Machine Learning for the Photonic Evaluation of Cranial and Extracranial Sites in Healthy Individuals and in Patients with Multiple Sclerosis

Abstract

This study aims to characterize short-wave infrared (SWIR) reflectance spectra at cranial (at the scalp overlying the frontal cortex and the temporal bone window) and extracranial (biceps and triceps) sites in patients with multiple sclerosis (MS) and age-/sex-matched controls. We sought to identify the diagnostic accuracy of wavelength-specific patterns in distinguishing MS from normal controls and spectral markers associated with disability (e.g., Expanded Disability Status Scale scores). To achieve these objectives, we employed a multi-site SWIR spectroscopy acquisition protocol that included measurements from traditional cranial locations as well as extracranial reference sites. Advanced spectral analysis techniques, including wavelength-dependent absorption modeling and machine learning-based classification, were applied to differentiate MS-related hemodynamic changes from normal physiological variability. Classification models achieved perfect performance (accuracy = 1.00), and cortical site regression models showed strong predictive power (EDSS: R²_CV = 0.980; FSS: R²_CV = 0.939). Variable Importance in Projection (VIP) analysis highlighted key wavelengths as potential spectral biomarkers. This approach allowed us to explore novel biomarkers of neural and systemic impairment in MS, paving the way for potential clinical applications of SWIR spectroscopy in disease monitoring and management. In conclusion, spectral analysis revealed distinct wavelength-specific patterns collected from cranial and extracranial sites reflecting biochemical and structural differences between patients with MS and normal subjects. These differences are driven by underlying physiological changes, including myelin integrity, neuronal density, oxidative stress, and water content fluctuations in the brain or muscles. This study shows that portable spectral devices may contribute to bedside individuation and monitoring of neural diseases, offering a cost-effective alternative to repeated imaging.

1. Introduction

Multiple sclerosis (MS), a chronic inflammatory and neurodegenerative disorder of the central nervous system (CNS), is marked by demyelination, axonal loss, and heterogeneous clinical manifestations [1]. While magnetic resonance imaging (MRI) remains the gold standard for diagnosing and monitoring MS [2], its utility is constrained by high costs, limited accessibility, and an inability to capture dynamic, real-time metabolic changes. Near-infrared (NIR) spectroscopy, a non-invasive optical technique that provides insights into tissue composition, blood oxygenation, and metabolic processes, has emerged as a promising tool for probing tissue changes in neurological disorders in both cranial [3]) and extracranial sites [3,4,5]. Recent advances in NIR spectroscopy technology now enable high-resolution spectral analysis, possibly offering novel insights into systemic manifestations of MS, such as peripheral neuroinflammation, microvascular dysfunction, and secondary tissue-specific changes.

In MS, disrupted blood–brain barrier integrity, altered cerebral perfusion, and mitochondrial dysfunction contribute to disease progression [6], yet these processes remain challenging to monitor longitudinally. NIR spectroscopy-derived reflectance spectra provide wavelength-dependent information about chromophores (e.g., oxy/deoxyhemoglobin, lipids, or water) and scatterers (e.g., myelin or cellular membranes) [7], potentially serving as biomarkers for demyelination, inflammation, or metabolic stress. Extracranial measurements—such as those from skeletal muscles—may further reveal systemic correlates of CNS pathology, including endothelial dysfunction, autonomic dysregulation, and fiber hypotrophy, which are increasingly recognized as contributors to MS disability.

MS-associated central nervous system dysfunction precipitates secondary skeletal muscle oxidative distress and microvascular abnormalities through several interconnected mechanisms. CNS demyelination and axonal injury in MS disrupt neuromuscular transmission, leading to reduced neural activation of skeletal muscle. This results in muscle disuse, atrophy, and metabolic deconditioning, which in turn impairs mitochondrial function within muscle fibers, increasing the production of reactive oxygen species (ROS) and promoting oxidative stress [8,9,10]. Mitochondrial dysfunction in the CNS, a hallmark of MS, is associated with increased ROS generation and impaired energy metabolism, which can extend systemically and affect peripheral tissues, including skeletal muscle [11,12,13,14]. Experimental models demonstrate that early in the disease, skeletal muscle exhibits mitochondrial degeneration and increased oxidative stress markers, with these changes intensifying as neurogenic atrophy progresses [9]. The resulting oxidative distress impairs muscle contractility and promotes further muscle fiber damage. Additionally, chronic inactivity and altered muscle metabolism in MS patients contribute to microvascular abnormalities, such as reduced capillary density and impaired perfusion, further exacerbating muscle hypoxia and oxidative injury [8,10]. Systemic inflammation and metabolic dysfunction associated with neurodegeneration also play a role in promoting muscle oxidative stress and microvascular compromise [10]. Overall, CNS dysfunction in MS leads to impaired neuromuscular signaling, muscle disuse, mitochondrial dysfunction, and systemic inflammation, collectively driving secondary skeletal muscle oxidative distress and microvascular abnormalities [11,14].

Despite preliminary applications of NIR spectroscopy in MS for assessing motor cortex activation and fatigue [15], no study has comprehensively analyzed reflectance spectra across both cranial and extracranial sites in tandem. The present dual approach could address critical knowledge gaps, such as whether peripheral changes mirror CNS pathology or how non-neural tissue-specific spectral signatures correlate with disability.

This study aims to characterize short-wave infrared (SWIR) reflectance spectra at cranial (the scalp overlying the frontal cortex and the temporal bone window) and extracranial (upper limb biceps and triceps) sites in MS patients and age-/sex-matched controls. We sought to identify spectral markers associated with disability (e.g., Expanded Disability Status Scale (EDSS) or Fatigue Severity Scale (FSS) scores) and the diagnostic accuracy of wavelength-specific patterns in distinguishing MS from normal controls.

The findings of this study are anticipated to contribute to the growing body of research on NIR spectroscopy applications in neurology, providing new insights into the changes associated with MS. By addressing the methodological challenges of enhancing the specificity of NIR spectroscopy-derived biomarkers, our work supports the broader goal of improving non-invasive diagnostic and monitoring tools for dysfunction in MS and other neurological disorders.

2. Materials and Methods

In this study, we followed a structured, multi-step approach combining in vivo spectral data acquisition, chemometric preprocessing, multivariate statistical modeling, and model performance evaluation to differentiate patients with MS from healthy individuals. The methodological pipeline is summarized below and visually represented in Figure 1. In summary, SWIR reflectance spectra were collected using a portable spectroradiometer from both cranial (cortical and brainstem) and extracranial (biceps and triceps) sites in MS patients and matched controls. Raw spectra underwent standard preprocessing steps to reduce noise, baseline variation, and enhance relevant spectral features. Principal Component Analysis (PCA) was used to explore variance structure, visualize group separation, and detect potential outliers. Classifiers were built and validated in order to distinguish MS from healthy subjects based on spectral patterns. Multivariate regression analysis was carried out to predict clinical variables (e.g., EDSS and FSS) from spectral data. Performance metrics were adopted to assess the goodness of fit and the errors of the setup models. Finally, key wavelengths contributing most to class discrimination or prediction were analyzed and discussed.

2.1. Participants

The study included Caucasian Southern European individuals recruited from the Neurology Units at Policlinico Umberto I in Rome, Italy. The apparently healthy control group (Normal) consisted of 13 individuals (age 44 ± 16 years; 6 females, 7 males), selected from hospital personnel and relatives of patients. All patients with an established diagnosis of MS (n = 13, age 46 ± 13 years; 10 women, 3 men) underwent a detailed neurological examination performed by a senior neurologist expert in demyelinating disorders, confirming their classification as MS.

MS patients were assessed using the EDSS to quantify overall neurological disability and the FSS to evaluate perceived fatigue levels. The corresponding scores for each scale are shown in Table 1.

Written informed consent was obtained from all participants prior to their inclusion in the study, which received approval from the local ethics committee (Comitato Etico Lazio 2, protocol number 0167183/2018). All procedures adhered to applicable ethical guidelines and regulatory standards.

