methods

2.2.1. Preparation of Spirulina platensis Extracts

The ultrasound-assisted extraction of bioactive compounds from Spirulina platensis was conducted using an ultrasonic bath (DT 103 H, Bandelin Electronics, Berlin, Germany). Experiments were conducted with varying solid-to-liquid ratios (S/L = 15, 25, and 35 g/L), different pH levels (pH = 7, 8.5, and 10), various temperatures (T = 25, 35, and 45 °C), and different extraction durations (t = 20, 40, and 60 min) according to the experimental plan given in

Table 1. Experimental conditions for the preparation of Spirulina platensis extracts.

Exp.	S/L (g/L)	pH	T (°C)	t (min)
1.	15	8.5	35	20
2.	35	8.5	35	20
3.	15	8.5	35	60
4.	35	8.5	35	60
5.	25	7	25	40
6.	25	10	25	40
7.	25	7	45	40
8.	25	10	45	40
9.	25	8.5	35	40
10.	15	8.5	25	40
11.	35	8.5	25	40
12.	15	8.5	45	40
13.	35	8.5	45	40
14.	25	7	35	20
15.	25	7	35	60
16.	25	10	35	20
17.	25	10	35	60
18.	25	8.5	35	40
19.	15	7	35	40
20.	35	7	35	40
21.	15	10	35	40
22.	35	10	35	40
23.	25	8.5	25	20
24.	25	8.5	25	60
25.	25	8.5	45	20
26.	25	8.5	45	60
27.	25	8.5	35	40
28.	25	8.5	35	40
29.	25	8.5	35	40
30.	25	8.5	35	40

. Extraction conditions were selected based on the available literature data [14,27,28]. In total, 30 extracts were prepared. A specific mass of dried algae was dispersed in a specific volume of deionized water with adjusted pH in a 50 mL flask, and extraction was carried out. Following extraction, samples were centrifuged at 8000 rpm for 10 min, and the supernatant was utilized for analysis. All samples were stored at 4 °C prior to analysis.

2.2.2. Extraction Yield Measurement

The extraction yield of extracts, expressed as the total dry matter, was determined using the standardized drying method at 105 °C [29]. Quartz sand was placed into a metal container to cover the bottom, and the sand was dried at a temperature of 105 °C for 1 h. After drying, the containers were placed in a desiccator and weighed after cooling to room temperature. Then, approximately 3 mL of Spirulina platensis extract was added to the container with sand, weighed, and dried at 105 °C for 4 h. After drying the samples for 4 h, the containers were cooled to room temperature in the desiccator and weighed again. Measurements were conducted in five repetitions, and the results are presented as the mean value ± standard deviation.

2.2.3. Total Phenolic Content of the Extracts Measurement

The total polyphenol content (TPC) in the Spirulina platensis extracts was determined spectrophotometrically, following the method described by Pinelo et al. [30] which relies on the colorimetric reaction between phenol and the Folin–Ciocalteu reagent. Briefly, 7.9 mL of distilled water was mixed with 500 μL of Folin–Ciocalteu reagent (Folin–Ciocalteu reagent–water ratio of 1:2) and 100 μL of the sample. The reaction was initiated by adding 1.5 mL of a 20% Na₂CO₃ solution. After 2 h of incubation in a dark environment, the absorbance of the reaction mixture was measured at λ = 765 nm using a spectrophotometer (Biochrom Libra S11, Cambridge, UK). The polyphenol content was quantified using a calibration curve for gallic acid (GA) ranging from 0 to 500 mg/L. Measurements were conducted in five repetitions, and the results were expressed as mg GA equivalents (GAE)/g dry weight (dw).

2.2.4. Antioxidant Activity of the Extracts Measurement

Antioxidant activity was determined using the DPPH method [31], where the reaction mixture consisted of 100 μL of extract sample and 3.9 mL of DPPH radical (c = 0.094 mmol/L) dissolved in methanol. After homogenization and 30 min of incubation, the absorbance of the reaction mixture was measured at λ = 515 nm. Measurements were performed in five repetitions, and the results were derived from a calibration curve for Trolox (0–1 mmol/L) and expressed as mmol Trolox equivalents/g dry weight (dw).

