ATR-FTIR Spectroscopy for Detection of Anti-interferon-Gamma Autoantibodies

Abstract

Introduction: Adult-onset immunodeficiency (AOID) unrelated to HIV is commonly linked to anti-interferon-gamma autoantibodies (AIGAs), predisposing patients to life-threatening opportunistic infections. Current AIGA detection methods are time-consuming and resource intensive. This study aimed to develop and validate an attenuated total reflectance-Fourier transform infrared (ATR-FTIR) technique for detecting and differentiating AIGA-positive patients. Methods: We recruited individuals into three groups: AIGA-positive patients (n = 45), AIGA-negative patients (n = 15), and healthy controls (n = 15). Heparinized plasma or serum samples underwent ATR-FTIR spectroscopy. Spectral profiles were examined using principal component analysis (PCA) and partial least-squares-discriminant analysis (PLS-DA). The reference standard for AIGA concentration was determined by the inhibitory enzyme-linked immunosorbent assay (ELISA). Results: ATR-FTIR spectra revealed significant differences, especially in the amide and fingerprint regions, particularly at key wavenumbers 1632 cm^–1 and 1535 cm^–1. The Amide I/Amide I + II + III ratio is significant between AIGA-positive and others and a secondary structure protein conformational change α-helix to β-sheet. PCA and PLS-DA enabled effective separation of patient groups; PLS-DA and k-NN achieved 95% accuracy, 93.3% sensitivity, and 100% specificity for differentiating AIGA-positive from negative/healthy individuals (NPV 83.3%; PPV 100%). The amide region (1700–1500 cm^‑1) delivered the best prediction accuracy. Conclusion: ATR-FTIR spectroscopy demonstrates significant potential as a rapid and effective diagnostic tool for differentiating AIGA-positive patients from healthy individuals, providing a valuable addition to clinical diagnostic practices. Further validation in larger, diverse populations is warranted.

Introduction

Adult-onset immunodeficiency (AOID) patients who are negative for human immunodeficiency virus (HIV) typically exhibit anti-interferon-gamma (IFN-γ) autoantibodies or AIGAs. AOID patients are often susceptible to microbial infections, particularly intracellular pathogens. − The presence of neutralizing AIGA fundamentally disrupts cellular immunity by preventing IFN-γ from engaging with its receptors. The resulting immunodeficiency manifests severely impaired intracellular bacterial clearance, rendering patients particularly susceptible to infections with opportunistic organisms such as nontuberculous mycobacteria (NTM). −

Previous research has demonstrated the utility of AIGA levels as a laboratory parameter for diagnosing, monitoring, and predicting disease reactivity in AOID patients. Inhibitory enzyme-linked immunosorbent assay (ELISA) has been successfully validated as an appropriate methodology. However, to conduct this specialized test, a minimum number of samples is required for batch processing, causing a prolonged turnaround time.

Infrared spectroscopy is a vibrational technique that obtains the infrared spectrum of tested samples. A unique fingerprint is generated based on the chemical composition of the biomedical sample. This approach provides results rapidly, requires minimal sample volume, and is suitable for integration into a medical laboratory. Analysis of the chemical fingerprint using validated algorithms can inform clinical diagnosis. The integration of the attenuated total reflectance (ATR) and Fourier-transform infrared (FTIR) spectroscopy, also known as ATR-FTIR spectroscopy, enhances data resolution. The total reflectance signal from a crystal module with a high reflective index improves classification performance. Vibrational techniques combined with machine learning algorithms have previously demonstrated excellent efficacy in distinguishing disease states from healthy individuals. ,, FTIR analysis has been combined with various computational frameworks such as principal component analysis (PCA), partial least-squares-discriminant analysis (PLS-DA), and other multivariate methods. These approaches have been used to analyze molecular fingerprints in patients with rheumatoid arthritis, cholangiocarcinoma, and ovarian cancer. ,, With significant advantages and supportive evidence, the ATR-FTIR technique is a promising tool to further develop for the detection of plasma AIGA.