2.2. Statistical Analysis

To test whether the MS and control groups had similar demographics, we performed statistical tests comparing their distributions for age, sex, weight, height, and Body Mass Index (BMI). If data were normally distributed, we used unpaired Student t-tests for comparisons; otherwise, we used Mann–Whitney U tests. The Chi-Square test was performed for categorical variables (sex).

The statistical tests were performed using Python 3.8 with packages including Pandas and scipy.stats, within a Conda environment.

2.3. Spectra Collection and Analysis

The setup and procedures of the portable spectroradiometer utilization (including instrument calibration, in vivo spectral acquisition, data handling, and analytical workflow) have been previously detailed [5].

Briefly, we employed an ASD FieldSpec 4 Standard-Res spectroradiometer (ASD Inc., Boulder, CO, USA) covering a spectral range of 350–2500 nm [16].

The instrument is equipped with three separate detectors: one for the visible (VIS) spectrum (350–1000 nm), one for the first short-wave infrared (SWIR) region (SWIR1, 1001–1800 nm), and one for the second SWIR region (SWIR2, 1801–2500 nm). A contact probe with an internal halogen light source was used, delivering illumination at a 12° angle, with a 35° measurement angle and a 10 mm spot size.

Spectral calibration and acquisition were performed using the ASD’s proprietary RS3 software. The calibration procedure involved dark current correction and acquisition of a white reference standard (Spectralon from LabSphere™, North Sutton, NH, USA). The collected spectra (.asd format) were imported into MATLAB (MATLAB R2022a, ver. 9.12; The Mathworks, Inc., Natick, MA, USA) through a custom script. Data preprocessing and analysis were performed using the PLS_toolbox (ver.9.0; Eigenvector Research, Inc., Wenatchee, WA, USA) in combination with MATLAB’s Statistics and Machine Learning Toolbox (Ver 12.3.; The Mathworks, Inc., Natick, MA, USA). Spectral files were structured into dataset objects and labeled by class accordingly.

Before data collection, both the contact probe and the participants’ skin were cleaned using disposable skin wipes. Reflectance spectra in the VIS-SWIR range were then recorded from cranial and extracranial anatomical regions using the contact probe applied directly to the skin. To reduce variability in measurement conditions, all spectra were acquired in a temperature-controlled environment (22–24 °C), and participants were seated comfortably for 5–10 min before measurement to stabilize hemodynamics. A trained operator manually adjusted the length of the contact probe holder’s extensible rod and applied the contact probe using consistent and gentle pressure across all anatomical sites, based on visual confirmation of probe–skin contact and reflectance stability. The spectroradiometer was operated remotely via a connected laptop. Participants were instructed to immediately notify the operator if they experienced any discomfort during the procedure. Cranial spectral acquisition was performed at two standardized scalp locations: a cortical site (electrode position FC3 in the 10–20 EEG system) and a brainstem-associated site at the temporal acoustic window, opposite to the most affected limb. The latter location, above the zygomatic arch and anterior to the tragus, corresponds to a region commonly used in transcranial Doppler studies, benefiting from a thinner cranial bone structure. The contact probe was aligned parallel to the zygomatic arch. From each cranial site, 20 spectra were collected per subject (for a total of 40 spectra), with a total acquisition time of approximately 40 s per individual.

The extracranial sites were located on the dorsal and ventral portions of the arm, as determined by the motor point of the biceps and triceps muscles. The muscle was kept relaxed during the spectral data collection, with the segment fully supported and the limb held in a fixed position (elbow angle of 90°). In total, 50 spectra were collected from each extracranial site (100 spectra per patient), in about 100 s per subject.

Each subject contributed 50 spectra per extracranial site and 20 spectra per cranial site, resulting in 650 and 260 spectra, respectively, for each group.

Chemometric analyses were performed exclusively within the SWIR wavelength range of 1000 to 2500 nm (each spectrum contains 1501 wavelengths in the 1000–2500 nm range), excluding the visible range, for several key reasons. Primarily, the visible region (400–1000 nm) is highly susceptible to variability caused by skin pigmentation, hemoglobin concentration, and other superficial chromophores, which can introduce significant inter-subject noise and reduce the robustness and generalizability of the models.

The approach starts with a specific set of preprocessing techniques designed to eliminate physical phenomena and improve multivariate analysis [17]. The preprocessing workflow included the application of Extended Multiplicative Scatter Correction (EMSC), class-based Generalized Least Squares Weighting (GLS-W) with α = 0.002 to enhance spectral filtering, followed by mean centering (MC) algorithms.

Principal Component Analysis (PCA) was then applied to the reflectance spectra to reduce data dimensionality by extracting orthogonal components formed from linear combinations of the original variables, effectively capturing the underlying variance structure [18]. The number of principal components (PCs) retained was determined by evaluating the eigenvalues plot, while potential outliers were detected and excluded on the basis of Hotelling’s T2 versus Q residuals analysis.

A total of four separate PCAs were conducted on the primary dataset, each corresponding to a pairwise comparison of measurements obtained from one of four anatomical regions (biceps, triceps, brainstem, and cortical regions):

MS biceps (n = 650 spectra) vs. Normal biceps (n = 650 spectra);

MS triceps (n = 650 spectra) vs. Normal triceps (n = 650 spectra);

MS brainstem (n = 260 spectra) vs. Normal brainstem (n = 260 spectra);

MS cortical (n = 260 spectra) vs. Normal cortical (n = 260 spectra).

We applied Partial Least Squares Discriminant Analysis (PLS-DA) to classify and predict whether participants belonged to the MS or control group, based on reflectance spectra collected from both cranial and extracranial anatomical sites. The calibration and validation sets were generated using the Kennard-Stone (K-S) algorithm [19]. Model cross-validation and the selection of the optimal number of latent variables (LVs) were carried out using the Venetian Blinds (VBs) method. To evaluate the performance of the classification models, metrics such as sensitivity, specificity, error rate, precision, and overall accuracy were calculated from the confusion matrix [20].

Variable Importance in Projection (VIP) scores were computed for each classification model to assess the relative contributions of individual wavelengths to class separation. VIP scores provide a way to evaluate the importance of each variable within the model’s projection [21]. Variables with VIP values around or above 1 were considered influential, while those exceeding a threshold of 2 were highlighted as particularly important. VIP score trends were compared across all models, with special attention to peaks above 2, which were interpreted as key indicators of discriminative spectral features.

We developed eight PLS regression (PLS-R) models to relate NIR reflectance spectra —acquired at four anatomical sites (biceps, triceps, brainstem, and cortical regions)—to clinical disability (EDSS) and fatigue (FSS) scores in patients. The mean EDSS score was 3 ± 2, and the mean FSS score was 30 ± 15. PLS is a multivariate chemometric technique ideally suited to cases where a large, collinear set of predictor variables (spectral wavelengths) must be used to predict one or more dependent variables. Model calibration and cross-validation were performed by employing the VBs method to determine the optimal number of LVs. Predictive performance was evaluated by the root mean square error (RMSE) for calibration and cross-validation, by the coefficient of determination (R²), and by assessment of bias in both calibration and cross-validation phases to ensure minimal systematic error. Additionally, to assess the effect size of the PLS-R models, we calculated Cohen’s f² [22], a widely accepted metric for quantifying effect size in multiple regression frameworks where both the dependent and independent variables are continuous. Cohen’s f² was computed for each regression model (EDSS and FSS) to evaluate the practical significance of the prediction. This index, calculated on R²cv in our case, allows for the evaluation of the relationship magnitude, with conventional thresholds indicating small (f² = 0.02), medium (f² = 0.15), and large (f² = 0.35) effects.

Finally, VIP scores were examined to highlight the specific wavelengths that contributed most strongly to the prediction of EDSS and FSS scores.

PLS-DA and PLS-R were selected due to their robustness in handling collinear, high-dimensional spectral data with small sample sizes, as well as their interpretability via latent variables and VIP scores.

3. Results

3.1. Statistical and Exploratory Analyses

A Shapiro–Wilk normality test showed that weight (MS: p = 0.251, control: p = 0.220), height (MS: p = 0.366, control: p = 0.463, and BMI (MS: p = 0.234, control: p = 0.414) were normally distributed and did not differ between groups. Age was not normally distributed (MS: p = 0.017, control: p = 0.077), but was similar between groups (Mann–Whitney U Test U = 89.0, p = 0.837). A Chi-Square test showed that sex had a similar distribution between groups (χ² = 1.463, p = 0.227), thereby confirming the validity of the sex- and age-matching procedure.

