Capillary serum-based spectroscopic signature for multiple sclerosis: A novel diagnostic perspective

Abstract

Objective

To analyze the biochemical characteristics of capillary blood-derived serum using Fourier-transform infrared (FTIR) spectroscopy, aiming to differentiate individuals diagnosed with Multiple Sclerosis (MS) from a matched control group (CG).

Methods

The study included 52 patients with confirmed MS and 52 healthy control participants, matched by sex and age. Serum samples were obtained via capillary blood collection and analyzed using mid-infrared ATR-FTIR spectroscopy. Chemometric analyses were performed on the full spectrum (3700–900 cm⁻¹) and on the fingerprint (1800–900 cm⁻¹) and high-wavenumber (3700–2800 cm⁻¹) regions separately. Preprocessing included baseline correction and smoothing, followed by multivariate modeling using Partial Least Squares Discriminant Analysis (PLS-DA).

Results

FTIR spectroscopy combined with supervised machine learning revealed metabolic alterations between groups, including increased concentrations of phospholipids/lipids, nucleic acids, and Amide I/II in MS patients. The high-wavenumber region showed enhanced differentiation of control individuals, whereas the fingerprint region provided better separation between MS and CG. The combination of both spectral regions (full spectrum) yielded balanced performance metrics, sensitivity, specificity, and accuracy of 87.5%, highlighting their complementary diagnostic value.

Conclusion

FTIR spectroscopy of capillary blood serum provides distinct and complementary biochemical signatures for MS diagnosis. When integrated with machine learning algorithms, this approach demonstrated high discriminatory power. This is the first study to employ serum for capillary blood for FTIR-based MS classification, supporting its potential as a non-invasive, rapid, and accessible tool to assist in clinical decision-making.

Graphical abstract

Highlights

• •A less invasive analysis using capillary serum combined with FTIR spectroscopy.
• •Were distinguished between Multiple Sclerosis and the control group FTIR profiles.
• •Bands of proteins, lipids, and nucleic acids provided essential information.
• •The model achieved 87.5% sensitivity, specificity, and accuracy.
• •Biospectroscopy is proposed as a promising complementary tool for MS screening.

1. Introduction

Multiple sclerosis (MS) is a chronic, immune-mediated, inflammatory, demyelinating and neurodegenerative disease of the central nervous system (CNS). It is one of the leading cause of non-traumatic neurological disability in young adults [1]. Its etiology is complex and multifactorial, involving genetic, infectious, and environmental interactions that trigger an abnormal immune response, leading to demyelination and axonal injury [2]. Disease progression is typically assessed using the Expanded Disability Status Scale (EDSS), which quantifies neurological impairment over time across eight functional domains: visual, brainstem, pyramidal, cerebellar, sensory, bowel and bladder, cerebral, and ambulation. Higher EDSS scores correlate with greater disability severity [3].

The prevalence of MS varies geographically, with regions classified into low, medium and high prevalence zones. It is estimated that approximately 2.8 million people live with MS worldwild, with a ratio of 1:3000 individuals between the ages of 20 and 50 years. In Brazil, the average prevalence is about 15 cases per 100,000 inhabitants [4]. MS diagnosis has undergone several refinements over recent decades. Although no definitive biomarker exists, current diagnostic criteria are based on the integration of clinical history with radiological and laboratory findings [5]. According to the McDonald criteria revised in 2017, diagnosis is based on objective evidence of lesions dissemination in time and space, supported by magnetic resonance imaging (MRI), cerebrospinal fluid (CSF) analysis and visual evoked potentials [6,7].

Vibrational spectroscopy, particularly Fourier-transform infrared (FTIR) spectroscopy, has emerged as a promising tool for biomedical diagnostics due to its label-free nature, minimal sample requirements, and ability to capture global biochemical fingerprints. FTIR detects subtle molecular alterations in protein secondary structure, lipid content, and nucleic acid composition, biomarkers that shift during demyelination, oxidative stress, and neuroinflammation. Biofluids such as blood, saliva, and CSF are increasingly explored as sources for such diagnostic signatures. In MS, demyelination results in significant changes in the FTIR spectrum, reflecting myelin degradation and metabolite accumulation [8].

