2.1 Sample preparation

Analytical-grade D-glucose (C₆H₁₂O₆, ≥ 99% purity) was used to prepare aqueous solutions at five concentrations: 0.1, 0.2, 10, 20, and 40 g/mL.

The required mass of glucose was accurately weighed, dissolved in double-distilled water, and stirred magnetically until complete dissolution. To ensure reliability and reduce the risk of microbial contamination or glucose degradation, solutions were prepared immediately prior to analysis [5,6,8].

2.2 UV-visible spectrophotometric measurements

Spectral data were acquired using a HIGHTOP UV-Visible-NIR spectrophotometer equipped with 1 cm quartz cuvettes, following procedures consistent with previous studies [7,9].

Double-distilled water was used as the blank for calibration. Absorbance spectra were recorded from 200 to 1100 nm at 1 nm resolution. Each sample was measured in triplicate, and the mean values were analyzed. All measurements were performed at ~25°C under stable laboratory conditions to ensure reproducibility.

2.3 Spectral analysis

The absorbance spectra of each glucose solution were plotted as a function of wavelength to investigate patterns in the UV and visible regions. Because glucose lacks sharp absorbance peaks within this range, emphasis was placed on minor intensity fluctuations, especially below 400 nm where variations were more pronounced [2,4].

Prior to visualization, baseline correction was applied to remove instrumental offsets, followed by Savitzky–Golay smoothing (window size = 7 points, polynomial order = 2). These preprocessing steps were implemented to improve signal quality while preserving subtle absorbance variations. Although smoothing reduces random noise, care was taken to avoid over-smoothing, which could potentially mask fine spectral features relevant to quantitative analysis.

2.4 Data processing and visualization

The raw spectral data were initially exported to Microsoft Excel for preliminary visualization. Subsequent analysis was conducted in MATLAB, where baseline correction and smoothing were applied to improve data quality. Absorbance spectra for each concentration were then visualized to highlight concentration-dependent trends across spectral regions.

2.5 Statistical and computational modeling

To explore the predictive relationship between glucose concentration and spectral absorbance, a feed-forward artificial neural network (ANN) was implemented and trained using the Levenberg–Marquardt algorithm [10]. Spectral data were normalized with the map minmax function, and divided into training (70%), validation (15%), and testing (15%) subsets according to MATLAB protocols [11]. Model performance was assessed using mean squared error (MSE) and correlation coefficient (R) values comparing predicted and actual concentrations.

To complement ANN modeling, linear regression [12] and principal component analysis (PCA) [13] were also employed. These approaches enabled extraction of key spectral features, dimensionality reduction, and cross-validation of model robustness and interpretability.

2.6 Ethical approval

All authors have read, understood, and complied with the journal’s “Ethical responsibilities of Authors” statement. The study did not involve human participants or animals, and therefore no specific ethical approval was required. The research was conducted in accordance with institutional research standards.

3.1 UV-visible spectral characteristics of glucose solutions

The UV-Vis absorbance spectra of aqueous glucose solutions at concentrations of 0.1, 0.2, 10, 20, and 40 g/mL were recorded over the 200–1100 nm range using the HIGHTOP spectrophotometer.

As expected, glucose did not show distinct absorbance peaks due to the absence of strong chromophoric groups. However, absorbance intensity increased with concentration, especially in the UV region (200–400 nm) and near-visible wavelengths, with the most pronounced changes observed below 350 nm [2,4].

At low concentrations (0.1 and 0.2 g/mL), absorbance values were low and relatively stable across the spectrum. In contrast, higher concentrations (10, 20, and 40 g/mL) exhibited noticeably greater absorbance.

This indicates that, although glucose does not undergo direct electronic transitions within this range, higher molecular density contributes to light attenuation through scattering, refractive index changes, and modifications in hydrogen bonding within the aqueous medium [2,14].

The spectral baseline remained stable across all measurements, confirming both proper instrument calibration and the consistency of the experimental setup. Fig 1 illustrates these concentration-dependent spectral differences.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. UV–Visible absorbance spectra of aqueous glucose solutions at concentrations ranging from 0.1 to 40 g/mL, measured across the 200–1020 nm range.

Absorbance intensity increased progressively with concentration, with the most pronounced differences observed in the ultraviolet region.

https://doi.org/10.1371/journal.pone.0335807.g001

As illustrated in Fig 1, the absorbance spectra of glucose solutions at concentrations of 0.1, 0.2, 10, 20, and 40 g/mL were recorded across the 200–1020 nm wavelength range.