The spectral data were collected from both extracranial and cranial sites without causing any discomfort to the participants. A total of 140 spectral acquisitions were obtained from every subject, consisting of 100 recordings from extracranial sites and 40 from cranial sites. This resulted in a comprehensive dataset comprising 3640 spectra, with 2600 spectra obtained from extracranial sites and 1040 from cranial sites.

Figure 2 shows the mean reflectance spectra recorded from the analyzed sites, while pre-processed spectra are reported in Supplementary Materials (Figure S1). A comparison between the mean and standard deviation of the raw reflectance and pre-processed (EMSC + GLS-W + MC) spectra for each analyzed anatomical site is reported in Figure S2 (Supplementary Materials). A visual inspection of the average spectra revealed differences between groups and measurement locations. For cranial sites, at the cortical site the most evident spectral features are in the range 1050–1100 nm, around 1500 nm, and in the ranges 1900–2000 nm and 2200–2400 nm; at the brainstem site, in the range 1100–1300 nm, around 1800 nm, and between 2200 and 2300 nm. For the extracranial sites corresponding to the biceps and triceps, the most prominent spectral variations were observed at specific wavelengths, particularly around 1050 nm, within the 1300–1400 nm range, in the range 1600–1800 nm, around 1900 nm, and around 2200–2300 nm.

PCA performed on the dataset associated with the examined site pairs demonstrated that spectral scores were grouped according to specific class pairs: “MS cortical/Normal cortical”, “MS brainstem/Normal brainstem”, “MS biceps/Normal biceps”, and “MS triceps/Normal triceps”. Across all cases, PC1 accounted for the highest variance among the analyzed classes. The PCA score plots indicated that the explained variance (EV) for PC1 varied across the different anatomical site pairs, with values of 38.85% for the biceps (Figure 3a), 47.96% for the triceps (Figure 4a), 25.93% for the cortical site (Figure 5a), and 33.33% for the brainstem site (Figure 6a).

3.2. PLS-DA Classification Models

To further investigate predictive relationships between spectral inputs and class labels, PLS-DA was carried out. This approach enabled an efficient classification of spectral data derived from both cranial and extracranial sites. The K-S algorithm was used to select representative samples that captured the variability within the dataset and to define the calibration subset for model development. The dataset was partitioned into training (70%) and testing (30%) subsets for model calibration (training) and validation (test), respectively (as reported in Table 2). By adopting this approach, the risk of possible overfitting was mitigated.

All four classification models showed outstanding classification performance. As reported in Table 2, each model achieved perfect scores across all key performance indicators (including sensitivity, specificity, precision, and accuracy) during calibration (C), cross-validation (CV), and external prediction (P), with an error rate of zero in every case.

Although the models achieved perfect classification performance, the lack of an external validation dataset remains a limitation, and future studies will aim to assess generalizability using independent test sets.

Variables with a VIP score exceeding 2 played an important role in distinguishing tissue types. The VIP score plots for the “MS Biceps/Normal Biceps” and “MS Triceps/Normal Triceps” comparisons are shown in Figure 7, while those for “MS Cortical/Normal Cortical” and “MS Brainstem/Normal Brainstem” are shown in Figure 8.

Table 3 shows the key VIP score peaks derived from the PLS-DA models.

At the cortical site, relevant peaks occur at 1050 nm (indicative of second overtone C–H vibrations linked to lipids and proteins [23,24]), 1500 nm (associated with a combination band of C–H and N–H stretching, as well as O–H overtones, commonly found within protein and water content in tissue [23,24]), 1900–2000 nm (related to water and, partially, amide group absorptions [3,23,24]), and 2200–2250 nm (corresponding to O–H, N–H, and C–H combination bands, vibrations, which are linked to proteins, amide groups, and lipids, respectively [3,23,25]).

For the brainstem site, key spectral regions include peaks around 1300 nm (that corresponds to the first overtone of O–H stretching and is sensitive to water and hydroxyl-rich compounds [24]), 1800 nm (that captures water and some C–H stretching modes, possibly indicating biochemical variation in hydration and lipid content [3,24]), 2200 nm, and 2300 nm (representing combination bands involving C–H, N–H, and O–H groups, corresponding to complex molecular vibrations in proteins, lipids, and residual water [3,23,25]), which align with molecular absorptions characteristic of biochemical changes in this brain region.

These spectral features provide critical insight into the biochemical variations between different tissue types, reinforcing the efficacy of the PLS-DA classification models.

For the biceps site, the PLS-DA model highlights important spectral peaks at 1900 nm (strongly related to water absorption [24]), 2200–2290 nm (associated with N–H and O–H combination bands), and 2400 nm, which corresponds to C–H overtone and combination bands [23].

For the triceps site, significant peaks appear at 1000 nm (linked to second overtone C–H vibrations [23]), 1800 nm (first overtone of C–H stretching, related to lipid content [23]), 1950 nm (associated with water combination bands [23]), 2200 nm, and 2390 nm (covering combination bands of C–H, N–H, and O–H [3,23]), representing key molecular absorptions relevant to the classification.

3.3. EDSS-Based PLS Regression Models

PLS-R models were developed to predict EDSS scores based on NIR spectra acquired from the four anatomical sites (Figure 9). All models demonstrated strong performance, with minimal bias and a high coefficient of determination in both the calibration and cross-validation phases (Table 4). The cortical model achieved the best predictive accuracy (RMSECV = 0.293, R²_CV = 0.980) despite requiring seven latent variables, indicating the high informational content of the cortical spectra. The biceps model followed closely, with RMSECV = 0.399 and R²_CV = 0.962, using only two LVs. The triceps model also showed good performance (RMSECV = 0.475, R²_CV = 0.946), while the brainstem model, although requiring more LVs (five), maintained acceptable accuracy (RMSECV = 0.597, R²_CV = 0.916).

Notably, each model’s RMSECV remained below the EDSS standard deviation (2.0), indicating that prediction errors fell within the intrinsic variability of the clinical scores. The calculated f² values for each PLS-R model confirmed that the models exhibited large effect sizes, supporting their robustness despite the limited sample size.

VIP analysis revealed key wavelengths contributing to the prediction, with consistent peaks at 1100 nm and 1700 nm across all sites, suggesting these regions may serve as potential spectral biomarkers of disease severity. Site-specific features (e.g., 1050 nm in cortical, 1300 nm in brainstem) highlight regional spectral differences that may reflect localized physiological or pathological characteristics.

3.4. FSS-Based PLS Regression Models

PLS-R models were developed to predict FSS scores from NIR spectra obtained at the four anatomical sites (Figure 10). Overall, the models displayed moderate to high predictive power, with performance varying across anatomical sites (Table 5). The cortical model yielded the best results, showing the lowest RMSECV (3.504) and the highest R²_CV (0.939), indicating strong predictive ability despite the use of 7 LVs. Similarly, the brainstem model performed well with RMSECV = 4.675 and R²_CV = 0.891, requiring only three LVs, which suggests a good balance between model complexity and performance. The biceps and triceps models showed lower predictive accuracy, with RMSECVs of 7.295 and 8.779, respectively, and R²_CV values of 0.735 and 0.617, indicating limited ability to explain variability in FSS scores from spectra at these sites. The low bias across all models suggests minimal systematic error in predictions. Because all RMSECV values are substantially lower than the FSS standard deviation (15), these models capture meaningful spectral to clinical correlations despite differences in latent variable complexity. Moreover, the strong predictive power of the NIR spectral data for estimating FSS scores was confirmed by the Cohen’s f² values, particularly in the brainstem and cortical sites.

VIP score analysis revealed key wavelength regions relevant to FSS prediction. Notably, 1200 nm was consistently important across biceps, triceps, and cortical sites, while 1050 nm appeared across triceps, brainstem, and cortical regions. These wavelengths likely correspond to specific tissue components or metabolic features associated with fatigue in MS. The broader and more informative spectral features observed in the cortical and brainstem regions support their value for assessing fatigue severity using optical spectroscopy.