Biofluids such as blood, saliva and urine are increasingly studied in this context for their accessibility and diagnostic relevance [9]. Capillary blood, obtained from small blood vessels, contains various components and is widely used in rapid tests, such as glucose measurement. Its ease of collection and spectroscopic analysis offer valuable insights for diagnostics and advances in biomedical and clinical research [10,11].

Previous studies have demonstrated that FTIR can identify spectral differences between MS patients and healthy controls, particularly in regions corresponding to proteins, lipids, and nucleic acids. Kołodziej et al. [12] identified alterations in protein and lipid regions, including increased β-sheet proteins and lipid fractions, along with decreased tyrosine residues, indicating the technique's feasibility for differentiating MS patients. Similarly, Yonar et al. [8] distinguished patients with MS and Clinically Isolated Syndrome from controls using CSF analysis, identifying spectral differences in lipid, RNA, and carbohydrate regions that enabled discrimination between patients and controls.

Despite promising results in related fields, to the best of our knowledge, no studies have evaluated FTIR spectroscopy of capillary serum for the investigation of multiple sclerosis. This study aims to assess the feasibility of using FTIR spectroscopy on capillary serum samples to distinguish MS patients from healthy controls. Additionally, we investigate which spectral regions (fingerprint, high-wavenumber or the full spectrum) provide the most relevant biochemical information for the classification model.

2. Methods

2.1. Population study

This study included 52 participants diagnosed with MS, recruited from the Neurology Outpatient Clinic of the Cassiano Antônio Moraes University Hospital (HUCAM), and 52 healthy control participants, age- and gender-matched (CG). Clinical data from MS patients were obtained from their medical records. Inclusion criteria consisted of confirmed diagnosis of MS according to the McDonald criteria [7], age over 18 years, and voluntary participation, confirmed by a signed Informed Consent Form (ICF). For the CG, residents from the Greater Vitória region were randomly selected, provided they met the study's eligibility criteria. Exclusion criteria included individuals with any active infections, pregnant women, and those diagnosed with other neurological disorders.

The study was approved by the Research Ethics Committee of the Hospital Universitário Cassiano Antônio Moraes, Federal University of Espírito Santo (HUCAM/UFES), under CAAE No. 44387521.8.0000.5071.

2.2. Sample collection

Capillary blood samples were collected by obtaining three microcapillary tubes from each participant's fingertip puncture. After collection, samples were centrifuged at 1500 x g for 5 min at room temperature. After centrifugation, 50 μL of serum was transferred into a 500 μL Eppendorf microtube. The microtubes were stored in a freezer at −80°C until spectral analysis.

For spectral acquisition, samples were thawed for 1 h at room temperature (22.6 ± 1.3°C) and relative humidity of 35 ± 3.1%. The samples were then homogenized using a vortex mixer. Then, 10 μL of serum was deposited in triplicate onto aluminum microplates (5 mm × 5 mm) and air-dried for at least 2 h at room temperature before FTIR data acquisition.

Spectra were acquired using the ALPHA II Compact FTIR spectrometer (Bruker Optics, Ettlingen, Germany), equipped with a diamond attenuated total reflectance (ATR) accessory. Spectra were recorded using OPUS software (version 7.8, Bruker Corporation, Germany) in absorbance mode covering the 4000 to 400 cm⁻¹ range, with a spectral resolution of 4 cm⁻¹ and 32 scans for each sample and background. The dried serum films on aluminum plates were directly positioned on the ATR diamond crystal and a pressure accessory was used to ensure optimal contact between the dried film and the crystal surface. The crystal surface window was cleaned with Milli-Q® water and 70% v/v (Sigma-Aldrich, St Louis, MO, USA) ethanol before each analysis and dried with absorbent paper.

2.3. Chemometric analysis

All spectra were imported into Matlab software (version 2023a, MathWorks®), Baseline correction was applied using the Adaptive Iteratively Reweighted Penalized Least Squares (airPLS) [13] algorithm, followed by smoothing with a Savitzky-Golay filter and averaging of triplicates. Multivariate analysis was performed to discriminate samples based on spectral differences, focusing on the following regions of interest: full spectral region (3700–900 cm⁻¹), high wavenumber region (3700–2800 cm⁻¹) and fingerprint (1800–900 cm⁻¹). The commonly referred to “silent region” (2800–1800 cm-1) in biospectroscopy studies was excluded because it lacked relevant signals. Different preprocessing methods were tested, including vector normalization, Savitzky-Golay derivatives, standard normal variate (SNV), multiplicative scatter correction (MSC) and their combinations to enhance model performance and interpretability.

