Submit to this Journal Review for this Journal Propose a Special Issue

Article Menu

Open AccessArticle

Peer-Review Record

High-Wavenumber Infrared Spectroscopy of Blood Plasma for Pre-Eclampsia Detection with Machine Learning

Photonics 2024, 11(10), 937; https://doi.org/10.3390/photonics11100937

by Gabriela Reganin Monteiro^1,*, Sara Maria Santos Dias da Silva¹, Jaqueline Maria Brandão Rizzato¹, Simone de Lima Silva¹, Sheila Cavalca Cortelli¹, Rodrigo Augusto Silva¹

, Marcelo Saito Nogueira^2,†

and Luis Felipe das Chagas e Silva de Carvalho^1,†

Reviewer 1:

Johannes Dominikus Pallua

Reviewer 2:

Ruslan E. Asfin

Reviewer 3: Anonymous

Photonics 2024, 11(10), 937; https://doi.org/10.3390/photonics11100937

Submission received: 14 August 2024 / Revised: 25 September 2024 / Accepted: 26 September 2024 / Published: 5 October 2024

(This article belongs to the Special Issue Clinical Translation of Novel Photonics Technologies: From Fundamental Research to Clinical Practice)

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Editor,

I am submitting my review of the manuscript titled "High wavenumber infrared spectroscopy of blood plasma for pre-eclampsia detection" (photonics-3183452). This study investigates the potential of high-wavenumber FTIR spectroscopy as a diagnostic tool for pre-eclampsia by analyzing biochemical interactions in blood plasma. This research holds significant promise because current detection methods often lack sensitivity and specificity, leading to delayed diagnoses.

The study examined dried blood plasma samples from 33 pregnant women and their newborns, employing machine learning to classify control and pre-eclampsia cases. The results demonstrated moderate sensitivity, specificity, and accuracy. These findings suggest that FTIR spectroscopy, particularly when integrated with other diagnostic methods, could improve diagnostic accuracy by identifying hydrophilic molecular interactions in blood plasma, offering valuable clinical insights.

However, the manuscript requires major revisions to enhance its clarity and comprehensiveness. The following issues should be addressed:

Abstract:

The abstract should mention FTIR peaks' main vibrational modes and biomolecular components.

Introduction:

Include and cite other infrared spectroscopic methods used in clinical practice.
Discuss the advantages and disadvantages of various infrared techniques, comparing FTIR and Raman spectroscopy, to justify why FTIR might be superior.
Provide a clear explanation of high-wavenumber infrared spectroscopy.
Illustrate how machine learning classification is applied to FTIR data analysis with examples.
Mention the clinically established routine methods used for screening in both the introduction and the results section.

Materials and Methods:

For FTIR spectroscopy, include missing details such as background measurement, wavenumber range, spectral resolution, coadded scans, and the software used.
Expand the description of spectral data analysis, particularly the machine learning classification, and consider adding a graphical representation to aid understanding. Machine learning should also be reflected in the title of the manuscript.

Results:

Provide a graphic overview of the workflow to improve clarity.
In Figure 1, ensure that the peaks are marked, and citations should be provided to support the band assignments in Table 1.
Replace the confusion matrices in the tables with illustrative figures.
Incorporate and compare clinically established routine screening methods results with the FTIR findings.

Discussion:

Expand the discussion section to incorporate data from the clinically established routine methods.

Addressing these revisions will significantly improve the manuscript's quality and contribution to the field.

Best regards,

Author Response

For the point-by-point responses including figures and indications of manuscript modifications, we suggest checking the PDF file attached.

Comments to the Author

Abstract:

The abstract should mention FTIR peaks' main vibrational modes and biomolecular components.

Response: We appreciate the reviewer's suggestion to expand the discussion section by incorporating data from clinically established routine methods. We have revised the section to highlight the limitations of traditional diagnostic approaches and the potential of FTIR spectroscopy as a complementary tool.

In terms of clinical diagnostics, pre-eclampsia is a complex systemic disorder with multiple factors involved in its pathogenesis, including hypoxia, shallow placentation, endothelial cell damage, and immunological factors. These factors contribute to the altered levels of various substrates, many of which are in the early stages of understanding their roles in the development of pre-eclampsia. The condition is characterized by hypertension, proteinuria, and/or edema occurring after 20 weeks of gestation. Maternal clinical risk factors include advanced maternal age, previous history of pre-eclampsia, family history of pre-eclampsia, nulliparity, obesity, multifetal gestation, diabetes mellitus, chronic hypertension, and chronic renal disorders. To aid in early diagnosis, various analyses using FTIR have been conducted.

FTIR spectroscopy has potential for pre-eclampsia diagnostics due to its significant advantages in precision and minimal sample preparation. The high-wavenumber FTIR analysis of pregnant women's plasma enables the identification of significant differences in biochemical composition, highlighting biomolecular changes mainly in lipids, bound water, proteins, and nucleic acids. Importantly, while traditional methods such as urinalysis and blood pressure measurements serve as the foundation for diagnosing pre-eclampsia, they often fall short in sensitivity and specificity, especially in early detection.

Paper text modified:

Abstract: Early detection of pre-eclampsia is challenging due to the low sensitivity and specificity of current clinical methods and biomarkers. This study investigates the potential of high-wavenumber FTIR spectroscopy as an innovative diagnostic approach capable of providing comprehensive biochemical insights with minimal sample preparation.

Blood samples were collected from 33 pregnant women and their corresponding 33 newborns during induction or spontaneous labor. By analyzing dried blood plasma samples, we identified biomarkers associated with FTIR vibrational modes, including 2853.6 cm−1 (CH2 stretching in lipids), 2873.0 cm−1 (CH3 stretching in lipids and proteins), and 3279.7 cm−1 (O–H stretching related to water and proteins).

Machine learning classification revealed 76.3% ± 3.5% sensitivity and 56.1% ± 4.4% specificity in distinguishing between pre-eclamptic and non-pre-eclamptic pregnant women, along with 79.0% ± 3.5% sensitivity and 76.9% ± 6.2% specificity for newborns. The overall accuracy for classifying all pregnant women and newborns was 71.8% ± 2.5%.