4. Discussion

This study demonstrates that non-invasive in vivo NIR reflectance spectroscopy in humans can effectively differentiate the clinical condition of individuals, thereby enabling the identification of MS patients.

This capability is attributed to the unique optical characteristics of the tissue volumes examined at both extracranial and cranial locations.

4.1. Cranial Sites

The NIR reflectance spectra acquired from cerebral cortical regions reveal distinct wavelength-specific differences between patients with MS and healthy controls, offering novel insights into the biochemical and microstructural alterations underpinning MS pathology. These spectral variations, observed at 1050, 1500, 1900–2000, and 2200–2250 nm, align with known disease mechanisms and provide a non-invasive means to probe both focal and diffuse tissue damage.

Between-group spectral differences emerge prominently at 1050 nm. This wavelength is associated with aromatic (ArCH) or primary amine (RNH₂) vibrations and reflects the distribution of proteins and lipids, essential components for the structural integrity of neural tissue. The finding likely reflects demyelination-driven lipid depletion, particularly the loss of myelin-specific galactocerebrosides and sulfatides [26], that exhibit strong absorption due to second overtone C–H stretching vibrations in their methyl (-CH₃) and methylene (-CH₂-) groups. This finding corroborates histopathological studies demonstrating a 20–30% reduction in myelin lipids within chronic active lesions [27]. Concurrently, elevated scattering coefficients at this wavelength may arise from neurofilament accumulation [28], a hallmark of axonal injury in MS. These observations align with prior work linking neurofilament light chain levels in cerebrospinal fluid to disease activity [29], suggesting NIR spectroscopy could serve as a surrogate for tracking axonal degeneration.

Spectral differences found at 1300 nm collected from the cranial brainstem site reflect lipid loss but are mediated by distinct molecular mechanisms and spectral contributions due to a more complex interplay of lipid loss and water accumulation. Third overtone C–H vibrations in long-chain fatty acids (e.g., C24:0 sulfatides) dominate this region, but the signal is confounded by overlapping O–H stretching vibrations from water molecules. In MS, lipid depletion reduces absorption, while concurrent vasogenic edema, driven by blood–brain barrier disruption, increases water content [30], resulting in a flattened or shifted spectral profile. This duality complicates interpretation but provides critical complementary information; the 1300 nm band captures both structural demyelination and dynamic inflammatory processes.

The combined analysis of these wavelengths may enhance diagnostic specificity. For instance, a reduced 1050/1300 nm absorption ratio could potentially distinguish chronic demyelination (predominantly lipid loss) from acute inflammation (predominant edema). This differentiation would have clinical relevance, possibly aiding in the identification of lesion subtypes that might require distinct therapeutic approaches.

Absorption at 1500 nm (the first overtone of NH, OH, and CH vibrations) corresponds to elevated water content that may be secondary to vasogenic edema from blood–brain barrier disruption or neuroinflammatory infiltrates, consistent with MRI studies showing transient T2 hyperintensities in acute lesions. However, no relapsing patient was recruited in the study; therefore, other factors should be invoked. The contribution of collagen deposition to scattering at this wavelength implicates reactive astrogliosis, a pathological feature of MS that promotes fibrotic scar formation [31]. The dual role of water and collagen in this spectral band underscores the utility of NIR spectroscopy in capturing dynamic tissue changes that conventional imaging may overlook.

At 1800 nm (a wavelength identified from the brainstem site), the first overtone of SH and CH vibrations indicates possible alterations in oxidation processes and the metabolism of sulfur-containing proteins, which are critical for neuronal functionality. Hydrogen sulfide-induced S-sulfhydration regulates blood–brain barrier permeability and cerebral blood flow, and it intervenes in biological pathways, such as inflammatory response and oxidative stress [32]; in addition, protein S-sulfhydration regulates mitochondrial integrity, long-term potentiation, and calcium homeostasis [33], mechanisms found altered in MS patients.

The spectral features found in the 1900–2000 nm range align with elevated lactate levels, a metabolite indicative of anaerobic glycolysis in metabolically stressed cells. In MS, this likely reflects activated microglia and macrophages within inflamed lesions, which shift to glycolytic metabolism under hypoxic conditions [34]. The simultaneous detection of hydroxyl group vibrations suggests concomitant oxidative stress, a mechanism implicated in both mitochondrial dysfunction and neurodegeneration [35]. These metabolic perturbations mirror magnetic resonance spectroscopy findings of lactate peaks in chronic active lesions [36], positioning NIR spectroscopy as a complementary tool for monitoring inflammatory activity.

Changes in the spectral features at 2200–2300 nm may correspond to the loss of long-chain fatty acids in myelin membranes [37], a hallmark of MS demyelination. At 2200 nm, NH+OH combinations suggest changes in hydration levels and protein balance, while the signal at 2300 nm, linked to CH+CH combinations, highlights changes in lipid composition and membrane stability. Notably, the spectral shifts in this region may also arise from β-sheet-rich protein aggregates, such as amyloid-like deposits observed in progressive MS [38]. Finally, the 2400 nm signal, attributed to CH+CH and CH+CC vibrational combinations, suggests a reorganization of lipid components and modifications in cellular membrane stability.

Overall, the multispectral approach employed here enables differentiation of MS-specific changes from age- or comorbidity-related confounding factors. These differences are driven by underlying physiological changes, including lipid and myelin integrity, neuronal density, oxidative stress, and water content fluctuations.

4.2. Extracranial Sites

The human arm is a heterogeneous structure composed of layers with distinct optical properties. It is covered by skin and appendages, has a subcutaneous adipose layer between the skin and muscles, and contains an underlying bone, the humerus, all of which are variably vascularized. Due to its low absorption, NIR light can penetrate several centimeters into biological tissues, allowing for non-invasive extraction of information from deeper structures. Previous studies have already demonstrated that NIR reflectance spectra acquired from the ventral and dorsal aspect of the arm mainly reflect the current biology of the biceps and triceps, and that the ability to distinguish between muscles is based on features like muscle architecture, metabolic enzyme content, and contractile protein expression [3,4]. Spectral differences observed in biceps and triceps between patients with MS and healthy controls highlight significant biochemical and structural changes linked to the disease. These alterations at specific wavelengths (i.e., 1000 nm, 1800–1950 nm, and 2200–2300 nm) shed light on systemic manifestations of MS extending beyond CNS pathology.

Spectral features at 1000 nm reflect changes in lipid and protein composition within muscle tissue. This wavelength corresponds to second overtone C–H stretching vibrations in fatty acids and methyl/methylene groups in various lipids. In MS, muscle atrophy and inflammatory stress alter lipid metabolism and protein density [39]. Muscle atrophy arises from decreased mobility [40], leading to reduced muscle mass and altered storage of lipids, while systemic inflammation promotes lipolysis and protein breakdown. These changes contribute to the observed spectral shifts at 1000 nm, providing a window into the systemic effects of MS on muscle tissue.

In 1800–1950 nm, the absorption is primarily due to elevated water content and collagen remodeling. The first overtone of O–H stretching vibrations in this range is sensitive to changes in muscle edema, which is enhanced by systemic inflammatory cytokines in MS [41]. Concurrently, chronic inflammation leads to increased collagen deposition, affecting the scattering properties of muscle tissue. This dual effect (edema and fibrosis) underscores how systemic inflammation in MS impacts muscle integrity, contributing to fatigue and weakness [42].

The spectral changes at 2200–2300 nm highlight alterations in lipid peroxidation and oxidative stress within muscle tissue. Lipid peroxidation, a hallmark of oxidative damage, leads to the formation of lipid peroxides absorbing in this range. Moreover, the loss or alteration of long-chain fatty acids due to oxidative stress affects the absorption peak. MS-related mitochondrial dysfunction in muscle cells [43] enhances reactive oxygen species (ROS) production, exacerbating lipid peroxidation. This oxidative stress is central to muscle fatigue and weakness in MS patients [44], illustrating how NIR spectroscopy can capture systemic metabolic dysregulation.