Statistical multivariate analysis was carried out using pattern recognition methods, such as Principal Component Analysis (PCA), and machine learning with a Random Forest (RF). A combined unsupervised random forest with principal coordinate analysis and supervised random forest (URF-PCoA-RF) was used to evaluate natural groupings and identify relevant variables for discriminating CG and MS groups. Partial Least Squares Discriminant Analysis (PLS-DA) was also applied for classification models using latent PLS variables instead of the original variables.

To validate the models, the dataset was split into training and test subsets using the Kennard-Stone algorithm, ensuring uniform sample distribution. Approximately 70% of the data was allocated for training, while 30% was reserved for an external test (unseen samples to the model). Training samples were used to determine the model's optimal number of latent variables through Monte Carlo cross-validation. The number of LVs was selected based on the highest AUC and stabilization of performance metrics during Monte Carlo cross-validation. External test samples were reserved to evaluate the model's predictive performance, yielding metrics such as Sensitivity, Specificity, False Positive Rate, False Negative Rate, F1-Score, and Accuracy. The corresponding equations are presented in Table 1.

Table 1.

Equations for the performance parameters applied in the PLS-DA model.

Performance Parameters	Equation
Sensitivity (SENS)	$SENS=\frac{TP}{TP+TN}$
Specificity (SPEC)	$SPEC=\frac{TN}{FP+TN}$
Accuracy (ACC)	$ACC=\frac{TP+TN}{TP+FN+FP+TN}$
False Positive Rate (FPR)	$FPR=\frac{FP}{FP+TN}$
False Negative Rate (FNR)	$FNR=\frac{FN}{FN+TP}$
Precision (PREC)	$PREC=\frac{TP}{TP+FP}$
F-score	$F1=2·\frac{\left(\text{Precision}·\text{Sensitivity}\right)}{\left(\text{Precision}+\text{Sensitivity}\right)}$

Legend: TP: True Positive; TN: True Negative; FP: False Positive; FN: False Negative.

Additionally, the Receiver Operating Characteristic (ROC) curve was used to evaluate the accuracy of the PLS-DA models in each spectral region. The PLS-DA model also allowed for the interpretation of relevant variables through the Variable Importance in Projection (VIP) analysis. The VIP score was obtained for each variable in classifying the samples, summarizing the importance and significance of these variables. A VIP value greater than 1 was adopted as the criterion for considering a variable to significantly contribute to the model [14].

2.4. Statistical analysis

Socio-demographic and clinical data were analyzed using GraphPad Prism 8 (GraphPad Software). The normality of continuous variables was assessed using the Shapiro-Wilk test. Results are presented as mean ± standard deviation. Fisher's exact test and the Mann-Whitney test were used for comparisons. A p-value < 0.05 was considered statistically significant.

3. Results and discussion

The sociodemographic characteristics for both MS and CG groups are presented in Table 2. No statistically significant differences were observed between the groups regarding gender and age. These results indicate an adequate matching between the groups in terms of age and gender. A complete individual description of the characteristics of all participants can be found in Tables S1 and S2 of the supplementary data.

Table 2.

Comparison of sociodemographic characteristics between the Multiple Sclerosis and Control groups.

		MS (n = 52)	CG (n = 52)	p-value
Gender	Male	17 (32.7%)	16 (30.8%)	> 0.99
	Female	35 (67.3%)	36 (69.2%)
Age (mean ± SD)		39.5(±12.7)	39.3 (±12.4)	0.96

The clinical characteristics of the MS group are presented in Table 3. Among the 52 patients, 43 (82.7%) were identified as having the relapsing-remitting MS phenotype, while 9 (17.3%) were identified as having the progressive MS phenotype. Regarding disability status, 38 (73.1%) were classified as having a mild functional impairment (EDSS scores 0-3), while 14 (26.9%) had moderate to severe disability (EDSS scores 3.5-7.5). The median age at diagnosis was 29 years, with a median disease duration of 7 years. The number of reported relapses ranged from 1 to 12. Regarding treatment, the most commonly used medications were Natalizumab (31.1%) and Dimethyl Fumarate (28.9%), followed by other agents such as Fingolimod Hydrochloride (11.1%) and Interferon Beta (6.7%).