Furthermore, antioxidant activity was measured by the FRAP method [32]. The FRAP reagent was prepared by combining V = 25 mL of acetate buffer (c = 300 mmol/L), V = 2.5 mL of TPTZ solution (c = 10 mmol/L), and V = 2.5 mL of iron (II) chloride hexahydrate solution (c = 20 mmol/L). The reaction mixture consisted of V = 50 µL of the extract and V = 950 µL of FRAP reagent. After incubation for 4 min, the absorbance was measured at λ = 593 nm. Antioxidant capacity was calculated from the calibration curve and expressed as mmol FeSO₄·7H₂O equivalents per g dry weight. Measurements were performed in five repetitions, and the results are presented as the mean value ± standard deviation.

2.2.5. Measurement of Protein Concentrations in the Extracts

Protein concentrations were assessed using the Bradford colorimetric method, employing bovine serum albumin (BSA) as a standard at a concentration of 1 mg/mL [33]. Each sample, 0.5 mL in volume, was mixed with 0.5 mL of Bradford reagent, and after 30 min of incubation, the absorbance was measured at 595 nm using a UV–VIS spectrophotometer Biochrome Libra S12 (Biochrom Ltd., Cambridge, UK). The data were expressed as mg/L. Measurements were conducted in five repetitions, and the results are presented as the mean value ± standard deviation.

2.2.6. NIR Spectra Gathering

A portable NIR spectrometer (NIR-M-R2, InnoSpectra, Hsinchu, Taiwan) was utilized to collect spectra of Spirulina platensis extracts, recording in diffuse reflective mode in wavelengths in the range of 900–1700 nm. The spectra were analyzed using ISC-NIRScan software v. 3.10 (InnoSpectra, Taiwan). The NIR instrument used has an optical resolution of 10 nm and a signal-to-noise ratio of 5000:1. It employs two 0.7 W tungsten filament lamps as the illumination source. The samples were placed in the quart cuvette for spectra gathering. The optical path length was 10 mm. For each prepared extract, NIR spectra were recorded five times.

2.2.7. Basic Statistical Analysis and Correlation Matrix

All measurements of the properties of Spirulina platensis extracts were conducted in triplicate, and basic statistical analysis (including mean values, standard deviation, ranges, and coefficients of variation (CVs)) was performed using the Statistica 14.0 software package (TIBCO^® Statistica, Palo Alto, CA, USA). The correlations or associations between the extraction conditions and Spirulina platensis extract properties were analyzed using Spearman correlation matrices in the Statistica 14.0 software package. Spearman’s matrix was utilized due to the analysis showing that the data did not follow a normal distribution.

2.2.8. Partial Least Square (PLS) Modeling

Partial least square (PLS) models for the prediction of the extraction yield, total phenolic content, antioxidant activity measured by the DPPH method, antioxidant activity measured by the FRAP method, and protein concentration based on the raw NIR spectra were developed in Unscrambler X 10.4 software (CAMO Software, AS, Oslo, Norway). Model development and validation were performed using NIR wavelengths in the range of 1394–1699 nm. The dataset was randomly split into calibration and prediction datasets in a 70:30 ratio. Within the calibration dataset, a random cross validation method implemented into Unscrambler X 10.4 software was applied. The dataset was divided into 20 segments and each segment included 7 samples. Applicability of the developed models was estimated based on the following: (i) the coefficients of determination for calibration (R_cal²), cross-validation (R_cval²), and prediction (R_pred²), (ii) root mean square error for calibration (RMSEC), cross-validation (RMSECV), and prediction (RMSEP), (iii) standard error for calibration (SEC), cross-validation (SECV), and prediction (SEP), and (iv) average value of the difference between predicted and measured values (Bias).

2.2.9. Artificial Neural Network (ANN) Modeling

Artificial neural networks (ANN) modeling was employed for predicting properties of Spirulina platensis extracts, including extraction yield, total phenolic content, antioxidant activity measured by the DPPH method, antioxidant activity measured by the FRAP method, and protein concentration, based on the raw NIR spectra collected by a portable NIR instrument.