Overall, these findings suggest that NIR spectroscopy-derived biomarkers could complement CNS-centered diagnostic tools by offering insights into systemic disease activity. Muscle weakness in MS is a result of the central neural changes from the primary disease process [45], but over time, patients experience accelerated muscle loss and dysfunction because of the interaction between weakness, fatigue, decreased mobility of paretic limbs [46], and age-related decline. An NIR spectroscopy approach might facilitate the monitoring of systemic MS effects and therapeutic responses beyond traditional CNS measures.

4.3. Correlation with Clinical Data

The spectral changes collected from cranial and extracranial sites in patients strongly correlate with clinical measures of disease severity. PLS models displayed high predictive power for both EDSS and FSS scores, with performance varying across anatomical sites. The finding that spectra collected at both the cranial and extracranial sites are predictive of disability and fatigue reinforces the notion that both central neural and peripheral muscular dysfunction contributes to clinical disability and to the most debilitating symptom of MS, fatigue.

4.4. Comparison with Functional NIRS and Multimodal Perspectives

To contextualize the complementary roles of optical neuroimaging in multiple sclerosis, it is important to contrast functional near-infrared spectroscopy (fNIRS) with broadband NIR reflectance spectroscopy. The primary distinction lies in the nature of the signals measured. fNIRS in MS assesses task-evoked changes in oxygenated and deoxygenated hemoglobin within cortical regions, reflecting neurovascular and hemodynamic responses during functional or cognitive tasks [47,48]. Such studies reveal altered cortical activation, reduced neurovascular coupling, and impaired compensatory mechanisms, linking abnormal prefrontal activation to cognitive fatigue [15]. In contrast, broadband NIR reflectance spectroscopy captures a richer spectral fingerprint across visible and shortwave infrared bands, probing biochemical and microstructural properties at cranial and extracranial sites [3]. By analyzing absorption features linked to myelin integrity, oxidative stress, water, and protein/lipid composition, it identifies molecular alterations undetectable by hemodynamic changes alone. Unlike fNIRS, it also assesses extracranial muscle, informing systemic MS manifestations. Thus, fNIRS excels in functional monitoring, while broadband NIR spectroscopy reveals static biochemical and structural abnormalities, even at rest, detecting subclinical injury. Machine learning applied to multisite spectra shows high accuracy in differentiating MS from controls, supporting diagnostic and monitoring potential. Recent studies advocate combining functional and biochemical optical biomarkers for enhanced MS phenotyping and personalized care [49,50].

4.5. Limitations and Future Directions

Despite these promising findings, current studies are limited by small sample sizes and heterogeneity in MS subtypes. Future research should focus on refining spectral unmixing algorithms, such as multivariate curve resolution, to enhance specificity for myelin-derived signals. Additionally, while discarding the <1000 nm wavelength range reduces melanin-related variability, and recruiting participants who are regionally homogenous implies similar skin types, we recognize that the absence of quantitative skin tone characterization (e.g., using the Fitzpatrick scale or objective colorimetric measures) may represent a limitation.

Other limitations are the cross-sectional design, the lack of longitudinal reproducibility data, and the potential influence of disease-modifying therapies (DMTs) on the metabolic status of participants. The cross-sectional nature inherently limits the ability to assess temporal stability and causality of the observed near-infrared spectral changes. Without longitudinal follow-up, we cannot determine whether these spectral biomarkers track disease progression or therapeutic response over time. Furthermore, by altering the inflammatory activity and neurodegeneration, the heterogeneity of DMT regimens may represent a potential confounder that we were unable to control statistically given the limited sample size.

Although the spectral analysis identified wavelength-specific patterns potentially related to tissue composition and physiological status, the present study relied on in vivo spectra without direct biochemical validation. Spectral assignments were based on the established NIR literature, yet no ex vivo tissue phantoms or reference liquids (e.g., lipid emulsions or water) were employed to confirm molecular attributions. Incorporating such validation in future work will be essential to strengthen the biochemical interpretation of key spectral features identified by VIP analysis. A final limitation of the present approach is that spectra were treated as independent observations, despite multiple spectra being acquired from each subject. This approach is common in NIR chemometrics but may introduce pseudo-replication, potentially inflating statistical power. Nonetheless, future studies should explore the use of mixed-effects models that treat subject identity as a random factor, which would allow for a more accurate generalization of findings at the participant level.

5. Conclusions

The observed differences in SWIR reflectance spectra between MS patients and healthy controls are driven by pathophysiological changes in brain and muscle tissues. These differences primarily involve changes in lipid content, water absorption, and tissue scattering properties, all of which reflect central neural (neurovascular dysfunction, demyelination, and inflammation), and peripheral muscular (hypotrophy, fibrosis, and fatigue) mechanisms.

By integrating multisite NIR spectroscopy data with clinical parameters, this work seeks to establish a framework for using portable optical spectroscopy as a complementary tool in MS management, with potential applications in therapeutic monitoring by enabling phenotype stratification and personalized therapeutic interventions. Rigorous analytical and clinical validation, including standardized protocols for device calibration, data acquisition, and processing, is essential to ensure accuracy, sensitivity, specificity, and reproducibility across diverse populations. Early engagement with regulatory agencies would facilitate qualification and define the biomarker’s context of use. Integration with routine MRI could enhance diagnostic confidence, enable more frequent bedside monitoring, and guide personalized treatment adjustments. Furthermore, its potential cost-effectiveness would warrant formal economic modeling to assess healthcare impact and optimize resource utilization.

Author Contributions

Conceptualization, A.C. and G.B.; methodology, A.C., C.M., G.B., S.S., R.G., L.M. and D.G.; software, R.G. and D.G.; validation, A.C., A.M., C.M., G.B., S.S., P.M., R.G., F.F., L.M. and L.P.; formal analysis, A.C., C.M., G.B., R.G. and D.G.; investigation, A.C., C.M., G.B., R.G., D.G. and C.T.; resources, A.C. and G.B.; data curation, A.C., C.M., G.B., R.G. and D.G.; writing—original draft preparation, A.C., C.M., G.B., R.G., C.T., L.M. and L.P.; writing—review and editing, A.C., G.B., R.G., D.G., L.M., N.L.B. and L.P.; visualization, A.C., A.M., C.M., F.F., G.B., S.S., P.M., R.G., N.L.B. and D.G.; supervision, A.C., C.T. and G.B. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the local ethics committee (Comitato Etico Lazio 2, protocol number 0167183/2018).

Informed Consent Statement

Informed consent was obtained from all subjects involved in the study.

Data Availability Statement

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

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1 Workflow of the adopted methodology.

Figure 2 Mean reflectance spectra acquired from extra-cranial and cranial anatomical locations: (a) MS biceps vs. Normal biceps, (b) MS triceps vs. Normal triceps, (c) MS cortical vs. Normal cortical, and (d) MS brainstem vs. Normal brainstem.

Figure 3 Principal Component Analysis (PCA) score plots of the first two principal components derived from spectra obtained at the extra-cranial biceps site in both multiple sclerosis (MS) patients and healthy controls (Normal) are presented in (a). The corresponding loading plot for the first principal component (PC1) is shown in (b).

Figure 4 Principal Component Analysis (PCA) score plots of the first two principal components (a) derived from spectra collected at the extra-cranial triceps site in both patients (MS) and healthy subjects (Normal). Panel (b) presents the loadings plot for the first principal component.

Figure 5 Principal Component Analysis (PCA) score plots of the first two principal components (a) based on spectra collected from the cranial/cortical site in both patients (MS) and healthy subjects (Normal). Panel (b) shows the loadings plot for the first principal component.

Figure 6 Principal Component Analysis (PCA) score plots of the first two principal components (a) based on spectra collected from the cranial/brainstem site in both patients (MS) and healthy subjects (Normal). Panel (b) presents the loadings plot for the first principal component.

Figure 7 VIP score plots for “MS biceps/Normal biceps” (a) and “MS triceps/Normal triceps” (b).

Figure 8 VIP score plots for “MS cortical/Normal cortical” (a) and “MS brainstem/Normal brainstem” (b).

Figure 9 EDSS-based PLS regression model results and corresponding VIP scores plot of biceps (a), triceps (b), brainstem (c), and cortical (d) sites.