This study aimed to develop ATR-FTIR with a validated analytical pipeline to distinguish patients with AIGA from patients without autoantibodies and healthy individuals. We examined the detection performance of our ATR-FTIR technique for differentiating between AIGA-positive and AIGA-negative lymphadenitis patients. We hope that our results will highlight the appropriateness of ATR-FTIR for screening patients with AIGA-positive infections from other diseases.

Results

No Significant Spectral Differences Observed between Heparinized Plasma and Serum Using ATR-FTIR Spectroscopy

To investigate potential differences in ATR-FTIR spectroscopic profiles between heparinized plasma and the corresponding sera, we analyzed samples from healthy individuals (n = 5 for each sample type). Our analysis revealed that the average spectral patterns obtained from heparinized plasma and serum samples were virtually identical (Figure ).

1.

Comparison of heparinized plasma and serum samples from the same healthy individuals. Average spectra from ATR-FTIR spectroscopic analysis of heparinized plasma (blue) and serum (yellow), (a) showing the fingerprint region (1800–800 cm^–1) and (b) CH-stretching region (3000–2800 cm^–1). PCA data clustering between heparinized plasma and serum, (c) depicted with PC-1 and PC-2 scores from the fingerprint region and (d) CH-stretching region.

ATR-FTIR Spectroscopic Analysis Reveals Distinct Spectral Characteristics in AIGA Patients Compared to Healthy Controls

To explore the spectral characteristics of samples from patients with AIGA, we initially analyzed heparinized plasma from AIGA patients (n = 10) and healthy controls (n = 10). Significant differences in the average spectra within the fingerprint region (1800–800 cm^–1) were observed between the two groups (Figure ). Notably, nine differential peak positions were identified: 1632, 1535, 1451, 1395, 1310, 1233, 1168, 1073, and 1033 cm^–1 (Figure a). These peaks correspond to various biochemical components, summarized in Table , with three significant peak intensities: 1632 cm^–1, 1073 cm^–1, and 1033 cm^–1.

2.

Comparison of ATR-FTIR spectra from patients and healthy controls. Average spectra from 10 AIGA-positive patients (orange) were compared to 10 healthy controls (blue). The experiments were performed in triplicate in each sample. (a) Significant peaks are labeled in the fingerprint region (1800–800 cm^–1). (b) Average second derivative spectra of patients with AIGA (orange line) versus healthy controls (blue line) in the fingerprint region. (c) Data clustering in the fingerprint region, and (d) distinction in the amide region depicted by PCA with a dot plot from each replicate result.

1. Functional Groups of Differential Peak Areas between AIGA-Positive Patients and Healthy Controls .

		Average absorbance (a.u.) ± SEM
Peak (cm–1)	Molecular vibration	Patients	Healthy controls	P-value	References
1632	β-sheet structure of amide	0.423 ± 0.010	0.388 ± 0.012	0.042
1535	Amide II (the N–H bending vibration)	0.391 ± 0.010	0.391 ± 0.012	0.969
1451	Amino acid, lipids (CH3 bend)	0.169 ± 0.003	0.159 ± 0.005	0.129
1395	Amino acids (−COO symmetric stretch and Protein–CH3 group)	0.192 ± 0.004	0.179 ± 0.007	0.153
1310	Amides (C–N stretching), Amide III	0.177 ± 0.004	0.175 ± 0.005	0.718
1233	Antisymmetric phosphate stretch	0.129 ± 0.002	0.122 ± 0.004	0.203
1168	C–O stretch of tyrosine	0.072 ± 0.001	0.077 ± 0.002	0.103
1073	C–O, deoxyribose/ribose DNA, RNA	0.169 ± 0.005	0.147 ± 0.004	0.002
1033	Collagen and C–O stretch of β-anomer	0.158 ± 0.005	0.130 ± 0.005	0.001

^aa.u.: arbitrary unit, SEM: standard error of the mean. Comparisons reaching statistical significance are underlined. P-value <0.05 is considered statistically significant.