As anticipated, glucose showed no distinct absorbance peaks in this range, reflecting the absence of strong chromophoric groups [2,4]. However,

absorbance intensity increased progressively with concentration, most notably in the ultraviolet region (200–400 nm) and extending into near-visible wavelengths.

3.2 Spectral trends and relationship between concentration and absorbance

Visual inspection of the spectral data revealed that absorbance values were generally low in the visible region (>400 nm), while measurable variations appeared in the lower UV range (200–400 nm).

This observation aligns with earlier studies showing that glucose lacks strong chromophores in this spectral range. Consequently, the absorbance patterns result mainly from indirect effects such as light scattering, refractive index changes, and hydrogen bonding, rather than distinct electronic transitions [2,4].

Fig 2 illustrates the average absorbance at selected representative wavelengths (220 nm, 260 nm, and 300 nm) as a function of glucose concentration.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Absorbance vs. Glucose Concentration at Selected Wavelengths 220 nm, 260 nm, and 300 nm.

https://doi.org/10.1371/journal.pone.0335807.g002

At these wavelengths, absorbance increased in quasi-linear or nonlinear fashion with concentration, especially in the UV region. This concentration-dependent trend highlights the potential of using absorbance data for indirect quantification of glucose concentration through statistical methods or machine learning models, despite the absence of sharp spectral features.

Fig 2. illustrates the relationship between glucose concentration and absorbance at three representative wavelengths (220, 260, and 300 nm). Absorbance values, obtained using UV–Visible spectroscopy, were recorded for glucose concentrations of 0.1, 0.2, 10, 20, and 40 g/mL. The results demonstrate a clear concentration-dependent trend, most evident at 300 nm, highlighting the potential of this wavelength for indirect glucose quantification.

3.3 Artificial neural network modeling and implications for non-destructive glucose detection

To exploit the subtle spectral variations in the UV–Visible data, an artificial neural network (ANN) model with a feed-forward architecture was developed and trained using standard MATLAB procedures [12]. The ANN demonstrated strong predictive performance, achieving correlation coefficients of R > 0.98 across training, validation, and testing datasets, along with low mean squared error (MSE) values.

This confirms the model’s ability to capture complex, non-linear relationships between absorbance patterns and glucose concentration [2,4].The regression plot (Fig 3) illustrates the close agreement between predicted and actual glucose concentrations, highlighting the ANN’s capacity to interpret subtle spectral differences that are challenging to resolve using traditional univariate methods. These findings emphasize the value of data-driven approaches, particularly machine learning, for non-destructive glucose detection even in the absence of distinct spectral peaks [15].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Regression plot comparing actual and predicted glucose concentrations obtained with the ANN model.

The feed-forward ANN achieved high accuracy, with correlation coefficients (R) > 0.98 and low mean squared error (MSE) across training, validation, and testing sets. The strong alignment of predicted and measured values confirms the ANN’s ability to capture subtle spectral differences.

https://doi.org/10.1371/journal.pone.0335807.g003

Overall, the results demonstrate the feasibility of ANN-based models as reliable analytical tools for glucose quantification in aqueous solutions [16,17].

Fig 3 Regression plot comparing predicted glucose concentrations obtained from the ANN model with experimentally measured values. The red dashed line represents the ideal fit (y = x), indicating perfect agreement. The close clustering of data points along this line confirms the high predictive accuracy of the ANN.

These findings highlight the potential of data-driven methods for non-destructive glucose analysis in aqueous systems. Despite the absence of distinct absorption peaks, UV–Vis spectra of glucose solutions contained sufficient information to enable reliable quantification.This supports the feasibility of integrating UV–Vis spectroscopy with machine learning algorithms for medical diagnostics, food quality control, and environmental monitoring, where reagent-free and real-time analysis is required.

3.4 Principal component analysis and statistical validation

Complementary statistical analyses were performed to further evaluate the data. Linear regression models at selected wavelengths (e.g., 220 and 260 nm) produced reasonable but less accurate estimates compared to the ANN [12].

Principal component analysis (PCA) reduced spectral dimensionality while preserving key variance and revealed clear clustering of spectra according to glucose concentration levels (Fig 4). These complementary methods confirmed the robustness of the data and underscored the advantage of combining machine learning with traditional statistical techniques to improve interpretability and reliability.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Principal component analysis (PCA) of UV–Vis absorbance spectra (200–1020 nm) for glucose solutions at concentrations of 0.1, 0.2, 10, 20, and 40 g/mL.