Figure 10 FSS-based PLS regression model results and corresponding VIP score plots of biceps (a), triceps (b), brainstem (c), and cortical (d) sites.

Table 1

Demographic and anthropometric characteristics of the study participants, including Expanded Disability Status Scale (EDSS) and Fatigue Severity Scale (FSS) scores for subjects with multiple sclerosis (MS). Control subjects (Normal) did not undergo EDSS or FSS evaluation (indicated by “-”). Body Mass Index: BMI; Disease Duration: DD; disease-modifying therapy: DMT; Annualized Relapse Rate: ARR. Sex is coded as 0 = male, 1 = female.

ID	Condition	Age [yrs]	Sex	Weight [kg]	Height [cm]	BMI	DD [mths]	EDSS	FSS	DMT	ARR
D01	RRMS	36	1	75	156	31	8	1	18	Peginterferon beta-1a	0.2
D02	RRMS	41	0	104	178	33	6	1	11	Dimethyl fumarate	0.4
D03	RRMS	24	1	80	167	29	4	1	33	Ocrelizumab	0.5
D04	RRMS	51	1	51	170	18	24	2.5	20	Peginterferon beta-1a	0.3
D05	RRMS	19	1	55	168	19	0.09	2	21	Dimethyl fumarate	1.0
D06	RRMS	58	1	48	160	19	14	3	43	Fingolimod	0.2
D07	RRMS	42	1	48	162	18	4	1.5	47	Peginterferon beta-1a	0.4
D08	RRMS	55	1	70	170	24	2	2.5	28	Dimethyl fumarate	0.6
D09	RRMS	51	0	85	182	26	20	6	28	Ocrelizumab	0.2
D10	RRMS	54	1	95	157	39	7	5	61	Dimethyl fumarate	0.5
D11	RRMS	56	1	62	167	22	17	6	37	Fingolimod	0.2
D12	RRMS	57	0	85	181	26	20	3.5	32	Cladribine	0.2
D13	SPMS	54	1	47	167	17	24	7.5	9	Ofatumumab	0.3
NT03	Normal	55	1	58	168	21	-	-	-	-	-
NT04	Normal	27	1	54	160	21	-	-	-	-	-
NT05	Normal	24	1	58	167	21	-	-	-	-	-
NT07	Normal	36	0	75	174	25	-	-	-	-	-
NT10	Normal	66	0	74	168	26	-	-	-	-	-
NT11	Normal	53	1	75	156	31	-	-	-	-	-
NT12	Normal	50	0	68	170	24	-	-	-	-	-
NT13	Normal	58	0	83	173	28	-	-	-	-	-
NT14	Normal	26	0	72	178	23	-	-	-	-	-
NT15	Normal	27	0	80	177	26	-	-	-	-	-
NT16	Normal	33	0	57	160	22	-	-	-	-	-
NT17	Normal	64	1	62	159	25	-	-	-	-	-
NT18	Normal	54	1	76	168	27	-	-	-	-	-

Table 2

Statistical metrics derived from PLS-DA models for extracranial sites (biceps and triceps) and cranial sites (cortex and brainstem). Model evaluation was performed during the calibration (C), cross-validation (CV), and prediction (P) phases.

Model	Model phase	Class	Sensitivity	Specificity	Number of Spectra	Error Rate	Precision	Accuracy
MS biceps/ Normal biceps	C	MS biceps	1	1	458	0	1	1
		Normal biceps	1	1	452	0	1	1
	CV	MS biceps	1	1	458	0	1	1
		Normal biceps	1	1	452	0	1	1
	P	MS biceps	1	1	192	0	1	1
		Normal biceps	1	1	198	0	1	1
MS triceps/ Normal triceps	C	MS Triceps	1	1	452	0	1	1
		Normal triceps	1	1	458	0	1	1
	CV	MS Triceps	1	1	452	0	1	1
		Normal triceps	1	1	458	0	1	1
	P	MS Triceps	1	1	198	0	1	1
		Normal triceps	1	1	192	0	1	1
MS cortical/ Normal cortical	C	MS cortical	1	1	182	0	1	1
		Normal cortical	1	1	182	0	1	1
	CV	MS cortical	1	1	182	0	1	1
		Normal cortical	1	1	182	0	1	1
	P	MS cortical	1	1	78	0	1	1
		Normal cortical	1	1	78	0	1	1
MS brainstem/ Normal brainstem	C	MS brainstem	1	1	186	0	1	1
		Normal brainstem	1	1	178	0	1	1
	CV	MS brainstem	1	1	186	0	1	1
		Normal brainstem	1	1	178	0	1	1
	P	MS brainstem	1	1	74	0	1	1
		Normal brainstem	1	1	82	0	1	1

Table 3

Wavelengths identified based on VIP scores from the PLS-DA models for both cranial and extracranial sites.

Wavelengths (nm)	Cortical	Brainstem	Biceps	Triceps
1000–1100	1050			1000
1100–1200
1200–1300		1300
1300–1400
1400–1500	1500
1500–1600
1600–1700
1700–1800
1800–1900		1800		1800
1900–2000	1900–2000		1900	1950
2000–2100
2100–2200
2200–2300	2200–2250	2200	2200–2290	2200
2300–2400		2300		2390
2400–2500			2400

Table 4

Performance parameters for the EDSS-based PLS regression models of the spectra collected on biceps, triceps, brainstem, and cortical sites.

PLS Model	Biceps EDSS	Triceps EDSS	Brainstem EDSS	Cortical EDSS
LVs	2	2	5	7
RMSEC	0.34791	0.474	0.58775	0.28884
RMSECV	0.39863	0.475	0.59679	0.29334
Bias C	0	0	0	0
Bias CV	0.00015	0.000123	0.0016795	0.0007989
R2C	0.971	0.947	0.918	0.98
R2CV	0.962	0.946	0.916	0.98
Cohen’s f2	25.316	17.519	10.905	49.000

LVs: latent variables; RMSEC: root mean square error in calibration; RMSECV: root mean square error in cross-validation; Bias C: calibration bias; Bias CV: cross-validation bias; R²_C: coefficient of determination in calibration; R²_CV: coefficient of determination in cross-validation.

Table 5

Performance parameters for the FSS-based PLS regression models of the spectra collected on biceps, triceps, brainstem, and cortical sites.

PLS Model	Biceps FSS	Triceps FSS	Brainstem FSS	Cortical FSS
LVs	6	3	3	7
RMSEC	7.233	8.7686	4.6228	3.4493
RMSECV	7.295	8.7793	4.6745	3.5037
Bias C	0	0	0	0
Bias CV	−0.006235	−0.000297	0.033647	−0.000959
R2C	0.737	0.618	0.894	0.941
R2CV	0.735	0.617	0.891	0.939
Cohen’s f2	2.774	1.611	8.174	15.393

LVs: latent variable; RMSEC: root mean square error in calibration; RMSECV: root mean square error in cross-validation; Bias C: calibration bias; Bias CV: cross-validation bias; R²_C: coefficient of determination in calibration; R²_CV: coefficient of determination in cross-validation.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/app15158534/s1, Figure S1: EMSC + GLS-W + MC pre-processed spectra: (a) MS biceps vs. Normal biceps, (b) MS triceps vs. Normal triceps, (c) MS cortical vs. Normal cortical, and (d) MS brainstem vs. Normal brainstem. Figure S2: Comparison between mean and standard deviation of raw reflectance and pre-processed (EMSC + GLS-W + MC) spectra for: Normal biceps (a), MS biceps (b), Normal triceps (c), MS triceps (d), Normal cortical (e), MS cortical (f), Normal brainstem (g), and MS brainstem (h).