Comparison of the second derivative spectra in the fingerprint region revealed prominent IR peaks with higher intensities in AIGA patients (Figure b). Key peaks included 1632, 1513 (amide II), 1451, 1395, 1310, 1233, 1073, and 1033 cm^–1. In contrast, healthy individuals exhibited distinct peak patterns at 1649 cm^–1 (amide I, α-helical protein structures), 1535 cm^–1, and 1168 cm^–1.

Principal component analysis (PCA) of the second derivative spectra in the fingerprint region effectively distinguished patients with AIGA from healthy controls, with PC-1 and PC-2 explaining 77% and 11% of the variance, respectively (Figure c). Furthermore, an analysis of differential peak areas in the amide region (1700–1500 cm^–1 and 1350–1240 cm^–1) associated with immunoglobulins and proteins demonstrated distinct clustering between the groups, with PC-1 explaining 86% of the variance and PC-2 only 7% (Figure d).

ATR-FTIR Spectroscopy Distinguishes AIGA Patients from Patients without AIGA and Healthy Controls

The sample spectra were divided into training and testing data sets at a 70:30 ratio, as outlined in Figure S1. The training data set was employed to develop classification models aimed at differentiating between the groups. To further elucidate the distinctions between groups, we examined 30 patients with AIGA compared with 10 healthy controls and 10 other diseases. The median age of AIGA-positive individuals was 57 years, compared to 51 years for AIGA-negative individuals and 43 years for healthy controls, showing a statistically significant difference (P = 0.003) (Table ). Gender distribution did not differ significantly among groups (P = 0.921; Table ).

2. General Demographic Data of Participants in the Training Set .

	Patients (n = 40)
	AIGA POS	AIGA NEG
Characteristics	(n = 30)	(n = 10)	Healthy Control (n = 10)	P-value
Age in years, median (IQR)	57	51	43	0.003,
	(48–64)	(48–56)	(38–53)
Females, n (%)	12 (40%)	3 (30%)	4 (40%)	0.921
Clinical diagnosis, n (%)
Suspected AOID	28	1	0	<0.001,
	(93.3%)	(10%)
Lymphadenitis	25	4	0	0.014,
	(83.3%)	(40%)
Reactive skin diseases	13	1	0	0.070,
	(43.3%)	(10%)
Malignancy	0	2 (20%)	0	0.058,
Postinfectious bronchiectasis	0	1 (10%)	0	0.250,
Meningitis	0	1 (10%)	0	0.250,
Positive AIGA titer, n (%)	30	0	0	<0.001,
	(100%)
Titer at 1000–5000	5	0	0	ND
	(16.7%)
Titer at 10,000–50,000	15	0	0	ND
	(50.0%)
Titer at ≥100,000	10	0	0	ND
	(33.3%)

^aAIGA POS: AIGA-positive group, AIGA NEG: AIGA-negative group, IQR: interquartile range, ND: not determined.

^bStatistical test by one-way ANOVA.

^cStatistical test by Fisher’s exact test.

^dComparison only between the AIGA-positive versusthe healthy control group.

^eComparison only between the AIGA-positive versus the negative group.

Significantly, 93.3% (n = 28) of AIGA-positive cases in the training set were suspected as AOID, compared to only one case in the AIGA-negative group (P < 0.001) (Table ). The prevalence of lymphadenitis was higher in AIGA-positive individuals (83.3%) compared to AIGA-negative individuals statistically (P = 0.014). The incidence of reactive skin lesions was 43.3% in AIGA-positive individuals. Other AIGA-negative cases were diagnosed as malignancy, postinfectious bronchiectasis, and meningitis (Table ).