The results indicate that high-wavenumber FTIR spectroscopy can enhance classification performance when combined with other analytical methods. Our findings suggest that investigating hydrophilic sites may complement plasma analysis in clinical settings.

Introduction:

Include and cite other infrared spectroscopic methods used in clinical practice.

Response: We have taken this into consideration and expanded our discussion to incorporate relevant techniques and their applications.

FTIR spectroscopy provides information on the composition of substances according to their molecular content. This capability has been utilized in different research and topics since the early 1990s to distinguish differences between normal and diseased tissues. Initial studies were conducted for in vivo applications using attenuated total reflectance infrared (ATR-IR) combined with optical fibers. Other infrared spectroscopic methods, such as near-infrared (NIR) and Raman spectroscopy, have also been employed in clinical practice. For example, NIR spectroscopy has been used for non-invasive monitoring of glucose levels in diabetic patients, while Raman spectroscopy has shown promise in cancer diagnostics by enabling the identification of tumor characteristics through tissue analysis.

Recent studies further illustrate the potential of ATR-FTIR spectroscopy in clinical applications. Bunaciu et al. (2013) provide a review of infrared microspectroscopy applications that highlight its utility in differentiating tissue types. Cameron et al. (2020) demonstrated the application of ATR-FTIR serum spectroscopy to stratify brain tumor histological sub-types, indicating the method's relevance in secondary care settings. Additionally, Magalhães et al. (2021) discussed the broader implications of FTIR spectroscopy in biomedical research, emphasizing strategies to maximize its potential for clinical diagnostics.

Paper text modified: FTIR spectroscopy provides information on the composition of substances according to their molecular content. This fact has been used by various groups since the early 1990s to distinguish the difference between normal and diseased tissues. Some initial studies were conducted for in vivo applications using ATR-IR combined with optical fibers.

Bunaciu, A., Fleschin, S., & Aboul-Enein, H. (2013). Infrared Microspectroscopy Applications - Review. Current Analytical Chemistry, 10(1), 132–139. doi:10.2174/1573411011410010011

Cameron, J.M.; Rinaldi, C.; Butler, H.J.; Hegarty, M.G.; Brennan, P.M.; Jenkinson, M.D.; Syed, K.; Ashton, K.M.; Dawson, T.P.; Palmer, D.S.; et al. Stratifying Brain Tumour Histological Sub-Types: The Application of ATR-FTIR Serum Spectroscopy in Secondary Care. Cancers 2020, 12, 1710. https://doi.org/10.3390/cancers12071710

Magalhães, S., Goodfellow, B. J., & Nunes, A. (2021). FTIR spectroscopy in biomedical research: how to get the most out of its potential. Applied Spectroscopy Reviews, 1–39. doi:10.1080/05704928.2021.1946822

Discuss the advantages and disadvantages of various infrared techniques, comparing FTIR and Raman spectroscopy, to justify why FTIR might be superior.

Response: Fourier Transform Infrared (FTIR) spectroscopy offers several advantages over Raman spectroscopy, particularly in the analysis of biological materials. According to Baker et al. (2014), FTIR provides higher sensitivity in detecting biomolecular changes, making it a powerful tool for identifying biomarkers in complex biological samples, such as plasma. One key advantage of FTIR over Raman is its ability to minimize issues related to fluorescence interference, which can significantly complicate the interpretation of Raman spectra. Additionally, FTIR is particularly effective in analyzing functional groups within biomolecules, which is crucial for diagnosing conditions like pre-eclampsia. While Raman spectroscopy is often preferred for its ability to analyze samples with little or no preparation and for its compatibility with aqueous environments, its susceptibility to fluorescence remains a significant drawback. Therefore, when analyzing biological fluids, FTIR might be considered superior due to its higher sensitivity and fewer complications from fluorescence, as highlighted by Baker et al. (2014).

Paper text modified:

3rd paragraph of “1. Introduction” section:

FTIR spectroscopy offers distinct advantages over Raman spectroscopy, particularly in the analysis of biological fluids such as plasma. FTIR provides higher sensitivity in detecting biomolecular changes, with fewer issues related to fluorescence interference that often complicate Raman spectra. This makes FTIR an ideal tool for identifying biomarkers in complex biological samples, which is crucial in the diagnosis of conditions such as pre-eclampsia.

Baker MJ, Trevisan J, Bassan P, Bhargava R, Butler HJ, Dorling KM, et al. Using Fourier transform IR spectroscopy to analyze biological materials. Nature Protocols, 2014. Vol 9(8), page 1771–1791. doi:10.1038/nprot.2014.110

5th paragraph of “4. Discussion” section:

Comparative studies between FTIR and Raman spectroscopy demonstrate that both offer similar practical advantages, particularly in prenatal diagnostics. FTIR spectroscopy is a more mature and cost-effective technology, making equipment compact and easier to translate to clinical practice. The ease of implementation and low operational cost of FTIR reinforce its feasibility as a valuable tool for diagnosing and monitoring complex conditions such as pre-eclampsia.

Provide a clear explanation of high-wavenumber infrared spectroscopy.

Response: Thank you for your comment, we have explained the definition of high-wavenumber infrared spectroscopy and its advantages in the 6th paragraph of the “1. Introduction section”.

Paper text modified: The use of the high wavenumber (HW) region, between 2600 and 3800 cm⁻¹, for the diagnosis of pathologies offers several advantages over the fingerprint region (400 to 1800 cm⁻¹). Recent studies have shown that the HW region contains diagnostic information similar to the fingerprint region, making it effective in discriminating tissue structures with different molecular compositions, as observed in brain and bladder tumors. The use of FTIR spectroscopy in the HW region can be defined as high-wavenumber infrared spectroscopy or HWIR spectroscopy. When investigating biological samples, HWIR spectroscopy captures molecular vibrations of the hydrophilic sites within biomolecules such as lipids, proteins, carbohydrates and nucleic acids. Given the maturity of FTIR as a technology, not only portable equipment can be produced, but also biofluid analysis and sample classification can be achieved in minutes, limited mostly by the sample drying time. Little to no professional training required to use FTIR equipment. In addition, clinical translation of bound water analysis can be facilitated by automatic analysis of the HW region of infrared spectra by using machine learning. Thus, HW spectroscopy presents a more practical and economical alternative for diagnostics that can complement biochemical analysis and potentially introduce new methods for time-sensitive clinical decision-making.