Table 3.

Clinical characteristics of participants diagnosed with multiple sclerosis.

MS Phenotypes
Relapsing-remitting	n (%)	43 (82.7%)
Secondary progressive	n (%)	9 (17.3%)
EDSS
0 – 3	n	38
3.5 -7.5	n	14
Age at diagnosis	Median (min – máx)	29 years (12 – 58)
Disease duration	Median (min – máx)	7 years (1 – 24)
Number of relapses	Median (min – máx)	4 (1 – 12)
Medication
Dimethyl fumarate	n (%)	13 (28.9%)
Natalizumab	n (%)	14 (31.1%)
Fingolimod hydrochloride	n (%)	5 (11.1%)
Interferon beta 1	n (%)	3 (6.7%)
Others	n (%)	10 (22.2%)

Legend: EDSS: Expanded Disability Status Scale.

FTIR spectroscopy was used to acquire the spectral data followed by pattern recognition analysis. Fig. 1 represents the average spectral profile of the capillary serum samples from the GC and MS groups. This spectrum represents the spectral profile of chemical bonds of functional groups present in the samples and was analyzed in two regions: fingerprint (1800 – 900 cm⁻¹) and high wavenumber (3700 – 2800 cm⁻¹), as well as the combination of these two regions together (4000 – 2800 cm⁻¹ | 1800 – 900 cm⁻¹).

Fig. 1.

Mean FTIR-ATR absorbance spectra of capillary serum from individuals with multiple sclerosis (MS, red) and healthy controls (CG, blue). Shaded areas represent ±1 standard deviation. The spectral region between approximately 1800 and 2700 cm⁻¹ was omitted due to the absence of diagnostically relevant bands. Prior to analysis, spectra were preprocessed by baseline correction using the adaptive iteratively reweighted penalized least squares (airPLS) algorithm, followed by smoothing with a Savitzky-Golay filter.

The unsupervised exploratory analysis (URF-PCoA), shown in Fig. 2, revealed that the samples were separated into different groups based on PCo2, with an explained variance of 69.1% from PCo1 to PCo3 in the full spectrum (Figs. 2A), 72.8% in the high wavenumber region (Fig. 2B), and 79.7% in the fingerprint region (Fig. 2C). Supervised analysis (URF-PCoA-RF) was conducted to identify the most influential variables, as illustrated in Fig. 2D, E, and 2F, thereby enabling the distinct MS and CG spectral groups.

Fig. 2.

Machine learning analysis of FTIR spectral data. (A–C): Score plots generated by the unsupervised random forest with principal coordinate analysis (URF-PCoA) using normalized spectra: (A) Full spectrum (3700–900 cm⁻¹), (B) High-wavenumber region (3700–2800 cm⁻¹) and (C) Fingerprint region (1800–900 cm⁻¹). (D–F): Relevant spectral variables identified by the supervised random forest model for: (D) Full spectrum, (E) High-wavenumber region and (F) Fingerprint region. Vertical dashed lines indicate variables with high discriminatory power between the MS (red line) and control (blue line) groups.

Subsequently, applying the PLS-DA method enables the assessment of the predictive performance for group classification. All spectral regions analyzed showed satisfactory performance, with detailed parameters provided in Table 4. In the high-wavenumber region, the model achieved sensitivities, specificities, and accuracies of 87.5%. In the fingerprint region, sensitivity was 100%, specificity 75%, and accuracy 87.5%. In the full spectrum, sensitivity, specificity and accuracy also reached 87.5%. The confusion matrices for the described models are available in the supplementary material (Tables S3, S4, and S5).

Table 4.

Performance metrics of the PLS-DA model for classification of MS and CG using ATR-FTIR spectra from capillary serum samples (n = 104).