The PCA of the second derivative spectra in the fingerprint region distinctly separates patients with AIGA from both healthy controls and patients without AIGA (Figure ). The variance proportion explained by PC-1 was 60%, with PC-2 accounting for 15% (Figure a). Additionally, we selected the amide region to decrease the influence of the nonprotein chemical structure while providing PCA of the second derivative spectra in only the amide region, showing that PC-1 explains 80% of the variance, while PC-2 accounts for 9% (Figure b). The loading plot of the fingerprint region revealed differential spectra, consistent with those presented in Table . Notably, peaks at 1632 cm^–1, 1395 cm^–1, 1310 cm^–1, 1233 cm^–1, and 1033 cm^–1 showed positive regression coefficients, while peaks at 1535 cm^–1, 1168 cm^–1, and 1073 cm^–1 exhibited negative regression coefficients (Figure c). Additionally, when analyzing only the amide region, the loading plot identified a positive regression coefficient at 1649 cm^–1 (amide I α-helical protein structures), along with negative regression coefficients at 1632 cm^–1 and 1233 cm^–1 (Figure d).

3.

ATR-FTIR spectra analysis in the fingerprint and selected amide regions from each sample group. Principal component analysis of results (a) fingerprint region and (b) amide region from patients with AIGA spectra (orange), patients without AIGA spectra (purple), and healthy control spectra (blue). Regression coefficients analysis from spectra loading from the (c) fingerprint region and (d) amide region.

Patients with AIGA Show Altered Protein Content in the Amide I Region, Exhibit Changes in Secondary Protein Structure, and Have Differences in Immunoglobulin G Levels Compared to Patients without AIGA

Our PCA loading plot of the amide I region revealed distinct peaks at 1649 cm^–1 and 1632 cm^–1 (Figure d). To further assess these changes, we compared the absorbances from the amide I region (1700–1600 cm^–1) to those from the wider amide I + II + III region (1700–1000 cm^–1) calculating the absorbance area ratio A _amide I /A _{amide I+II+III}. The median ratio in AIGA-positive patients was 0.90 (IQR, 0.86–0.99), which was significantly lower than the ratios found in the AIGA-negative group (median = 1.07, IQR = 1.02–1.09, P = 0.0001) and healthy controls (median = 1.04, IQR = 1.01–1.08, P = 0.0007) (Figure a). This finding suggests a substantial redistribution of the protein components. Furthermore, analysis of the IgG-associated region (1560–1464 cm^–1) demonstrated a marked increase in absorbance intensity in AIGA-positive patients (P < 0.0001) (Figure b).

4.

Comparison of protein content in the amide and IgG spectral regions among sample groups. (a) Ratio of the Amide I peak area relative to the total area of Amide I, II, and III. (b) Absorbance intensity in the IgG-associated region (1560–1464 cm^‑1). Data are presented as scatter plots with the median and interquartile range (IQR) for healthy control (HC), AIGA-negative (NEG), and AIGA-positive (POS) groups. Statistical significance was determined using one-way ANOVA with a Kruskal–Wallis test. (c, d) Representative deconvoluted Amide I spectra from an AIGA-negative and an AIGA-positive patient, respectively.

To elucidate the specific changes in protein secondary structure, we examined the second derivative spectra of the Amide I region including β-turns (1689–1660 cm^–1), α-helices (1660–1650 cm^–1), β-sheets (1639–1620 cm^–1), and intermolecular β-sheets (1619–1610 cm^–1). We then applied a Gaussian deconvolution method to the spectra to quantify the contributions of these structures. Representative deconvoluted spectra from AIGA-negative and AIGA-positive patients are presented in Figure c and d, visually demonstrating the conformational shifts associated with the disease.

Development and Validation of a Predictive Model for AIGA Using Partial Least-Squares-Discriminant Analysis (PLS-DA)

To develop a predictive model for AIGA, we applied PLS-DA to distinguish between patients with AIGA and those without. The model was trained on a calibration set consisting of 30 AIGA-positive and 10 AIGA-negative patients, utilizing spectral data from the amide region. The PLS-DA score plot demonstrates a clear separation between the groups, with Factor-1 accounting for 87% of the total variance (Figure a). The analysis excluded the 1800–1700 cm^–1 range to minimize noise, in accordance with previous studies. ,

5.

PLS-DA model for discriminating AIGA-positive from AIGA-negative patients using ATR-FTIR spectra. (a) Score plot based on the amide-region spectra showing the separation between AIGA-positive and AIGA-negative patients. (b) Regression coefficients indicate the wavenumbers that contribute most to class separation. (c) Predicted classes for patients in the test set; the misclassified sample is marked with an asterisk (*).