ANN models, specifically Multiple Layer Perceptron (MLP) networks, were developed using the software Statistica 14.0. (TIBCO^® Statistica, Palo Alto, CA, USA). The ANNs consisted of three layers: input, hidden, and output. The input layer comprised 5 neurons representing the coordinates of the first five factors obtained from PCA analysis. These five principal components accounted for more than 99.99% of the data variability and were selected as inputs. The number of neurons in the hidden layer varied between 4 and 13 and was randomly selected by the algorithm. The activation functions for both the hidden and output layers were randomly chosen from the following set: Identity, Logistic, Hyperbolic Tangent, and Exponential.

Two types of ANN models were developed for predicting properties of Spirulina platensis extracts:

(1)

Simultaneous Output ANN Model: In this model, all physicochemical properties of the Spirulina platensis extracts were used as outputs simultaneously. Each property, including extraction yield, total phenolic content, antioxidant activity measured by the DPPH method, antioxidant activity measured by the FRAP method, and protein concentration, was predicted together using the same ANN architecture.

(2)

Individual Output ANN Models: In these models, each physicochemical property of the Spirulina platensis extracts was used as the output individually. Separate ANN models were developed for each property, allowing for independent prediction of each property using dedicated ANN architectures.

The dataset for ANN development had dimensions of 150 × 10, where 150 rows represented the total number of Spirulina platensis extract samples, 5 columns referred to the 5 PCA coordinates (factors), and 5 columns represented the measured properties of the Spirulina platensis extract samples.

A total of 2000 networks were generated for each output. Model training was conducted using the back error propagation algorithm, with the error function being the sum of squares. The dataset was randomly split into calibration and prediction datasets in a 70:30 ratio. Within the calibration dataset, 70% was allocated for network training, 15% for network testing, and 15% for model validation. The backpropagation algorithm was utilized for model training. To avoid model overfitting, subsampling cross-validation was applied. The random subsampling method and the seed for subsampling of 1000 were used.

The developed calibration models’ applicability was evaluated using the coefficient of determination for calibration (R_cal²), the adjusted coefficient of determination for calibration (R_cal²_adj), and the root mean square error for calibration (RMSE). Prediction performance of the models was assessed based on the coefficient of determination for prediction (R_pred²), the adjusted coefficient of determination for calibration (R_pred²_adj), the root mean square error for prediction (RMSEP), the standard error of prediction (SEP), the ratio of prediction to deviation (RPD), and the ratio of the error range (RER) [34].

3. Results and Discussion

3.1. Statistical Analysis of the Spirulina platensis Extracts’ Properties

Descriptive statistics of the measured results of the extraction yield, total polyphenols, antioxidant activity measured by the DPPH method, antioxidant activity measured by the FRAP method, and protein concentration in all 150 samples are presented in

Table 2. Descriptive statistics of Spirulina platensis extracts’ properties for calibration set and prediction set.

	Calibration Set (105 Samples)			Prediction Set (45 Samples)
	Mean ± st.dev.	Range	CV (%)	Mean ± st.dev.	Range	CV (%)
EY (%)	2.889 ± 2.454	0.043–9.178	84.944	2.873 ± 2.464	0.320–9.178	85.765
TPC (mgGAE/gdw)	11.621 ± 4.354	4.548–30.051	37.467	12.765 ± 5.900	5.900–30.235	39.453
DPPH (mmolTrolox/gdw)	0.016 ± 0.007	0.002–0.031	43.206	0.014 ± 0.006	0.003–0.030	42.660
FRAP (mmolFeSO4 7H2O/gdw)	0.009 ± 0.006	0.002–0.035	59.165	0.010 ± 0.002	0.002–0.035	64.916
TP (mg/mL)	43.100 ± 6.825	26.575–60.15	15.856	42.375 ± 7.125	32.05–60.150	16.797