Spectral Region	LV	Group	SENS (%)	SPEC (%)	FPR (%)	FNR (%)	F–score	Accuracy
High-Wavenumber	6	Train	97.2	97.2	2.78	2.78	0.972	97.2
		Test	87.5	87.5	12.5	12.5	0.875	87.5
Fingerprint	5	Train	88.8	80.5	19.4	11.1	0.853	84.7
		Test	100	75.0	25.0	0	0.888	87.5
Full Spectrum	4	Train	91.6	80.5	19.4	8.33	0.868	86.1
		Test	87.5	87.5	12.5	12.5	0.875	87.5

Latent variables; SENS, Sensitivity; SPEC, Specificity; FPR, False Positive Rate; FNR, False Negative Rate. Filled row: training subset; unfilled row: external test subset.

The ROC curve was used to calculate AUC values, which served as a comparison parameter among the PLS-DA models, and the ROC curve demonstrated satisfactory performance. Fig. 3A refers to the training subset, while Fig. 3B corresponds to the external validation (test) subset. The models achieved AUCs of 0.96 in the high-wavenumber region, 0.96 in the fingerprint region, and 0.94 across the full spectrum on the test set. The ROC curve data and AUC values confirmed the model's effectiveness in distinguishing between the groups. AUC values above 0.90 for all models indicate excellent discrimination capacity, consistent with clinically useful diagnostic tools. Notably, the high-wavenumber and fingerprint regions individually reached AUCs of 0.96, reinforcing their robustness even in isolation. The combination of fingerprint and high-wavenumber regions, which span the full spectrum, balanced sensitivity, specificity, and accuracy, indicating that complementary information from both regions are valuable for classification.

Fig. 3.

ROC curve and characteristics of the PLS-DA model based on the Multiple Sclerosis class (Fig. A, training subset and Fig. B, external test subset).

To the best of our knowledge, this is the first study utilizing capillary blood serum for MS classification via FTIR spectroscopy, representing an innovation compared to existing literature, which traditionally focuses on other biofluids such as plasma, venous serum, skin fibroblasts, and tears, or more invasive samples like cerebrospinal fluid and brain tissue [8,[12], [13], [14], [15], [16], [17], [18], [19]].

The predictive performance observed in this study, particularly the external validation metrics with sensitivity and specificity reaching 87.5%, underscores the potential of FTIR spectroscopy as a complementary diagnostic approach for multiple sclerosis. While not intended to replace established methods such as MRI or CSF analysis, the moderate-to-high accuracy achieved with minimally invasive capillary serum samples positions this technique as a valuable screening tool, especially in primary care or low-resource settings where access to neuroimaging and specialized testing is limited. Furthermore, the rapid acquisition, low operational cost, and reproducibility of the FTIR method make it attractive for longitudinal monitoring or early triage, thereby contributing to faster clinical decision-making and more accessible neurological care. Previous studies reinforce the technique's potential in different contexts. Yonar et al. [8] reported 95% sensitivity and 92% specificity in differentiating MS, Clinically Isolated Syndrome (CIS), and control group using cerebrospinal fluid. El Khoury et al. [15] reported 50% sensitivity and 100% specificity for MS when analyzing venous serum, whereas the ALS group showed 75% sensitivity and 75% specificity. Crocco et al. [18] reported 78% sensitivity and 83% specificity in distinguishing MS from CG using plasma. El Khoury et al. [19] also achieved 100% sensitivity and specificity when differentiating between Neuromyelitis Optica, MS, and CG using venous blood serum.

The analysis of the most relevant spectral variables for building the PLS-DA models, shown in Fig. 4, highlights the main biochemical components responsible for distinguishing between the groups. The variables highlighted in the VIP analysis are described in Table 5.

Fig. 4.

Variable Importance in Projection (VIP) analysis of spectral variables by the PLS-DA model (A): high wavenumber region; (B): Fingerprint region; (C): full spectrum.

Table 5.

Spectral regions identified by the VIP analysis are relevant in the capillary serum spectra.