Illustrate how machine learning classification is applied to FTIR data analysis with examples.

Response: Thank you for your insightful suggestion regarding the expansion of the description of the spectral data analysis and the incorporation of machine learning classification in our manuscript. In line with your feedback, we will elaborate further on how machine learning was applied to the FTIR data. Specifically, we will provide more details about the algorithms used, such as Neural Networks and Support Vector Machines (SVM), which have proven effective in previous studies for analyzing FTIR spectra and distinguishing between healthy and diseased samples. These methods not only enhance the diagnostic capabilities by identifying specific biomarkers but also facilitate the simultaneous analysis of multiple metabolites, which is vital for early detection of conditions such as pre-eclampsia.

Additionally, we will include graphical representations, such as flowcharts or figures illustrating the classification process, to improve reader comprehension. We will also revise the manuscript title to reflect the application of machine learning, as this approach is central to the analysis and findings of our study.

Paper text modified: Machine learning classification applied to FTIR spectroscopy data analysis has shown great potential in detecting conditions such as pre-eclampsia by identifying specific biomarkers in biological samples. For example, previous studies have demonstrated that machine learning algorithms, such as Neural Networks and Support Vector Machines (SVM), can be used to analyze FTIR spectra, allowing the distinction between healthy and diseased tissues as well as the identification of biochemical changes in fluids such as plasma. This approach not only improves diagnostic accuracy but also enables the simultaneous analysis of multiple metabolites, which is crucial for the early detection of pathologies. The integration of machine learning with FTIR provides additional information that can complement traditional clinical examinations, contributing to more informed decision-making in clinical practice

Mention the clinically established routine methods used for screening in both the introduction and the results section.

Response: In response to your request for a discussion of clinically established routine methods used for screening in both the introduction and results sections, we have revised these sections to address the use of traditional diagnostic techniques for preeclampsia.

In the introduction, we now highlight the current routine methods, such as blood pressure monitoring and proteinuria analysis, following the guidelines of the American College of Obstetrics and Gynecology (ACOG). While these methods are widely used in clinical practice, we emphasize their limitations, particularly in identifying early or nonproteinuric preeclampsia. This sets the context for our study, which explores the potential of FTIR spectroscopy to provide more sensitive and specific diagnostic insights.

In the results section, we compare the performance of FTIR spectroscopy with these established screening methods. By demonstrating that FTIR, combined with machine learning, achieves a higher classification accuracy, particularly in complex cases where traditional methods might be insufficient, we justify the use of this novel approach. The findings indicate that FTIR could complement or enhance existing screening protocols, offering earlier detection of preeclampsia and improving clinical outcomes.

These additions clarify the relevance of our study within the framework of current clinical practices and underscore the potential impact of FTIR spectroscopy as a diagnostic tool.

Parte superior do formulário

Parte inferior do formulário

Paper text modified: Preeclampsia is traditionally diagnosed based on clinical parameters such as elevated blood pressure (≥140/90 mmHg) and the presence of proteinuria (≥300 mg/24 hours) after 20 weeks of gestation, as recommended by the American College of Obstetrics and Gynecology (ACOG). While these methods are widely accepted and form the backbone of prenatal screening, they have limitations in terms of sensitivity and specificity, particularly in identifying cases of nonproteinuric or early-onset preeclampsia. Additionally, screening approaches that rely solely on clinical markers may fail to detect the complex molecular and biochemical changes associated with preeclampsia, contributing to delayed diagnosis and management. In light of these challenges, there is a growing interest in exploring more advanced diagnostic tools, such as Fourier Transform Infrared (FTIR) spectroscopy, which can detect subtle biomolecular alterations in blood plasma, potentially improving the early detection and differentiation of preeclampsia.

Vidaeff, A. C., Saade, G. R., & Sibai, B. M. (2020). Preeclampsia: The Need for a Biological Definition and Diagnosis. American Journal of Perinatology. doi:10.1055/s-0039-1701023

Section of results:

Our study compared the classification performance of FTIR spectroscopy with traditional clinical screening methods for preeclampsia, such as blood pressure monitoring and proteinuria analysis. The control group and the preeclampsia group were evaluated using standard criteria, with blood pressure and proteinuria levels used to classify pregnant women. In contrast, FTIR spectroscopy, combined with machine learning algorithms, successfully classified blood plasma samples with a mean accuracy of 89.5% across 20 iterations of 5-fold cross-validation. This accuracy surpasses the limitations of traditional screening, particularly in detecting nonproteinuric preeclampsia, where clinical methods may fail. The results indicate that FTIR can serve as a complementary tool to the current methods, providing earlier and more precise biochemical insights into the development of preeclampsia.

The diagnosis of preeclampsia has evolved over the years, with the removal of edema as a required diagnostic criterion in the early 2000s due to its lack of specificity and sensitivity. While proteinuria has often been associated with preeclampsia, approximately 10% of women may present with hypertension and multisystemic signs of the condition, such as new-onset thrombocytopenia, impaired liver function, renal dysfunction, and respiratory or cerebral disturbances, in the absence of proteinuria. This nonproteinuric form of preeclampsia is even more common postpartum. The recognition of the multisystemic nature of preeclampsia has led to the understanding that hypertension is merely one manifestation of endothelial dysfunction, with factors such as capillary leak and vasoconstriction affecting multiple systems, including the kidneys, liver, brain, heart, placenta, and hematological system. Although hypertension retains its central role in current diagnostic criteria, it is the dysfunction in other systems that is often associated with adverse maternal and neonatal outcomes, with hypertension absent in up to 18% of women with HELLP syndrome, one of the most severe forms of preeclampsia.

Reference: Vidaeff AC., Saade GR., Sibai BM. Preeclampsia: the need for a biological definition and diagnosis. American Journal of Perinatology, 2021. Vol 38(09), 976-982.

Materials and Methods:

For FTIR spectroscopy, include missing details such as background measurement, wavenumber range, spectral resolution, coadded scans, and the software used.