Each point represents a spectrum projected onto the first two principal components (PC1 and PC2), which together account for most of the variance in the dataset. The distinct clustering of samples by concentration confirms that PCA effectively differentiates glucose levels based on their spectral signatures.

https://doi.org/10.1371/journal.pone.0335807.g004

Nonetheless, some limitations were noted. The lack of distinct absorption bands limits direct spectroscopic quantification, while baseline drift and low signal-to-noise ratios in the near-infrared region (>800 nm) may affect model stability. Future studies should address these issues by applying signal enhancement, optimizing feature selection, and controlling parameters such as temperature and optical path length to improve model accuracy and robustness [18,19].

PCA also enabled visualization of clustering patterns among samples. Clear separation between glucose concentrations was observed in the PC1–PC2 space, demonstrating that the spectral data contained sufficient discriminatory information for classification.

Fig 4 PCA scatter plot showing samples projected onto the first two principal components. Together, PC1 and PC2 accounted for most of the variance in the dataset, confirming the robustness of the method.

Across 200–1020 nm, glucose solutions exhibited concentration-dependent absorbance patterns. At 0.1 and 0.2 g/mL, absorbance remained close to zero, indicating minimal optical activity. At higher concentrations (≥10 g/mL), absorbance increased markedly, particularly in the 200–400 nm region, reflecting enhanced molecular interactions with incident light.

At elevated concentrations, spectra exhibited more pronounced shape and intensity changes, consistent with Beer–Lambert behavior but also suggesting potential spectral shifts or saturation effects.

These trends were reinforced by PCA, which showed distinct clustering for each concentration, confirming that absorbance patterns encode sufficient discriminatory information for reliable classification.

Table 1 summarizes the performance metrics of ANN, PCA-based regression, and conventional regression models. The results indicate that PCA and linear regression achieved moderate predictive accuracy (R² ≈ 0.42), while ANN underperformed (R² ≈ 0.16), highlighting the influence of dataset size and spectral signal quality on model performance.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Performance comparison of ANN, PCA-based regression, and conventional regression models for predicting glucose concentrations from UV–Vis spectral data.

https://doi.org/10.1371/journal.pone.0335807.t001

The results in Table 1 indicate that both linear regression and PCA-based regression achieved moderate predictive accuracy, with R² values of ~0.42 and RMSE ≈ 15.1. In contrast, the ANN model exhibited poorer performance (R² ≈ 0.16, RMSE ≈ 18.3). This outcome suggests that, under the current dataset conditions (limited sample size and relatively weak spectral variations), traditional regression models are more effective than ANN in capturing the concentration–absorbance relationship. Nevertheless, the ANN framework retains potential for improved performance when applied to larger and more diverse datasets, where its ability to model nonlinear relationships may become more evident. These findings underscore the importance of selecting modeling approaches that are aligned with dataset characteristics and experimental constraints.

3.5 Implications for analytical and environmental applications

This study demonstrates the feasibility of integrating UV–Vis spectroscopy with machine learning models, particularly ANN, for reagent-free and non-destructive glucose quantification. The approach is promising for medical diagnostics (e.g., non-invasive glucose monitoring), food quality assurance, and environmental testing, where rapid and cost-effective analysis is critical.

Although the lack of distinct absorption peaks limits direct quantification, the observed spectral trends and strong model performance confirm that data-driven methods can overcome this challenge.

Future research should focus on signal enhancement, optimized feature selection, and tighter control of measurement conditions to further improve predictive robustness and analytical reliability.

3.6 Study limitations

Although the present study demonstrates the feasibility of UV–Vis spectral characterization and ANN-based modeling of aqueous sugar solutions, several limitations should be noted. First, the recorded spectra were occasionally affected by baseline drift and instrumental noise, especially at longer wavelengths where the absorbance was inherently weak. While preprocessing steps such as baseline correction and Savitzky–Golay smoothing were applied to mitigate these effects, subtle fluctuations may still influence the robustness of feature extraction. In future work, improved instrumentation and repeated calibration procedures could further enhance signal stability.

Second, the ANN models, although yielding strong predictive performance, are inherently data-driven and sensitive to training conditions. The absence of a large and diverse dataset may limit the generalizability of the models to more complex mixtures or environmental samples with multiple interfering species. Additional validation using independent datasets and comparisons with alternative machine learning algorithms would strengthen the reliability of the predictions. Despite these limitations, the study provides a valuable framework for expanding spectral-based sugar detection toward both clinical monitoring and broader analytical applications.