Regions	Assignment	Biomolecules	Fingerprint	High Wavenumber	Full Spectrum	p-Value
2960	υas(CH3)	Lipids and Proteins		x	x	0.002
2929	υas(CH2)	Lipids and Proteins		x	x	0.017
2923	Lipids (CH3 stretching)	Lipids		x	x	0.034
1741	Phospholipids	Lipids	x		x	0.003
1683	Amide I: antiparallel β-sheets and turns	Proteins	x		x	0.035
1653	Amide I (α-helical)	Proteins	x		x	0.025
1640	Amide I	Proteins	x		x	0.002
1636	Amide I: υs(C=O) coupled with υ(C–N) and δ(N–H)	Proteins	x		x	0.001
1626	Amide I (α-helical protein structures)	Proteins	x		x	0.003
1542	Amide II (α-helix and β-sheet)	Proteins	x		x	0.001
1535	Amide II	Proteins	x		x	0.002
1515	Symmetric stretching (C=C) of tyrosine	Proteins	x		x	0.001
1513	Amide II (N–H and C–N stretching)	Proteins	x		x	0.001
1379	δ (CH)3	Lipids and Proteins			x	0.019
1106	υ(C − N) e δ(C − H)	Proteins	x		x	0.009
1096	Lipids	Lipids	x		x	0.008
1085	Phospholipids	Lipids	x		x	0.001
1078	C–O stretching and bending of C–OH group/υs (PO2) of DNA, phospholipids, saccharides	Nucleic Acids	x		x	0.001
1047	υ e δ of groups (carbohydrate)	Carbohydrates	x		x	0.001
1041	DNA, RNA	Nucleic Acids	x		x	0.015
1031	δ(C − OH) Glucose, Polysaccharides	Carbohydrates	x		x	0.002
1016	υ(C–O)	RNA	x		x	0.002
1012	υ of DNA (nucleic acids)	Nucleic Acids	x		x	0.001

Legend: υ: stretching; υs: symmetric stretching; υas: asymmetric stretching; δ: in-plane deformation; x: bands present in different regions of the spectrum. Mann-Whitney test, p < 0.5.

The selected variables in Table 5 cover spectral regions related to lipids, proteins and nucleic acids. These molecules are associated with critical biological processes in the pathophysiology of MS, such as demyelination, oxidative stress and inflammation. As observed in previous studies, the p-values indicate statistical significance between the groups in the spectral regions analyzed [12,[15], [16], [17], [18], [19]].

The analyses considered the full spectrum (3700–2800 cm⁻¹; 1800–900 cm⁻¹), the specific fingerprint region (1800–900 cm⁻¹), and the high-wavenumber region (3700–2800 cm⁻¹). Relevant spectral bands include ∼1012, ∼1016, ∼1031, ∼1041, ∼1047, and ∼1078 cm⁻¹, associated with the absorption of carbohydrates, DNA, and RNA. These bands suggest the presence of damaged cellular fragments, programmed cell death, and other cellular processes characteristic of neurodegeneration. Other spectral bands (∼1106, ∼1513, ∼1515, ∼1542, ∼1626, ∼1636, ∼1640, ∼1653, and ∼1683 cm⁻¹) are related to proteins, particularly the Amide I and Amide II bands, which reflect structural conformations such as beta sheets and alpha-helix. These structural changes are associated with neurodegeneration and the presence of tyrosine, a marker of oxidative stress. The band ∼1741 cm⁻¹ was attributed to phospholipids, molecules essential for the composition and function of the myelin sheath, whose degradation directly impacts nerve conduction in patients with MS. Changes in lipid concentrations were observed in spectral regions such as ∼1085, ∼1096, ∼1379, ∼2923, ∼2929, and ∼2960 cm⁻¹, indicating an increase in these compounds in serum, possibly due to the demyelination process [12,[15], [16], [17], [18], [19], [20], [21], [22], [23]].

The prominence of bands in the 2923–2960 cm⁻¹ range reflects increased lipid content, possibly due to myelin sheath degradation, a central feature of MS. Likewise, changes in Amide I and II bands (∼1653, ∼1542 cm⁻¹) suggest altered protein secondary structure, consistent with neuroinflammation and protein misfolding processes.

It is important to highlight that FTIR spectroscopy is easy to handle, offers a good cost-benefit ratio, requires minimal sample preparation, and is highly reproducible, making it a more accessible method. However, the technique must be evaluated in comparative studies with diagnostic methods considered the gold standard to provide a more comprehensive understanding of its performance compared to conventional diagnostic approaches.