. The dataset was randomly divided into a calibration set (105 samples) and a prediction set (45 samples). It is notable that all analyzed physicochemical properties of the samples exhibit relatively high coefficients of variation. For all variables except protein concentration, the coefficient of variation was greater than 20% for both the calibration and prediction datasets. The extraction yield ranged from 0.043% to 9.178%, while the total polyphenol concentration in prepared extracts ranged from 4.548 to 30.051 mgGAE/gdw. Similar results were present by Aouir et al. [35] when analyzing the biochemical composition of different strains of Spirulina platensis from Algeria, Chad, and the USA. They reported total polyphenol concentrations ranging from 20 to 70 mgGAE/gdw, depending on the strain. Likewise, Alshuniaber et al. [36] reported that the amount of phenolic compounds in the Spirulina extract samples they analyzed ranged from 7.80 to 44.48 mgGAE/gdw. The authors explained that the quantity of phenolic compounds is influenced by the type of solvent and extraction conditions. Interestingly, Kumar et al. [37] reported three times higher concentrations of polyphenols in Spirulina aqueous extracts compared to Spirulina ethanolic extracts. The high phenolic compound content in the aqueous extracts may be correlated with significant antioxidant activity.

The antioxidant activity of the Spirulina extracts prepared in this work ranged from 0.002 to 0.031 mmol_Trolox/g_dw for the DPPH method and from 0.002 to 0.035 mmol_FeSO4·7H2O/g_dw for the FRAP method. These results are similar to those presented by Chu et al. [38] who indicated the protective effect of Spirulina extracts against free radicals. Abdel-Moneim et al. [39] compared antioxidant activity of methanol, acetone, and hexane Spirulina extract and showed that methanolic extracts had higher antioxidant activity than other extracts. Protein concentrations were also measured for the prepared Spirulina extract. The protein concentrations were in the range from 26.575 to 60.15 mg/mL. The extracts isolated from Spirulina platensis by Bleakley and Hayes [40] contained 85.50 ± 4.90% of proteins, similar to the results obtained in this work.

The relationship between extraction conditions and extract properties was analyzed using the Spearman correlation matrix (

Table 3. Spearman correlation matrix of relationship between Spirulina platensis extracts’ input and output variables. Significant correlations are marked with red color.

	S/L	t	pH	T	EY	TPC	DPPH	FRAP	TP
S/L	1.0000
t	−0.0000	1.0000
pH	−0.0000	0.0000	1.0000
T	0.0000	0.0000	−0.0000	1.0000
EY	−0.1911	−0.4946	0.0889	−0.4148	1.0000
TPC	−0.0231	0.0573	0.2375	−0.2181	0.0933	1.0000
DPPH	0.3483	0.0099	−0.0521	0.1070	−0.0525	0.0019	1.0000
FRAP	0.2930	−0.2350	0.3218	−0.1236	0.2877	0.0002	0.4019
TP	−0.0930	−0.3947	0.3498	−0.1780	0.2674	0.2433	−0.0670	0.2430	1.0000

). It can be noticed that the solid-to-liquid ratio has a significant negative effect on the extraction yield (r = −0.1911) and significant positive effect on the antioxidant activity measured by the DPPH (r = 0.3483) and FRAP methods (r = 0.2930). It is also evident that extraction time has a significant negative effect on extraction yield (r = −0.4946), antioxidant activity measured by FRAP (r = −0.2350), and protein concentration (r = −0.3947). Additionally, according to the correlation matrix, the extraction solvent pH has a significant positive effect on total polyphenolic content (r = 0.2375), antioxidant activity measured by the FRAP method (r = 0.3218), and protein concentrations (r = 0.3498). Furthermore, the results indicate that extraction temperature has a significant negative effect on extraction yield (r = −0.4148), total polyphenolic content (r = −0.2181), and protein concentrations (r = −0.1780). Analyzing the correlations between extraction conditions and extract properties can aid in selecting process conditions that ensure optimal composition of the prepared extract.