The effectiveness of the prediction model was then evaluated using independent samples in the test set (Figure S1). Characteristics of those samples are listed in Table . In the test set, the PLS-DA model correctly identified 14 AIGA-positive and all 5 AIGA-negative patients, with only one false negative (Figure b).

3. General Demographic Data of Participants in the Test Set .

	Patients (n = 20)
Characteristics	AIGA POS (n = 15)	AIGA NEG (n = 5)	Healthy Control(n = 5)	P-value
Age in years, median (IQR)	53 (46–63)	30 (26–51)	48 (32–54)	0.071
Females, n (%)	5 (26.7%)	1 (20%)	2 (40%)	>0.999
Clinical diagnosis, n (%)
Suspected AOID	15 (100%)	1 (20%)	-	0.001,
Lymphadenitis	7 (46.7%)	0	-	0.114,
Reactive skin diseases	5 (33.3%)	0	-	0.266,
Cryptococcal meningitis	0	1 (20%)	-	0.250,
Disseminated cryptococcosis	0	1 (20%)	-	0.250,
Drug allergy	0	1 (20%)	-	0.250,
Unidentified	0	1 (20%)	-	0.250,
Positive AIGA titer, n (%)	15 (100%)	0	0	<0.001,
Titer at 1000–5000	5 (33.3%)	0	0	ND
Titer at 10,000–50,000	5 (33.3%)	0	0	ND
Titer at ≥100,000	5 (33.3%)	0	0	ND

^aAIGA POS: AIGA-positive group, AIGA NEG: AIGA-negative group, IQR: interquartile range, ND: not determined.

^bStatistical tests were conducted by one-way ANOVA.

^cStatistical tests were conducted by Fisher’s exact test.

^dComparison only between the AIGA-positive versus the negative group.

This resulted in an overall accuracy of 95%, sensitivity of 98%, specificity of 100%, positive predictive value (PPV) of 100%, and negative predictive value (NPV) of 83.3%, as illustrated in the confusion matrices (Figure a and b). Among the machine learning approaches evaluated, PLS-DA and k-nearest neighbors (kNN) demonstrated the highest classification performance (Figure b).

6.

Classification performance of PLS-DA and other machine-learning models. (a) Confusion matrix for the PLS-DA model on the test set (five AIGA-NEG and 15 AIGA-POS patients). (b) Accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) with 95% confidence intervals (Wilson–Brown method) for each classifier.

Discussion

An effective diagnostic strategy is central to the clinical management of AOID due to the presence of AIGAs, given the heightened susceptibility to opportunistic infections in these patients. Our findings indicate that ATR-FTIR spectroscopy is a practical, rapid, and low-sample-volume screening approach for detecting AIGAs. ATR-FTIR has been successfully applied to biomarker analysis in several diseases, including cholangiocarcinoma, rheumatoid arthritis, and multiple sclerosis. , In this context, ATR-FTIR can function as a frontline screening tool to prioritize samples for confirmatory testing.

Importantly, ATR-FTIR should not be used as a standalone diagnostic method. Integration with established confirmatory assays remains essential within a comprehensive diagnostic workflow. Our analysis noted an age difference between healthy volunteers and patient groups; however, this did not materially influence the outcomes. The analytical objective was to maximize the comparability between AIGA-positive and AIGA-negative patients to ensure robust group matching and interpretation.