Response: Thank you for your valuable feedback. We have incorporated additional details regarding the FTIR spectroscopy methodology in the revised text. The changes provide clarity on background measurements, wavenumber range, spectral resolution, coadded scans, and the software used for data processing.

Paper text modified: In response to your request, we have now added key details regarding our FTIR methodology. FTIR spectra of blood plasma samples were collected using a Bruker Alpha II spectrometer, which was equipped with an ATR-FTIR diamond crystal and heating functions to aid sample drying. Background measurements were performed prior to each sample collection to ensure accuracy. The spectral range was from 4000 to 400 cm⁻¹, and data were acquired with a spectral resolution of 4 cm⁻¹. For each sample, 32 co-added scans were performed to enhance the signal-to-noise ratio.

To minimize cross-contamination, the ATR crystal was sanitized with 70% alcohol after each measurement, and we allowed adequate drying time. The collected spectra were processed using MATLAB scripts, where baseline correction was applied, and spectral data were smoothed using the Savitsky-Golay filter with an 11-point frame window and second polynomial order. Spectral normalization was performed using vector normalization to facilitate comparison between samples.

Expand the description of spectral data analysis, particularly the machine learning classification, and consider adding a graphical representation to aid understanding. Machine learning should also be reflected in the title of the manuscript.

Response: Thank you for your insightful comments regarding the assignment of the bands in the last three rows of Table 1. We have carefully revisited the assignments and would like to offer the following clarification to support the rationale behind our attributions:

Bands at 3456 cm⁻¹ and 3483 cm⁻¹: These bands are assigned to the symmetric OH stretching mode, which aligns with vibrations typically associated with water molecules. The band at 3483 cm⁻¹, in particular, corresponds to OH hydrogen bonds. Given the high concentration of water in biological fluids such as blood plasma, these assignments are expected and well-supported by previous studies.
Bands at 3280 cm⁻¹ and 3288 cm⁻¹: The 3280 cm⁻¹ band reflects O-H stretching in water and N-H stretching in proteins, while the 3288 cm⁻¹ band is specifically attributed to symmetric OH stretching in water. These bands are present across all study groups, indicating the common presence of water and proteins in plasma, consistent with their biochemical roles in these biological samples.
Bands at 3192 cm⁻¹ and 3194 cm⁻¹: These bands represent N-H stretching vibrations, primarily from peptides and proteins with cis-ordered substructures. The slight variation between the 3192 cm⁻¹ (observed in newborns) and 3194 cm⁻¹ (observed in pregnant women) may suggest conformational differences in proteins between these groups.

These assignments are consistent with existing FTIR spectroscopy literature, which associates these spectral regions with proteins, water, and other biomolecules found in blood plasma. Furthermore, the presence of these bands across specific study groups (as outlined) reflects the biochemical differences between control and pre-eclampsia groups, both in newborns and pregnant women, as highlighted in the classification results (Tables 4-9).

Paper text modified:

To ensure clarity and consistency, we presented only the classification performance metrics of the most accurate classifier across all study groups or pair of groups. The most important metrics for our feasibility study were sensitivity, specificity, accuracy and area under the receiver operating characteristic curve (AUC). These metrics were calculated as per the definitions and equations below:

True positive (TP): number of pre-eclampsia pregnants or newborns correctly classified
False positive (FP): healthy pregnants or newborns incorrectly classified
True negative (TN): healthy pregnants or newborns correctly classified
False negative (FN): number of pre-eclampsia pregnants or newborns incorrectly classified

(1)

(2)

(3)

It is worth noting that sensitivity and specificity are only pertinent for classification using two groups (one healthy group and one diseased group). The receiver operating characteristic curve (as known as ROC curve) features the true positive rate (TPR) on the Y axis, and false positive rate (FPR) on the X axis. In the classification using two groups, the TPR equals to sensitivity and FPR equals to (1-specificity). However, in the classification using more than two groups (e.g., our classification including all four study groups referred to in section 2.6), TP and FP considered the positive group as the selected group for calculating the AUC (e.g., AUC_{Control newborn} considered the newborn controls as positive group), while TN and FN included the remaining groups (e.g., when calculating AUC_{Control newborn} for the classification of all four study groups, the negative group included the control pregnants, pre-eclampsia newborns and pre-eclampsia pregnants. Accuracy could still be calculated for any number of study groups because it equals the number of correctly classified spectra (observations of this study) divided by the total number of correctly classified spectra. We reported the mean and standard deviation of these metrics calculated over 20 iterations of 5-fold cross-validation (CV). Each iteration applied random sampling, i.e., the dataset was randomly divided into training and validation sets with 80% and 20% of the total data, respectively. At each iteration, machine-learning model training and validation was repeated five times until all parts of the dataset were tested by using all the five validation sets. Therefore, the mean of 20 iterations reflects the average classification performance, and the standard deviation indicates the stability of the model based on the dataset's distribution. Figure 1 below indicates the steps for machine learning classification and validation based on processed FTIR spectra and their second derivative.

Figure 1. Steps of spectral data analysis and machine learning validation based on our FTIR spectra for all four study groups.

Results:

Provide a graphic overview of the workflow to improve clarity.

Response: Thank you for your comment. We have added the graphic overview of our workflow in section 2.7 of “2. Materials and Methods” section.

Paper text modified:

Figure 1. Steps of spectral data analysis and machine learning validation based on our FTIR spectra for all four study groups.

In Figure 1, ensure that the peaks are marked, and citations should be provided to support the band assignments in Table 1.

Response: Thank you for your suggestion. We have marked in Figure 2 (former Figure 1) the peaks used for the band assignments in Table 1. The marks can be checked in Figures 2A and 2B in the “3. Results” section.

Paper text modified:

Figure 2. Mean FTIR spectra of blood plasma samples of the control and pre-eclampsia groups for A) Pregnant women and B) Newborns after spectral smoothing and vector normalization, as well as the mean of the second derivative of the same FTIR spectra for the C) Pregnant women and D) Newborns. Spectra in A) and B) show the main vibrational modes for the FTIR peak/band assignment according to Table 1.

Replace the confusion matrices in the tables with illustrative figures.