3.2. NIR Spectra and PCA Analysis of the Spirulina platensis Extracts

NIR spectra were recorded for all 30 prepared Spirulina extracts with five repetitions in the wavelength range from 900 to 1700 nm. All recorded raw NIR spectra are presented in Figure 1. It can be noticed that all recorded spectra follow the same pattern. There are some differences regarding the maximum absorbance of the recorded samples. The increase in the absorbance can be noticed in the wavelength range from 1100 to 1200 nm and in the range from 1600 to 1700 nm. The wavelength range from 1100 to 1200 nm corresponds to the C–H second overtone (C–H, CH₂, and CH₃ bonds) while the wavelength range from 1600 to 1700 nm corresponds to the O–H first overtone and C–H first overtone [41]. These overtones are anharmonic, meaning that they do not behave simply, which complicates NIR spectra and renders them not directly interpretable as in other spectral regions. The absorbance peak maximum at the wavelength region from 1600 to 1700 nm could be explained by the selection of the extraction solvent. Currently, there is no available NIR spectra of Spirulina aqueous extract in the literature. An example can be found of using VIS-NIR spectroscopy for the discrimination of Spirulina plantensis in media [42] and also an example of using VIS-NIR spectroscopy for the quantification of Spirulina powder adulteration [43].

Taking into account the high similarity of obtained NIR spectra of Spirulina extracts, detailed interpretation of the NIR spectra requires chemometric tools. In this work, principal component analysis was applied to reduce the data dimension and to analyze the differences between the samples. According to Valinger et al. [23], PCA significantly reduces the number of variables by capturing the most variability. It starts with the first factor, which contains the greatest variance explained, and progresses through subsequent factors, with each successive factor explaining progressively less variance. After conducting PCA of the entire raw spectral dataset, a score plot of the first three principal components is presented for an initial inspection of the spectral data (Figure 2a). This plot demonstrates the potential for discrimination between the samples. It can be noticed that the first three principal components contribute to 98.21% of data variability. The first principal component contributes 88.30%, the second principal component 6.77%, and the third 3.13%. The results also showed that the first five PCs have eigenvalues greater than 1 (PC1-201.3205, PC2-15.3289, PC3-7.1515, PC4-3.3501, and PC5-1.2649), and those five components explain over 99.9% of the overall data variation. The score plot shows samples grouping in three specific groups based on the extraction temperature. The obtained results confirm the previously presented statement about the importance of the extraction temperature on the Spirulina extracts’ properties. Generally, interpreting PCA loadings involves understanding the relationship between variables and principal components [44]. The PC loadings show the dominant peaks influencing the trends in the scores plot. Loadings represent the correlation between the original variables and the principal components; (i) the magnitude (absolute value) of a loading indicates the strength of the correlation between the variable and the principal component. Larger absolute values suggest stronger relationships. (ii) The sign of the loading (+ or −) indicates the direction of the relationship. Positive loadings indicate a positive correlation, meaning when one variable increases, the other tends to increase as well. Negative loadings indicate an inverse correlation, meaning when one variable increases, the other tends to decrease. (iii) Variables with higher loadings contribute more to the principal component. Variables with lower loadings have less influence on the principal component. The wavelength regions having the largest deviation from zero are the most responsible for score values of the principal components; thus, the assigned peaks indicate the absorptions causing the difference between the extract samples. In this plot, the absorption bands that are important for discrimination could be identified. The loading plot (Figure 2b) shows that PC-1 is positively correlated with a wavelength range from 900 to 1700 nm, while PC-2 is positively correlated a wavelength range from 900 to 1624 nm. Maximum loadings for PC-1, PC-2, and PC-3 were obtained for the 1550–1700 nm interval representing characteristic C–H and O–H first overtones.

3.3. PLS Modeling of the Spirulina platensis Extracts’ Properties Based on the NIR Spectra

Partial least square regression was applied for calibration and prediction of the extraction yield, total phenolic content, antioxidant activity measured by the DPPH method, antioxidant activity measured by the FRAP method, and protein concentration based on the raw NIR spectra. The matrix dimension of all data used for calculation was 233 (selected wavelengths (228) and analyzed properties (5)) × 150 (analyzed samples). Model performance was evaluated based on the following: (i) the coefficients of determination for calibration (R_cal²), cross-validation (R_cval²), and prediction (R_pred²), (ii) root mean square error for calibration (RMSEC), cross-validation (RMSECV), and prediction (RMSEP), (iii) standard error for calibration (SEC), cross-validation (SECV), and prediction (SEP), and (iv) average value of the difference between predicted and measured values (bias). Results presented in

Table 4. PLS models for the prediction of Spirulina platensis extracts’ properties based on the raw NIR spectra.