Some limitations must be considered in our study. These include the broad age range of participants and the variation in the time of disease diagnosis. Using FTIR, it was possible to distinguish the groups; however, more comprehensive studies are needed to explore differences in spectral profiles among biofluids (blood, CSF, saliva, urine) and to identify specific bands. It is also crucial to compare these spectral profiles with those of other neurodegenerative diseases to establish a characteristic MS spectral profile that can serve as an adjuvant tool for disease diagnosis.

4. Conclusion

This study is the first to demonstrate the applicability of mid-infrared FTIR spectroscopy, combined with machine learning, for the classification of multiple sclerosis (MS) using capillary blood-derived serum. The approach proved effective in distinguishing MS patients from healthy controls, achieving high sensitivity, specificity, and overall accuracy. Spectral regions associated with lipids, proteins, and nucleic acids significantly contributed to this discrimination, reflecting key biological processes such as demyelination, oxidative stress, and neuroinflammation. The joint analysis of the fingerprint and high-wavenumber regions revealed complementary biochemical signatures, underscoring the benefit of integrating both spectral windows in diagnostic models.

Given its simplicity, cost-effectiveness, minimal sample volume requirement, and non-invasive collection, FTIR spectroscopy emerges as a promising complementary tool for MS screening and clinical decision-making, particularly in resource-limited settings. Future studies should investigate its applicability in differentiating MS from other neurodegenerative diseases, assess its performance in longitudinal and multicenter cohorts, and explore its integration into real-world diagnostic workflows.

CRediT authorship contribution statement

Raí S. Santiago: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Marcia H.C. Nascimento: Writing – review & editing, Validation, Software, Methodology, Investigation, Formal analysis, Data curation. Jaya M.C. Araújo: Writing – review & editing, Methodology, Investigation. Bruno B. Castrillo: Writing – review & editing, Methodology, Investigation. Paula Z.M. Dias: Writing – review & editing, Methodology, Investigation. Regina E. Albano: Writing – review & editing, Methodology, Investigation. Carla C. Nascimento: Writing – review & editing, Methodology, Investigation. Paulo R. Filgueiras: Writing – review & editing, Supervision, Methodology, Investigation, Funding acquisition. Valerio G. Barauna: Writing – review & editing, Supervision, Software, Resources, Project administration, Methodology, Investigation, Funding acquisition. Lívia C.M. Rodrigues: Writing – review & editing, Supervision, Software, Resources, Project administration, Funding acquisition.

Funding

This work was supported by the National Council for Scientific and Technological Development (CNPq) [grant number: 409700/2022-3, 310349/2021, 401870/2020]; CAPES [grant numbers 88887.487966/2020-00] and FAPES [grant numbers: 032/2023, 1036/2022, 003/2020, 691/2022, 300/2021, 113/2021, 442/2021, 492/2021, 165/2021].

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Reports a relationship with that includes:. Has patent pending to. If there are other authors, they declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Abbreviations:

ACC -

Accuracy

airPLS -

adaptive iteratively reweighted Penalized Least Squares algorithm

ATR -

Attenuated Total Reflectance

ATR-FTIR -

Attenuated Total Reflectance-Fourier Transform Infrared Spectroscopy

CG –

Control Group

CIS -

Clinically Isolated Syndrome

CNS -

Central Nervous System

CSF -

Cerebrospinal Fluid

EDSS -

Expanded Disability Status Scale

FNR -

False Negative Rate

FPR -

False Positive Rate

FTIR -

Fourier Transform Infrared Spectroscopy

MS -

Multiple Sclerosis

PCA –

Principal Component Analysis

PLS -

Partial Least Squares

PLS-DA -

Partial Least Squares Discriminant Analysis

SENS –

Sensitivity

SPEC -

Specificity

TP -

True Positive

TN -

True Negative

FP -

False Positive

FN -

False Negative

URF-PCoA-RF -

Unsupervised Random Forest with Principal Coordinate Analysis and Supervised Random Forest

VIP -

Variable Importance in Projection

Appendix A. Supplementary data

The following is the Supplementary data to this article:

Data availability

Data will be made available on request.