Response: Thank you for your comment. We understand that our text may have too many tables and hence the reviewer suggests condensing the content with illustrative figures. Still, please note that confusion matrices are standard among the clinical diagnostic and machine learning community, as it allows readers to observe which study groups were mistaken by others during sample classification by using our machine learning algorithm. We added confusion matrices in the text for completeness when showing the results of our classification models.

Illustrative figures simplifying the metrics of confusion matrices would only add difficulty to readers of our article. For comparing machine-learning classification between several methods illustrated in the paper, the reader should use classification performance metrics (i.e., results of Tables 2 and 3). These are the only tables that summarize our results. To simplify the results section, we moved tables 4-9 to supplementary material, added a fourth paragraph and the following figures in the “3.Results” section to summarize the results of our classification models (and illustrate the same results of tables 2 and 3):

Paper text modified: Tables 2 and 3 compare the classification performance metrics for 4-class models and 2-class models using either processed FTIR spectra or their second derivative. The overall most accurate model for all study groups. For the classification of all groups (i.e., 4-class model, Table 2 and Figure 3), the highest accuracy was observed when using the processed FTIR spectra instead of its second derivative. A similar behavior was observed for 2-class models only including the control and pre-eclampsia groups for pregnants and newborns separately (Table 3, Figures 4 and 5).

Table 2. Comparison between classification performance metrics achieved for 4-class models using either processed FTIR spectra or their second derivative. Please note that sensitivity and specificity cannot be defined when two control groups are present.

Classification performance metric	FTIR spectra	Second derivative of FTIR spectra
Accuracy	(71.8 ± 2.5)%	(63.8 ± 2.0)%
AUC_{Control newborn}	0.918 ± 0.013	0.923 ± 0.008
AUC_{Control pregnant}	0.901 ± 0.009	0.886 ± 0.009
AUC_{Preclampsia newborn}	0.933 ± 0.016	0.864 ± 0.015
AUC_{Preclampsia pregnant}	0.837 ± 0.020	0.810± 0.013

AUC: area under curve

Figure 3. Classification performance metrics achieved for the 4-class models using either processed FTIR spectra (solid blue bar) or their second derivative (striped orange bar).

Table 3. Comparison between classification performance metrics achieved for 2-class models using either processed FTIR spectra or their second derivative. Please note that sensitivity and specificity cannot be defined when two control groups are present.

	Pregnant		Newborn
Spectral type	FTIR spectra	Second derivative of FTIR spectra	FTIR spectra	Second derivative of FTIR spectra
Classification performance metric
Sensitivity (%)	(76.3 ± 3.5)%	(63.7 ± 5.9)%	(79.0 ± 3.5)%	(74.5 ± 4.2)%
Specificity (%)	(56.1 ± 4.4)%	(54.1 ± 5.1)%	(76.9 ± 6.2)%	(68.8 ± 4.0)%
Accuracy (%)	(66.3 ± 2.9)%	(58.9 ± 3.5)%	(78.0 ± 3.8)%	(71.6 ± 2.8)%
	0.692 ± 0.036	0.613 ± 0.032	0.83 ± 0.04	0.792 ± 0.025
	0.692 ± 0.036	0.613 ± 0.032	0.83 ± 0.04	0.792 ± 0.025

Figure 4. Classification performance metrics achieved for the 2-class model using either processed FTIR spectra from pregnants (solid blue bar) or the second derivative of the same spectra (striped orange bar).

Figure 5. Classification performance metrics achieved for the 2-class model using either processed FTIR spectra from newborns (solid blue bar) or the second derivative of the same spectra (striped orange bar).

Incorporate and compare clinically established routine screening methods results with the FTIR findings.

Response: In response to the comment to incorporate and compare clinically established routine screening methods with our FTIR findings, we have expanded the discussion to address this. Clinically established methods for pre-eclampsia screening, such as measuring blood pressure, proteinuria, and assessing levels of biomarkers like placental growth factor (PlGF) and soluble fms-like tyrosine kinase-1 (sFlt-1), have been useful in routine diagnostics. However, these methods are often limited by their sensitivity and specificity, as the results can vary due to symptom overlap with other conditions such as gestational hypertension, chronic hypertension, and renal diseases.

Our study shows that FTIR spectroscopy, especially in the high-wavenumber region, offers an advantage by providing a rapid and detailed molecular fingerprint of blood plasma, which allows for the identification of biochemical changes linked to pre-eclampsia at an earlier stage. When comparing the accuracy and sensitivity of FTIR-based classification models with traditional methods, FTIR has the potential to complement these established screening techniques. FTIR’s ability to detect subtle biochemical changes, such as shifts in lipid and protein content, gives it an edge in scenarios where routine methods might miss early biochemical markers.

The combination of FTIR with machine learning, as demonstrated in our study, highlights its ability to improve diagnostic accuracy for pre-eclampsia compared to routine methods alone. This suggests that FTIR could be integrated into clinical workflows as an additional or complementary tool, providing more comprehensive and accurate results for patient management and earlier intervention.

Paper text modified: This a pregnancy-specific condition that significantly contributes to maternal and perinatal mortality worldwide. The timely diagnosis of preeclampsia is crucial for implementing treatment; however, a concrete form of diagnosis is still discussed. The first-trimester screening algorithm has been developed and validated to predict preterm preeclampsia. The discussion for predicting term disease, where the majority of cases occur, is poor. Clinically established biomarkers such as soluble fms-like tyrosine kinase-1 (sFlt-1) and placental growth factor (PlGF) are utilized in cases of suspected preterm preeclampsia, providing a high negative predictive value for confidently excluding disease in women with normal results, but their sensitivity remains modest.

Discussion:

Expand the discussion section to incorporate data from the clinically established routine methods.

Response: Thank you for your valuable feedback regarding the expansion of the discussion section to include data from clinically established routine methods. We understand the importance of contextualizing our findings within the framework of current screening practices for pre-eclampsia to better highlight the contributions and potential benefits of our approach.

In response to your suggestion, we have expanded the discussion to provide a more in-depth comparison with established methods such as blood pressure monitoring, proteinuria testing, and the use of biomarkers like PlGF and sFlt-1. We specifically discuss how FTIR spectroscopy combined with machine learning algorithms can improve sensitivity and specificity, particularly in detecting early biochemical changes that might not be captured by these traditional methods.