FTIR spectroscopy is widely used to discriminate between healthy and diseased states and to identify putative spectral biomarkers. ,,,,,,− In this study, spectra from AIGA-positive patients differed significantly from those of the healthy controls. To strengthen clinical relevance, we further included an AIGA-negative cohort with similar clinical presentations. ,, This addition is critical as reliance solely on healthy controls is insufficient in AOID research. AIGA-negative patients often exhibit overlapping symptomatology with AIGA-positive patients. ,,,, Using AIGA-negative patients as comparators therefore provides a more rigorous framework for interpreting disease-specific spectral signatures. ,,,,

We also evaluated potential preanalytical effects related to the sample type. Despite prior reports suggesting anticoagulant-associated spectral variation, , spectra from heparinized plasma and serum in our data set showed closely similar peak positions, intensities, and overall profiles across the measured range. This finding indicates that sample type had minimal impact on ATR-FTIR outcomes. ,

ATR-FTIR has previously been reported as a valuable screening tool, particularly when used in conjunction with standard methods such as ELISA for detecting AIGAs. , Though ELISA is an effective assay, challenges remain regarding its relatively high costs and extended turnaround time. , In contrast, leveraging ATR-FTIR in combination with machine learning provides a more cost-efficient and rapid screening alternative. PCA distinguished AIGA-positive patients from both healthy and AIGA-negative groups and highlighted discriminating spectral regions, especially within the amide bands. The amide I/amide I + II + III absorbance ratio was significantly lower in AIGA-positive patients, indicating an altered protein composition. Within the Amide I region, features associated with the secondary structure suggested a shift from predominantly α-helix to increased β-sheet content among AIGA-positive individuals. This observation aligns with reports that healthy plasma typically exhibits a higher α-helix proportion. , In addition, increased absorbance in the 1560–1464 cm^–1 range, attributed to IgG, , was observed in AIGA-positive patients relative to both comparison groups. These approaches have precedents and collectively support protein-level alterations associated with AIGAs.

Supervised learning models, particularly PLS-DA, successfully discriminated AIGA-positive patients from AIGA-negative patients and outperformed unsupervised methods. This underscores their utility for differentiating patients with overlapping clinical features. Nonetheless, ATR-FTIR has its limitations. It is sensitive to metabolic background and environmental factors (e.g., crystal temperature and pressure), which may introduce artifacts. While similar FTIR-based diagnostic frameworks have been reported in other diseases, our study is the first to pair ATR-FTIR with multivariate modeling for AIGA detection, with encouraging results. Further optimization will be necessary to improve the sensitivity and specificity. Priorities include advanced preprocessing, refined feature selection, rigorous external validation, and multicenter calibration, especially when distinguishing among clinically similar AIGA-negative conditions.

In conclusion, ATR-FTIR is a nondestructive, label-free, and rapid platform that requires minimal sample volume. − , It shows promise as a screening tool that can complement the confirmatory immunoassays. The consistency of results across analyses supports the robustness of our approach and the feasibility of identifying AIGA-positive cases using spectral signatures. Future work should focus on broader validation, algorithmic refinement, longitudinal monitoring, and extension to related autoimmune disorders.

Materials and Methods

Samples and Definitions

Ethics approval for the research protocol was reviewed by the Institutional Review Board of Khon Kaen University (HE671128) in compliance with the Declaration of Helsinki. We used leftover heparinized plasma or serum samples from 60 patients made available by the routine laboratory service of Srinagarind Hospital, Khon Kaen University, Thailand. The samples were collected from 25th September 2018 to 24th May 2022. Medical records were available for these patients. All samples were kept frozen at −80 °C until the day of use. The Institutional Review Board granted a waiver of informed consent because the study used only anonymized leftover specimens from routine clinical testing, involved no additional interventions or sample collection, and posed minimal risk to participants.

Heparinized plasma samples were collected from patients undergoing routine AIGA detection tests. These samples were categorized based on their test results into two groups: AIGA positive (AIGA POS, n = 45) and AIGA negative (AIGA NEG, n = 15). Leftover samples from healthy control persons (HC, n = 15), with no signs of infection, who had attended the routine laboratory service of the hospital for a health check were also retrieved. Excluded from the study were samples from patients with a history of systemic lupus erythematosus (SLE) or rheumatoid arthritis, those suspected of HIV infection, those with a history of immunosuppressive agent usage not related to the treatment of AIGA-related diseases, pregnant women, persons with chronic diseases such as cancer or chronic kidney disease and those under 18 years old were excluded, as AOID primarily occurs in adults. ,,,

AIGA Titer Quantification by Inhibitory ELISA

The inhibitory ELISA was used to measure the AIGA level to be a reference standard for ATR-FTIR analysis. The protocol has been previously published. , Briefly, heparinized plasma samples from patients underwent serial dilution and were then mixed with recombinant human IFN-γ (BD Biosciences, US) at a final concentration of 300 pg/mL. Following incubation for an hour at 37 °C, the concentration of unbound IFN-γ was measured using a human IFN-γ ELISA kit (BD OptEIA; BD Biosciences, US) as per the manufacturer’s instructions.