References

1. Zainab, S.R.; Khan, J.Z.; Tipu, M.K.; Jahan, F.; Irshad, N. A review on multiple sclerosis: Unravelling the complexities of pathogenesis, progression, mechanisms and therapeutic innovations. Neuroscience; 2025; 567, pp. 133-149. [DOI: https://dx.doi.org/10.1016/j.neuroscience.2024.12.029]

2. Giovannoni, G.; Hetherington, S.; Jones, E.; Castro, P.D.; Karu, H.; Ansari, S.; Karlsson, G.; Heras, V.d.L.; Lines, C. MRI versus relapse: Optimal activity monitoring for management of progressive multiple sclerosis. Brain Commun.; 2025; 7, fcaf010. [DOI: https://dx.doi.org/10.1093/braincomms/fcaf010]

3. Currà, A.; Gasbarrone, R.; Gattabria, D.; Bonifazi, G.; Serranti, S.; Greco, D.; Missori, P.; Fattapposta, F.; Picciano, A.; Maffucci, A. . In Vivo Insights: Near-Infrared Photon Sampling of Reflectance Spectra from Cranial and Extracranial Sites in Healthy Individuals and Patients with Essential Tremor. Photonics; 2024; 11, 1025. [DOI: https://dx.doi.org/10.3390/photonics11111025]

4. Currà, A.; Gasbarrone, R.; Cardillo, A.; Fattapposta, F.; Missori, P.; Marinelli, L.; Bonifazi, G.; Serranti, S.; Trompetto, C. In vivo non-invasive near-infrared spectroscopy distinguishes normal, post-stroke, and botulinum toxin treated human muscles. Sci. Rep.; 2021; 11, 17631. [DOI: https://dx.doi.org/10.1038/s41598-021-96547-6] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34480037]

5. Currà, A.; Gasbarrone, R.; Cardillo, A.; Trompetto, C.; Fattapposta, F.; Pierelli, F.; Missori, P.; Bonifazi, G.; Serranti, S. Near-infrared spectroscopy as a tool for in vivo analysis of human muscles. Sci. Rep.; 2019; 9, 8623. [DOI: https://dx.doi.org/10.1038/s41598-019-44896-8] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31197189]

6. Orian, J.M. A New Perspective on Mechanisms of Neurodegeneration in Experimental Autoimmune Encephalomyelitis and Multiple Sclerosis: The Early and Critical Role of Platelets in Neuro/Axonal Loss. J. Neuroimmune Pharmacol.; 2025; 20, 14. [DOI: https://dx.doi.org/10.1007/s11481-025-10182-w] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39904925]

7. Sakudo, A. Near-infrared spectroscopy for medical applications: Current status and future perspectives. Clin. Chim. Acta; 2016; 455, pp. 181-188. [DOI: https://dx.doi.org/10.1016/j.cca.2016.02.009]

8. Willingham, T.B.; McCully, K.; Backus, D. Skeletal muscle dysfunction in people with multiple sclerosis: A physiological target for improving physical function and mobility. Arch. Phys. Med. Rehabil.; 2023; 104, pp. 694-706. [DOI: https://dx.doi.org/10.1016/j.apmr.2022.10.009]

9. Luque, E.; Ruz-Caracuel, I.; Medina, F.J.; Leiva-Cepas, F.; Agüera, E.; Sánchez-López, F.; Lillo, R.; Aguilar-Luque, M.; Jimena, I.; Túnez, I. . Skeletal muscle findings in experimental autoimmune encephalomyelitis. Pathol. Res. Pract.; 2015; 211, pp. 493-504. [DOI: https://dx.doi.org/10.1016/j.prp.2015.02.004]

10. Castelli, S.; Carinci, E.; Baldelli, S. Oxidative Stress in Neurodegenerative Disorders: A Key Driver in Impairing Skeletal Muscle Health. Int. J. Mol. Sci.; 2025; 26, 5782. [DOI: https://dx.doi.org/10.3390/ijms26125782]

11. Rajda, C.; Pukoli, D.; Bende, Z.; Majláth, Z.; Vécsei, L. Excitotoxins, mitochondrial and redox disturbances in multiple sclerosis. Int. J. Mol. Sci.; 2017; 18, 353. [DOI: https://dx.doi.org/10.3390/ijms18020353]

12. López-Muguruza, E.; Matute, C. Alterations of oligodendrocyte and myelin energy metabolism in multiple sclerosis. Int. J. Mol. Sci.; 2023; 24, 12912. [DOI: https://dx.doi.org/10.3390/ijms241612912]

13. Haider, L. Inflammation, iron, energy failure, and oxidative stress in the pathogenesis of multiple sclerosis. Oxidative Med. Cell. Longev.; 2015; 2015, 725370. [DOI: https://dx.doi.org/10.1155/2015/725370]

14. Adiele, R.C.; Adiele, C.A. Metabolic defects in multiple sclerosis. Mitochondrion; 2019; 44, pp. 7-14. [DOI: https://dx.doi.org/10.1016/j.mito.2017.12.005]

15. Borragán, G.; Gilson, M.; Atas, A.; Slama, H.; Lysandropoulos, A.; De Schepper, M.; Peigneux, P. Cognitive Fatigue, Sleep and Cortical Activity in Multiple Sclerosis Disease. A Behavioral, Polysomnographic and Functional Near-Infrared Spectroscopy Investigation. Front. Hum. Neurosci.; 2018; 12, 378. [DOI: https://dx.doi.org/10.3389/fnhum.2018.00378] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/30294266]

16. Danner, M.; Locherer, M.; Hank, T.; Richter, K. Spectral Sampling with the ASD FIELDSPEC 4; EnMAP Consortium: Potsdam, Germany, 2015.

17. Rinnan, Å.; van den Berg, F.; Engelsen, S.B. Review of the most common pre-processing techniques for near-infrared spectra. TrAC Trends Anal. Chem.; 2009; 28, pp. 1201-1222. [DOI: https://dx.doi.org/10.1016/j.trac.2009.07.007]

18. Wold, S.; Esbensen, K.; Geladi, P. Principal component analysis. Chemom. Intell. Lab. Syst.; 1987; 2, pp. 37-52. [DOI: https://dx.doi.org/10.1016/0169-7439(87)80084-9]

19. Kennard, R.W.; Stone, L.A. Computer aided design of experiments. Technometrics; 1969; 11, pp. 137-148. [DOI: https://dx.doi.org/10.1080/00401706.1969.10490666]

20. Fawcett, T. An introduction to ROC analysis. Pattern Recognit. Lett.; 2006; 27, pp. 861-874. [DOI: https://dx.doi.org/10.1016/j.patrec.2005.10.010]

21. Chong, I.-G.; Jun, C.-H. Performance of some variable selection methods when multicollinearity is present. Chemom. Intell. Lab. Syst.; 2005; 78, pp. 103-112. [DOI: https://dx.doi.org/10.1016/j.chemolab.2004.12.011]

22. Cohen, J. Statistical Power Analysis for the Behavioral Sciences; Routledge: New York, NY, USA, 2013.

23. Weyer, L.; Lo, S. Spectra-structure correlations in the near-infrared. Handb. Vib. Spectrosc.; 2002; 3, pp. 1817-1837.

24. Wilson, R.H.; Nadeau, K.P.; Jaworski, F.B.; Tromberg, B.J.; Durkin, A.J. Review of short-wave infrared spectroscopy and imaging methods for biological tissue characterization. J. Biomed. Opt.; 2015; 20, 030901. [DOI: https://dx.doi.org/10.1117/1.JBO.20.3.030901]

25. Caredda, M.; Dedola, A.S.; Pes, M.; Addis, M. The Use of NIR Spectroscopy and Chemometrics to Identify the Thermal Treatment of Milk in Fiore Sardo PDO Cheese to Detect Fraud. Foods; 2025; 14, 2288. [DOI: https://dx.doi.org/10.3390/foods14132288]

26. Podbielska, M.; Banik, N.L.; Kurowska, E.; Hogan, E.L. Myelin recovery in multiple sclerosis: The challenge of remyelination. Brain Sci.; 2013; 3, pp. 1282-1324. [DOI: https://dx.doi.org/10.3390/brainsci3031282]

27. Laule, C.; Pavlova, V.; Leung, E.; Zhao, G.; MacKay, A.L.; Kozlowski, P.; Traboulsee, A.L.; Li, D.K.B.; Moore, G.R.W. Diffusely abnormal white matter in multiple sclerosis: Further histologic studies provide evidence for a primary lipid abnormality with neurodegeneration. J. Neuropathol. Exp. Neurol.; 2013; 72, pp. 42-52. [DOI: https://dx.doi.org/10.1097/NEN.0b013e31827bced3] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23242281]