Furthermore, we have added a discussion on how our method can integrate into existing clinical workflows, complementing these routine screening practices by providing additional molecular insights that could help improve the early diagnosis and management of pre-eclampsia.

We hope these revisions adequately address your comment, and we remain open to further suggestions if needed.

Paper text modified:

For instance, routine biomarkers, including serum creatinine and uric acid, provide valuable information, but they can miss subtle biochemical changes occurring in the early stages of the disorder. FTIR spectroscopy, with its real-time extraction of digital spectral features, presents a promising complementary approach, allowing for a more comprehensive understanding of metabolic disruptions associated with pre-eclampsia. This method could facilitate early diagnosis and improve prenatal monitoring capabilities, particularly in resource-limited settings due to its minimal sample preparation and the absence of specialized reagents.

To understand the biological processes associated with pre-eclampsia plasma biomarkers, it is crucial to view pre-eclampsia biochemical changes as a placental disease. Placentas from women with pre-eclampsia show increased frequencies of villous infarctions, villous-free placental lakes, inflammation, fibrin deposition, syncytial knots, and abnormal cytotrophoblast proliferation. Additionally, it is associated with changes in genetic expression and DNA methylation in the placenta. Factors released from the placenta, including exosomes, pro-inflammatory cytokines, cell-free fetal DNA, and anti-angiogenic agents, disrupt maternal endothelial function, leading to the multi-systemic clinical syndrome of pre-eclampsia. However, the exact cause of this pathology remains undefined.

Many studies aim to identify biomarkers and maternal characteristics associated with hypertensive pathology through predictive algorithms. Yet, these biomarkers have not been translated into clinical practice, underscoring the potential of vibrational spectroscopy, specifically Raman spectroscopy and FTIR spectroscopy.

To interpret FTIR analysis results, it is essential to understand the metabolic characteristics of pre-eclampsia, which occurs in two phases. The first phase involves inadequate implantation and placentation, leading to poor uteroplacental perfusion, tissue hypoxia, and oxidative stress. This triggers the release of anti-angiogenic factors into the maternal circulation, causing a systemic inflammatory response. In the second phase, these factors induce widespread endothelial dysfunction, which is responsible for the hypertensive syndrome.

Changes in protein concentrations and specific expression patterns in coagulation cascades can favor thrombophilia and inflammation, linking pre-eclampsia to cardiovascular disease development. Identifying peptides associated with pre-eclampsia is crucial for determining cardiovascular alterations that may persist long-term, as changes in the proteome suggest cardiovascular and thrombotic risks in symptomatic and asymptomatic individuals six months post-pre-eclampsia.

Lipid concentration changes during pregnancy are another factor. Normal, uncomplicated pregnancies exhibit a progressive physiological increase in maternal serum lipid concentrations, essential for fetal development. Near the end of pregnancy, fatty acid storage in maternal adipose tissue increases due to physiological insulin resistance. Evidence suggests that pre-eclampsia is associated with dyslipidemia, an imbalance in lipid regulation.

Women with pre-eclampsia have insufficient adipose tissue expansion compared to healthy pregnant women. Additionally, their adipocytes become more insulin-resistant, increasing lipolysis. These processes result in ectopic fat accumulation in the liver and other tissues. Serum triglyceride levels are significantly higher in women with pre-eclampsia compared to those with normal pregnancies. Furthermore, LDL levels increase, HDL levels decrease, and serum free fatty acids rise. These lipid concentration changes are similar to those observed in obese patients.

In the high-wavenumber region, the spectrum is primarily attributed to lipid absorption for bands at 2849, 2917, and 3008 cm^-1, and dominated by water absorption bands at 3350 cm^-1. The spectral region 3050–2800 cm^-1 corresponds to asymmetric CH₃ asymmetric CH₂, symmetric CH₃, and symmetric CH₂ (symCH₂) distribution, observable in the control and pre-eclampsia sample results. This factor is essential for fetal development, especially in the late gestation period, a critical time for pre-eclampsia patients.

Alterations in proteases, enzymes that catalyze the hydrolysis of peptide bonds, can be observed during the pathology's development, disrupting cellular homeostasis. Previous studies reported inconclusive protein levels due to the low sensitivity of ELISA in measuring ultra-low protein levels and inflammatory cytokines with low positive predictive values for accurate pre-eclampsia diagnosis.

According to the absorbance range found for peptides in the high-wavenumber region, the bands between 3280 cm^-1 (H–O–H stretch), 2957 cm^-1 (asymmetric CH₃ stretch), 2920 cm^-1 (asymmetric CH₂ stretch), 2872 cm^-1 (symmetric CH₃ stretch), 1536 cm−1 (amide II of proteins), 1453 cm^-1 (CH₂ scissor), 1394 cm^-1 (C=O stretch of COO^-), 1242 cm-1 (asymmetric PO₂), 1171 cm^-1 (asymmetric C–O ester stretch), and 1080 cm^-1 (C–O stretch) were observed.

These protein patterns are important for diagnosing various pathologies. In pre-eclampsia, studies identifying molecular alterations causing endothelial dysfunction found amino acids, protein secondary structures, lipids, and fatty acids in higher amounts in the blood or urine of women with pre-eclampsia. As guidance for complementary and novel biochemical methods, we have described potential biochemical changes through FTIR band assignment that can be associated with effects on child growth and maternity complications/mortality associated with pre-eclampsia.

In addition, we found that changes in the FTIR high-wavenumber region are more prominent in newborn plasma compared to pregnant plasma, which may suggest that newborns are highly affected at the time of birth. Our future work will comprise new analyses with an increased number of patients and associations with 24-hour urine sample analysis for further correlation with plasma biochemical changes and understanding the origin of such changes.

Addressing these revisions will significantly improve the manuscript's quality and contribution to the field.

Response: Thank you for your review. We have implemented all changes in the manuscript text following your comments and suggestions, improving the content and readability.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The authors of "High wavenumber infrared spectroscopy of blood plasma for pre-eclampsia detection" studied the possibility of using FTIR spectroscopy to diagnose specific diseases. It may be interesting to readers, however some moments should be corrected.