	Calibration			Cross-Validation			Prediction
	Rcal2	RMSEC	SEC	Rval2	RMSECV	SECV	Rpred2	RMSEP	SEP	Bias
EY	0.0091	2.3793	2.3912	0.0055	2.4292	2.4414	0.0016	2.5403	2.5569	−0.2143
TPC	0.1482	0.8381	0.8423	0.0821	0.8754	0.8798	0.1556	0.8484	0.8570	0.0046
DPPH	0.0349	0.0063	0.0063	0.0041	0.0064	0.0064	0.0028	0.0067	0.0067	−0.0001
FRAP	0.4798	0.0041	0.0041	0.3501	0.0047	0.0047	0.3218	0.0049	0.0050	−0.0004
TP	0.0259	0.2407	0.2149	0.1699	0.2573	0.2586	0.2266	0.2338	0.2614	−0.0065

show low correlation coefficients for calibration, cross-validation, and prediction for all analyzed variables (R_cal² < 0.48, R_cval² < 0.35, R_pred² < 0.32). A low R² means that it is not possible to explain much variance in the dependent variables and therefore a nonlinear modeling approach should be applied to describe the analyzed variables. RMSEC, RMSECV, and RMSEP along with SEC, SECV, and SEP were slightly increasing going from calibration to prediction, indicating good robustness of the NIR models [45].

3.4. ANN Modeling of the Spirulina platensis Extracts’ Properties Based on the NIR Spectra

As previously mentioned, NIR spectroscopy offers non-destructive, rapid, effective, efficient, and environmentally friendly analysis, with high sensitivity and minimal sample requirements. However, to accurately analyze macromolecules within a specific matrix, NIR spectra need to be processed using chemometric tools. In this study, NIR spectra collected using a portable NIR spectroscope were modeled using an ANN approach. ANN models were developed using raw NIR spectra, simplifying the modeling procedure and enabling the development of reliable models for future process monitoring and optimization. Due to their high dimensionality and potential high correlation, analyzing and interpreting the NIR data can be challenging [46]. Preprocessing the data is often necessary to mitigate scattering effects, thereby enhancing the absorption components associated with chemical properties [47,48]. At times, both the scattering and absorption properties are crucial as they collectively elucidate the physicochemical properties of the samples. Consequently, preprocessing data to minimize scattering information in the spectra might result in models that perform inadequately [48]. Therefore, in this work, the potential of raw NIR spectra for the prediction of physicochemical properties of the Spirulina platensis extracts was analyzed.

In this work, two types of ANN models were developed: (i) a simultaneous output ANN model, where all physicochemical properties of the Spirulina platensis extracts were used as outputs simultaneously, and (ii) individual output ANN models, where each physicochemical property of the Spirulina platensis extracts was used as the output individually. NIR spectroscopy commonly employs partial least squares (PLS) regression analysis to construct linear regression models for predicting parameters of interest [48]. Nonetheless, linear regression models may not always effectively predict parameters unrelated to a specific compound or class of similar compounds. In such instances, utilizing nonlinear models like artificial neural networks (ANNs) has proven advantageous over PLS models for developing optimal prediction models [49,50]. As mentioned before, inputs for ANN models were coordinates of the first five factors obtained from PCA analysis. The use of scores reduced input nodes, so the training time of the network was shortened [51]. The architecture and prediction properties of both types of ANN models are presented in

Table 5. ANN model for simulations’ prediction of Spirulina platensis extracts’ properties based on the raw NIR spectra.