28. Fleischer, V.; Brummer, T.; Muthuraman, M.; Steffen, F.; Heldt, M.; Protopapa, M.; Schraad, M.; Gonzalez-Escamilla, G.; Groppa, S.; Bittner, S. . Biomarker combinations from different modalities predict early disability accumulation in multiple sclerosis. Front. Immunol.; 2025; 16, 1532660. [DOI: https://dx.doi.org/10.3389/fimmu.2025.1532660] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39958357]

29. Tolentino, M.; Pace, F.; Perantie, D.C.; Mikesell, R.; Huecker, J.; Chahin, S.; Ghezzi, L.; Piccio, L.; Cross, A.H. Cerebrospinal fluid biomarkers as predictors of multiple sclerosis severity. Mult. Scler. Relat. Disord.; 2025; 94, 106268. [DOI: https://dx.doi.org/10.1016/j.msard.2025.106268] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39832432]

30. Zierfuss, B.; Larochelle, C.; Prat, A. Blood–brain barrier dysfunction in multiple sclerosis: Causes, consequences, and potential effects of therapies. Lancet Neurol.; 2024; 23, pp. 95-109. [DOI: https://dx.doi.org/10.1016/S1474-4422(23)00377-0]

31. Ghorbani, S.; Yong, V.W. The extracellular matrix as modifier of neuroinflammation and remyelination in multiple sclerosis. Brain; 2021; 144, pp. 1958-1973. [DOI: https://dx.doi.org/10.1093/brain/awab059]

32. Chen, S.-M.; Tang, X.-Q. Homocysteinylation and sulfhydration in diseases. Curr. Neuropharmacol.; 2022; 20, pp. 1726-1735. [DOI: https://dx.doi.org/10.2174/1570159X20666211223125448]

33. Gupta, R.; Sahu, M.; Tripathi, R.; Ambasta, R.K.; Kumar, P. Protein S-sulfhydration: Unraveling the prospective of hydrogen sulfide in the brain, vasculature and neurological manifestations. Ageing Res. Rev.; 2022; 76, 101579. [DOI: https://dx.doi.org/10.1016/j.arr.2022.101579]

34. Soroush, A.; Dunn, J.F. A Hypoxia-Inflammation Cycle and Multiple Sclerosis: Mechanisms and Therapeutic Implications. Curr. Treat Options Neurol.; 2025; 27, 6. [DOI: https://dx.doi.org/10.1007/s11940-024-00816-4]

35. Su, K.; Bourdette, D.; Forte, M. Mitochondrial dysfunction and neurodegeneration in multiple sclerosis. Front. Physiol.; 2013; 4, 169. [DOI: https://dx.doi.org/10.3389/fphys.2013.00169] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/23898299]

36. Narayana, P.A. Magnetic resonance spectroscopy in the monitoring of multiple sclerosis. J. Neuroimaging; 2005; 15, pp. 46S-57S. [DOI: https://dx.doi.org/10.1177/1051228405284200] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/16385018]

37. Harbige, L.S.; Sharief, M.K. Polyunsaturated fatty acids in the pathogenesis and treatment of multiple sclerosis. Br. J. Nutr.; 2007; 98, pp. S46-S53. [DOI: https://dx.doi.org/10.1017/S0007114507833010] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/17922959]

38. Butnaru, D.; Chapman, J. The impact of self-replicating proteins on inflammation, autoimmunity and neurodegeneration—An untraveled path. Autoimmun. Rev.; 2019; 18, pp. 231-240. [DOI: https://dx.doi.org/10.1016/j.autrev.2018.09.009]

39. Wens, I.; Dalgas, U.; Vandenabeele, F.; Krekels, M.; Grevendonk, L.; Eijnde, B.O.; Asakura, A. Multiple sclerosis affects skeletal muscle characteristics. PLoS ONE; 2014; 9, e108158. [DOI: https://dx.doi.org/10.1371/journal.pone.0108158]

40. Kent-Braun, J.A.; Ng, A.V.; Castro, M.; Weiner, M.W.; Gelinas, D.; Dudley, G.A.; Miller, R.G. Strength, skeletal muscle composition, and enzyme activity in multiple sclerosis. J. Appl. Physiol.; 1997; 83, pp. 1998-2004. [DOI: https://dx.doi.org/10.1152/jappl.1997.83.6.1998]

41. Wischnewski, S.; Rausch, H.-W.; Ikenaga, C.; Leipe, J.; Lloyd, T.E.; Schirmer, L. Emerging mechanisms and therapeutics in inflammatory muscle diseases. Trends Pharmacol. Sci.; 2025; 46, pp. 249-263. [DOI: https://dx.doi.org/10.1016/j.tips.2025.01.005]

42. Dionyssiotis, Y.; Mavrogenis, A.; Trovas, G.; Skarantavos, G.; Papathanasiou, J.; Papagelopoulos, P. Bone and soft tissue changes in patients with spinal cord injury and multiple sclerosis. Folia Medica; 2014; 56, 237. [DOI: https://dx.doi.org/10.1515/folmed-2015-0002]

43. Barcelos, I.P.; Troxell, R.M.; Graves, J.S. Mitochondrial dysfunction and multiple sclerosis. Biology; 2019; 8, 37. [DOI: https://dx.doi.org/10.3390/biology8020037]

44. Garner, D.J.; Widrick, J.J. Cross-bridge mechanisms of muscle weakness in multiple sclerosis. Muscle Nerve Off. J. Am. Assoc. Electrodiagn. Med.; 2003; 27, pp. 456-464. [DOI: https://dx.doi.org/10.1002/mus.10346]

45. Marinelli, L.; Mori, L.; Canneva, S.; Colombano, F.; Currà, A.; Fattapposta, F.; Bandini, F.; Capello, E.; Abbruzzese, G.; Trompetto, C. The effect of cannabinoids on the stretch reflex in multiple sclerosis spasticity. Int. Clin. Psychopharmacol.; 2016; 31, pp. 232-239. [DOI: https://dx.doi.org/10.1097/YIC.0000000000000126] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/27003093]

46. Comi, G.; Solari, A.; Leocani, L.; Centonze, D.; Otero-Romero, S. Italian consensus on treatment of spasticity in multiple sclerosis. Eur. J. Neurol.; 2020; 27, pp. 445-453. [DOI: https://dx.doi.org/10.1111/ene.14110] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/31652369]

47. De Aratanha, M.A.; Balardin, J.B.; Cardoso do Amaral, C.; Lacerda, S.S.; Sowmy, T.A.S.; Huppert, T.J.; Thomaz, R.B.; Speciali, D.S.; Machado, B.; Kozasa, E.H. The use of functional near infrared spectroscopy and gait analysis to characterize cognitive and motor processing in early-stage patients with multiple sclerosis. Front. Neurol.; 2022; 13, 937231. [DOI: https://dx.doi.org/10.3389/fneur.2022.937231] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36105774]

48. Bonilauri, A.; Intra, F.S.; Pugnetti, L.; Baselli, G.; Baglio, F. A systematic review of cerebral functional near-infrared spectroscopy in chronic neurological diseases—Actual applications and future perspectives. Diagnostics; 2020; 10, 581. [DOI: https://dx.doi.org/10.3390/diagnostics10080581]

49. Kharati, M.; Foroutanparsa, S.; Rabiee, M.; Salarian, R.; Rabiee, N.; Rabiee, G. Early diagnosis of multiple sclerosis based on optical and electrochemical biosensors: Comprehensive perspective. Curr. Anal. Chem.; 2020; 16, pp. 557-569. [DOI: https://dx.doi.org/10.2174/1573411014666180829111004]

50. Condino, F.; Crocco, M.C.; Pirritano, D.; Petrone, A.; Del Giudice, F.; Guzzi, R. A Linear Predictor Based on FTIR Spectral Biomarkers Improves Disease Diagnosis Classification: An Application to Multiple Sclerosis. J. Pers. Med.; 2023; 13, 1596. [DOI: https://dx.doi.org/10.3390/jpm13111596]