First of all it is using of term 'High wavenumber infrared spectroscopy'. When I've read this term for the first time, I believed it referred the region close to visible light. However authors mean the region 2800-3600 cm⁻¹. I tried to find out how other scientists use this term, but without success. Probably authors introduced this term. I recommend changing it to something more suitable, for example, using term 'middle infrared spectroscopy'.

In Figure 1 one can see that the both curves on B and D panels are rather the same. The some differences are observed in panels A and C, however their significance can only be assessed after showing the standard deviation of this averaged curves. If these deviations are too small to be visible in the Figure, this should be stressed in the text.

Check please dimensions of ordinate axes in the panels B and D. Because it is the second derivative it should include cm², not only absorbance.

In Table one in the first row: what does mean the asterisk after 2853.6?

Isn't the number in the first column too precise? Should you give a cipher after the point? It particular depend on resolution of measurement. This information also should be given in the '2.5 FTIR spectroscopy' section.

It is better to use the symbol 'ν' (Greek nu) instead 'v' in vibration assignments. The symbols 's' and 'as', meaning 'symmetric' and 'asymmetric', are best written using a subscript. For example, ν_as.

In the second row 'CH 3' should be written without a space 'CH3'

The third row: 'Aromatic CH stretching, vC–H, Vas CH2 and CH3, Stretching C–H'. The symbol 'ν' is using for denoting the 'stretching vibrations', so here the same information are written at least two times.

In the ninth row: 'C2c–H2 aromatic stretching'. I cannot understand what kind of vibration you mean.

Check please the assignment the bands in the last three rows.

In lines 356-357 you introduce the designations for different asymmetric and symmetric vibrations of CH2 and CH3 groups (in parentheses). It is differ from designations in Table 1 and did you really need to do it here?

Line 379. '… wavenumber region are more prominent in newborn plasma compared to pregnant plasma…' As I see in Figure 1, everything should be the opposite. The well noticed changes are in the panels A and C…

Author Response

Reviewer: 2

Comments to the Author

Response: Thank you for pointing this out. High wavenumber is a common terminology within vibrational spectroscopy. “Middle infrared spectroscopy” is not specific to where “middle” is. That would add vagueness to the title of our paper. To accommodate the reviewers’ request, we specified the wavenumber range in the Abstract and second last paragraph of the 1. Introduction section.

Paper text modified:

“Abstract: We evaluated the diagnostic potential of high wavenumber FTIR spectroscopy (2800-3600 cm^-1) and captured critical biochemical information associated with the interaction between bound water and other metabolites within blood plasma metabolites.”

“1. Introduction

By providing details on biomarkers and the diagnostic relevance associated with high-wavenumber FTIR spectroscopy (2800-3600 cm^-1), we demonstrated its potential to be translated into clinical practice and/or to complement other clinical and laboratory tests.”

Response: Thank you for your comment. The significance can only be assessed with standard deviation when the data distribution is normal around the means. This is not the case of our dataset. Also, showing standard deviation of our FTIR spectra does not change the results of our machine learning classification. With this in mind, we considered that showing the standard deviation in FTIR spectra does not add any significance and any additional useful meaning to readers.

Finally, mean spectra have been used to show clearly where the main vibrational modes occur for results of Table 1. Adding standard deviation to FTIR spectra hinders observing where the main vibrational modes occur, since the main bands of vibrational modes are not based on where FTIR spectra vary the most. Instead, the main bands were selected based on the amplitude of the mean FTIR spectra. We clarified this in the first paragraph of the 3. Results section.

Paper text modified: 3. Results

“The main bands of vibrational modes (spectral peaks and shoulder) selected based on the amplitude of the mean FTIR spectra were used for the assignment of the main biochemical constituents of blood plasma as shown in Table 1 and discussed in the “4. Discussion” section.”

Check please dimensions of ordinate axes in the panels B and D. Because it is the second derivative it should include cm², not only absorbance.

Response: Thank you for pointing this out. We agree with the reviewer and corrected the units of the axis of panels B and D by adding “Absorbance x cm² (arb. units)” instead of only absorbance. The corrected figure is replaced in the “3. Results section”.

Paper text modified:

Figure 1. Mean FTIR spectra of blood plasma samples of the control and pre-eclampsia groups for A) Pregnant women and B) Newborns after spectral smoothing and vector normalization, as well as the mean of the second derivative of the same FTIR spectra for the C) Pregnant women and D) Newborns. Spectra in A) and B) show the main vibrational modes for the FTIR peak/band assignment according to Table 1.

In Table one in the first row: what does mean the asterisk after 2853.6?

Response: We apologize for the typo of the asterisk, as it does not have any particular meaning to the data of our table. We have now removed the asterist from the first row of Table 1.

Paper text modified:

Results

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Bands (cm-1)	Vibrational modes	Biomolecular components	Study groups where vibrational modes can be observed
2854
	ν_s CH₂ of lipids, νC–H, CH₂ symmetric stretching, Asymmetric CH₂ stretching mode of the methylene chains in membrane lipids	Lipids and small contribution of carbohydrates, nucleic acids and proteins with C-H bonds	All

2873	ν_s CH3, Stretching C–H and N–H, CH3 symmetric stretching, Symmetric stretching vibration of CH₃ of acyl chains	Lipids, peptides / proteins, and contribution from carbohydrates, nucleic acids with C-H and N-H bonds	All
2873
2927

	νC–H, ν_as CH₂and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Pregnant (control and pre-eclampsia)
2932	ν_as CH₂ and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Newborn (control and pre-eclampsia)
2932
2959	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	Newborn (pre-eclampsia)
2960	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	All
3011	ν =CH	Pregnant (control and pre-eclampsia)	Newborn (control and pre-eclampsia)
3013	ν =CH	Unsaturated lipids and cholesterol esters	All
3065	C^2c–H₂ aromatic stretching	Major contribution: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All

3083	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3114	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3192	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Newborn (control and pre-eclampsia)
3194	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Pregnant (control and pre-eclampsia)
3280	νO–H(water), νN–H (protein)	Water and proteins	All
3288	Stretching OH symmetric	Water	All
3456	Stretching OH symmetric	Water	All
3483	OH bonds	Water	All

Response: Thank you for pointing this out. We agree with the reviewer and rounded the wavenumbers of Table 1 to avoid the impression our spectrometer can reach wavenumber decimals of resolution.