Network	Calibration					Prediction
	Training Perf./Error	Test Perf./Error	Validation Perf./Error	Hidden Activation	Output Activation	Output	Rpred2	Rpred2adj	RMSEP	SEP	RPD	RER
MLP 5-11-5	0.9649 1.0018	0.8709 2.1231	0.8674 3.0579	Tanh	Identity	EY	0.7607	0.7265	0.7091	0.1069	2.003	6.1439
						TPC	0.8266	0.8142	0.8714	0.1314	2.0237	8.8296
						DPPH	0.8476	0.8367	0.0012	0.0002	2.4848	9.5606
						FRAP	0.7480	0.7300	0.0020	0.0003	1.9072	8.2517
						TP	0.6146	0.5871	0.1635	0.0246	1.5289	5.1176

and

Table 6. ANN models for individual prediction of Spirulina platensis extracts’ properties based on the raw NIR spectra.

		Calibration					Prediction
Output	Network	Training Perf./Error	Test Perf./Error	Validation Perf./Error	Hidden Activation	Output Activation	Rpred2	Rpred2adj	RMSEP	SEP	RPD	RER
EY	MLP 5-6-1	0.9971 0.0276	0.9862 0.1019	0.9860 0.1654	Logistic	Tanh	0.8355	0.8238	0.5149	0.0776	2.1519	8.4175
TPC	MLP 5-6-1	0.9872 0.3191	0.9827 0.4683	0.9824 0.5152	Logistic	Identity	0.9287	0.9236	0.5439	0.0819	3.7687	14.1618
DPPH	MLP 5-7-1	0.9889 0.0001	0.9823 0.0001	0.9762 0.0002	Logistic	Identity	0.8399	0.8285	0.0011	0.0002	2.4046	9.8211
FRAP	MLP 5-10-1	0.9946 0.0001	0.9921 0.0001	0.9746 0.0002	Logistic	Logistic	0.8143	0.8011	0.0016	0.0002	2.3143	11.4033
TP	MLP 5-10-1	0.8156 0.0111	0.7647 0.0126	0.7442 0.0279	Logistic	Exponential	0.7155	0.6952	0.1440	0.0211	1.8339	6.9056

, while the comparison between experimental data and ANN model predicted data is shown in Figure 3 and Figure 4. Table 3 displays the ANN model that was identified as optimal for simultaneously predicting Spirulina extracts’ properties, based on their R² and RMSE values for the training, test, and validation datasets, while also taking into account the number of neurons in the hidden layer. A lower number of neurons in the hidden layer was deemed advantageous as it indicates a simpler network structure [52]. MLP 5-11-5 was selected as optimal. That network was characterized by Tanh as the hidden activation function and the Identity function as the output activation function. Results showed that the best agreement between the experimental data and the data predicted by the ANN model was obtained for antioxidant activity measured by the DPPH method (R_pred² = 0.8476, R_pred²_adj = 0.8367, RMSEP = 0.0012, SEP = 0.0002%, RPD = 2.4848, RER = 9.5605 (Figure 3c)), followed by total phenolic content (R_pred² = 0.8266, R_pred²_adj = 0.8142, RMSEP = 0.8714, SEP = 0.1314%, RPD = 2.0237, RER = 8.8296 (Figure 3b)) and antioxidant activity measured by the FRAP method (R_pred² = 0.7480, R_pred²_adj = 0.7300, RMSEP = 0.0020, SEP = 0.0003%, RPD = 1.9072, RER = 8.2517 (Figure 3d)). The highest dissipation between model and experimental data was obtained for the protein concentration (R_pred² = 0.6146, R_pred²_adj = 0.75871, RMSEP = 0.1635, SEP = 0.0246%, RPD = 1.5289, RER = 5.1176 (Figure 3e)). The developed models’ applicability was estimated based on R_pred², RPD, and RER values. As previously described by Hair et al. [52], an R²_pred value of 0.75 is considered substantial, an R²_pred value of 0.50 is considered moderate, and an R²_pred value of 0.26 is considered weak. Models with RPD values below 1.4 are deemed unreliable, those falling between 1.4 and 2 are considered fair, while models surpassing an RPD value of 2 are labeled as excellent models [53]. Additionally, models with RER values exceeding 4 are suitable for data screening, those with RER values surpassing 10 can be utilized for quality control, and models with RER values exceeding 15 are applicable for quantification purposes [54,55]. Therefore, based on R_pred², the developed ANN can be considered substantial for the prediction of extraction yield, total phenolic content, and antioxidant activity measured by the DPPH method, while for the prediction of antioxidant activity measured by the FRAP method and protein concentration, the developed ANN model can be considered moderate. Based on the RPD value, the developed ANN can be considered excellent for the prediction of extraction yield, total phenolic content, and antioxidant activity measured by the DPPH method, while for prediction of antioxidant activity measured by the FRAP method and protein concentration, the developed ANN model can be considered fair. Furthermore, based on the RER value, the developed ANN model can be used for data screening, so additional improvement is needed to develop a model that can be used for quality control or even for quantification.