The titer of the AIGA was determined based on the highest plasma dilution showing greater than 50% inhibition of the expected IFN-γ concentration.

AIGA Detection by ATR-FTIR Spectroscopy

Plasma or serum samples were thawed and 5 μL was applied to the ATR crystal surface. After allowing the sample to air-dry for 5 min, analysis was performed using a Tensor II spectrometer (Bruker Optics, Germany) at 25 °C. The ATR crystal was cleaned using Kimwipes paper moistened with distilled water and methanol after each sample had been processed.

Spectral Acquisition from ATR-FTIR Spectroscopy

Background and sample spectra were obtained from 64 scans at wavenumbers in the 4000–400 cm^–1 range with a resolution of 4 cm^–1. Samples were analyzed in triplicate to ensure acceptable reproducibility. Spectral acquisition was performed through OPUS 8.2.28 software (Bruker Optics, Germany).

Analysis of ATR-FTIR Spectra

Unsupervised PCA was performed using Unscrambler XI (Camo Software, Norway). Spectra of each sample were sharpened using a second derivative (d2λ/dλ2) calculation with three-order polynomials and 17 smoothing points using the Savitzky–Golay algorithm to enhance the resolution of overlapping bands while controlling noise. These parameters were chosen based on prior ATR-FTIR biofluid studies and our preliminary tests, which indicated that they provided a good balance between noise reduction and preservation of peak shapes. Normalization was then performed by using vector unit normalization. The processed data were submitted to the analytical pipeline at two major spectral ranges; 3000–2800 cm^–1 (lipid region) and 1800–800 cm^–1 (fingerprint region). Unsupervised PCA was applied to model the association across all cohorts. PLS-DA (The Unscrambler XI, Camo Software) was supervised by integer coefficients, which assigned “1” as HC, “–1” as AIGA POS, and “0” as AIGA NEG.

Multivariable analysis such as Support Vector Machine (SVM), Random Forest (RF), Neural Network (NN), and k-Nearest Neighbor (kNN) was performed using Quasar version 1.11.1 (University of Ljubljana, Slovenia).

Model Validation

The confusion matrix was calculated. All spectra were randomly split into a training set (70%) and an independent test set (30%) using partial least-squares-discriminant analysis (PLS-DA). Within the training set, we used k-fold cross-validation (k = 10) to tune model hyperparameters and prevent overfitting. Final performance metrics were calculated for the independent test set.

Statistical Analysis

Statistical analysis was performed using Prism GraphPad version 10.4.1 (GraphPad). Differences between parametric variables were analyzed with t-tests while nonparametric variables were analyzed using Mann–Whitney U tests. Differences between multiple sample groups were analyzed with one-way analysis of variance (ANOVA) and differences of categorical data were tested using Fisher’s exact test. The level of statistical significance was set at P < 0.05. The accuracy, sensitivity, specificity, and negative and positive predictive values were also obtained using Prism GraphPad version 10.4.1 (GraphPad).

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acsomega.5c11499.

• Figure S1. Workflow for classification model development and testing; Spectra from the three groups were divided into training (70%) and testing (30%) data sets; The training data set was used to develop classification models, while the testing data set assessed these models’ performance (PDF)

Conceptualization and supervision: A.N. and M.W. Formal analysis and visualization: C.H. Funding acquisition: G.L. and A.N. Investigation, methodology, and validation: V.M., C.H., and A.N. Writingoriginal draft: C.H., P.S., and A.N. Writingreview and editing: all authors.

The authors declare no competing financial interest.