Analyzing the performance of the ANN models developed for the individual prediction of Spirulina extracts’ properties (Table 6 and Figure 4) based on the NIR spectra, it is evident that the best agreement between the experimental data and the data predicted by the ANN model was achieved for total phenolic content (R_pred² = 0.9287, R_pred²_adj = 0.9236, RMSEP = 0.5439, SEP = 0.0819%, RPD = 3.7687, RER = 14.1618 (Figure 4b)), followed by antioxidant activity measured by the DPPH method (R_pred² = 0.8399, R_pred²_adj = 0.8285, RMSEP = 0.0011, SEP = 0.0002%, RPD = 2.4046, RER = 9.8211 (Figure 4b)) and antioxidant activity measured by the FRAP method (R_pred² = 0.8143, R_pred²_adj = 0.8011, RMSEP = 0.0016, SEP = 0.0002%, RPD = 2.3143, RER = 11.4033). The highest dissipation between model and experimental data was obtained again for the protein concentration (R_pred² = 0.7155, R_pred²_adj = 0.6952, RMSEP = 0.1440, SEP = 0.0211%, RPD = 1.8339, RER = 6.9056 (Figure 4e)). Based on the presented results, it can be concluded that ANN models developed for the prediction of TPC and DPPH can be utilized for extraction process control (RER > 10), while the other three models can be employed for screening (RER > 4). Generally, the obtained results indicate significant potential for using portable NIR spectroscopy for the analysis of Spirulina platensis aqueous extracts.

As mentioned before, there are no available examples of using NIR spectroscopy for monitoring blue-green microalga extract in the literature. Wu et al. [43] applied VIS and NIR spectra for the quantification of Spirulina powder adulteration with flour and mungbean powder. The authors compared the efficacy of the partial least squares (PLS) and least-squares support vector machine (LS-SVM) modeling and showed that both modeling approaches were adequate in predicting either flour or mungbean-powder adulterant in Spirulina powder. Bataller et al. [56] developed PLS models for individual rapid prediction of biomolecules (proteins, carbohydrates, and lipids) in Spirulina platenisis based on the Fourier transform infrared (FTIR) spectra. Their results showed that PLS modeling was suitable for the prediction of proteins and carbohydrates, but it was not suitable for the prediction of lipid content, possibly due to its low lipid content, small sample size, and overlapping bands of amide and C=O ester bonds.

4. Conclusions

In this work, the potential of a portable NIR spectrometer coupled with ANN modeling was applied for the analysis of Spirulina platensis aqueous extracts specifically, extraction yield, total polyphenols, and antioxidant activity measured by the DPPH and FRAP methods and protein concentration. Based on the presented results, it can be concluded that ANN models developed for the prediction of TPC and DPPH can be utilized for extraction process control (RER > 10), while the other three models can be employed for screening (RER > 4). Generally, the obtained results indicate significant potential for using portable NIR spectroscopy for the analysis of Spirulina platensis aqueous extracts. However, in order for the established model to achieve more accurate results in later application, in future experiments, it is necessary to collect more batches of samples that can reflect the characteristics of samples on the market. The presented results can contribute to the advancement of analytical techniques in this field. This could include novel preprocessing methods, innovative feature selection techniques, or advanced multivariate analysis algorithms tailored specifically for NIR data. Furthermore, NIR spectroscopy is widely used in the agricultural and food industry for various purposes including crop quality assessment, food authentication, and nutritional analysis. By applying ANN to NIR spectra of agricultural products and food samples, researchers can differentiate between different varieties, detect adulteration, and assess nutritional composition, thereby ensuring food safety and quality.