Paper text modified:

Results

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Bands (cm-1)	Vibrational modes	Biomolecular components	Study groups where vibrational modes can be observed
2854
	ν_s CH₂ of lipids, νC–H, CH₂ symmetric stretching, Asymmetric CH₂ stretching mode of the methylene chains in membrane lipids	Lipids and small contribution of carbohydrates, nucleic acids and proteins with C-H bonds	All

2873	ν_s CH3, Stretching C–H and N–H, CH3 symmetric stretching, Symmetric stretching vibration of CH₃ of acyl chains	Lipids, peptides / proteins, and contribution from carbohydrates, nucleic acids with C-H and N-H bonds	All
2873
2927

	νC–H, ν_as CH₂and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Pregnant (control and pre-eclampsia)
2932	ν_as CH₂ and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Newborn (control and pre-eclampsia)
2932
2959	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	Newborn (pre-eclampsia)
2960	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	All
3011	ν =CH	Pregnant (control and pre-eclampsia)	Newborn (control and pre-eclampsia)
3013	ν =CH	Unsaturated lipids and cholesterol esters	All
3065	C^2c–H₂ aromatic stretching	Major contribution: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All

3083	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3114	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3192	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Newborn (control and pre-eclampsia)
3194	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Pregnant (control and pre-eclampsia)
3280	νO–H(water), νN–H (protein)	Water and proteins	All
3288	Stretching OH symmetric	Water	All
3456	Stretching OH symmetric	Water	All
3483	OH bonds	Water	All

In the second row 'CH 3' should be written without a space 'CH3'

Response: Thank you for your suggestion. We have corrected our symbols and notations as requested by the reviewer.

Paper text modified:

Results

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Bands (cm-1)	Vibrational modes	Biomolecular components	Study groups where vibrational modes can be observed
2854
	ν_s CH₂ of lipids, νC–H, CH₂ symmetric stretching, Asymmetric CH₂ stretching mode of the methylene chains in membrane lipids	Lipids and small contribution of carbohydrates, nucleic acids and proteins with C-H bonds	All

2873	ν_s CH3, Stretching C–H and N–H, CH3 symmetric stretching, Symmetric stretching vibration of CH₃ of acyl chains	Lipids, peptides / proteins, and contribution from carbohydrates, nucleic acids with C-H and N-H bonds	All
2873
2927

	νC–H, ν_as CH₂and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Pregnant (control and pre-eclampsia)
2932	ν_as CH₂ and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Newborn (control and pre-eclampsia)
2932
2959	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	Newborn (pre-eclampsia)
2960	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	All
3011	ν =CH	Pregnant (control and pre-eclampsia)	Newborn (control and pre-eclampsia)
3013	ν =CH	Unsaturated lipids and cholesterol esters	All
3065	C^2c–H₂ aromatic stretching	Major contribution: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All

3083	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3114	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3192	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Newborn (control and pre-eclampsia)
3194	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Pregnant (control and pre-eclampsia)
3280	νO–H(water), νN–H (protein)	Water and proteins	All
3288	Stretching OH symmetric	Water	All
3456	Stretching OH symmetric	Water	All
3483	OH bonds	Water	All

Response: We appreciate your observation regarding the redundancy in the third row, where both 'νC–H' and 'Stretching C–H' convey similar information about the stretching vibrations of C–H bonds. The use of the symbol 'ν' is a standard notation in spectroscopy to specifically denote vibrational modes, while the phrase 'Stretching C–H' provides a more descriptive understanding of the type of vibration occurring.

To clarify this for the readers, we revised the text to eliminate redundancy. We retained the use of the 'ν' notation for its scientific significance while omitting the repetitive description.

Paper text modified: Third row of Table 1

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Bands (cm-1)	Vibrational modes	Biomolecular components	Study groups where vibrational modes can be observed
2854
	ν_s CH2 of lipids, νC–H, CH₂ symmetric stretching, Asymmetric CH₂ stretching mode of the methylene chains in membrane lipids	Lipids and small contribution of carbohydrates, nucleic acids and proteins with C-H bonds	All

2873	ν_s CH3, Stretching C–H and N–H, CH3 symmetric stretching, Symmetric stretching vibration of CH₃ of acyl chains	Lipids, peptides / proteins, and contribution from carbohydrates, nucleic acids with C-H and N-H bonds	All
2873
2927

	νC–H, νas CH₂and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Pregnant (control and pre-eclampsia)

In the ninth row: 'C2c–H2 aromatic stretching'. I cannot understand what kind of vibration you mean.

Response: Thank you for your feedback regarding Table 1, which outlines the main vibrational modes and structural components identified in the FTIR peaks of blood plasma spectra across our study groups. We would like to clarify that the “2c” is superscript for the carbon within the aromatic ring. C^2c–H₂ aromatic stretching is associated with the vibrational modes of CH₂ groups found mainly in steroids and carbohydrates but also found in side chains of lipids and proteins. This had already been described in the ninth row of Table 1, and now it has a clear superscript to make sure readers understand the CH₂ notation. Finally, this notation is the same as used by Talari et al. and Movasaghi et. al., references the peak/band assignments were based on:

Talari ACS.; Martinez MAG.; Movasaghi Z.; Rehman S.; Rehman IU. (2017). Advances in Fourier transform infrared (FTIR) spectroscopy of biological tissues. Applied Spectroscopy Reviews, 2017, Vol52(5), page 456-506.
Movasaghi Z.; Rehman S.; Rehman DI. Fourier transform infrared (FTIR) spectroscopy of biological tissues. Applied Spectroscopy Reviews, 2008, Vol 43, page 134-79.

Paper text modified: “3. Results” section (Table 1, ninth row)

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Bands (cm-1)	Vibrational modes	Biomolecular components	Study groups where vibrational modes can be observed
2854
	ν_s CH₂ of lipids, νC–H, CH₂ symmetric stretching, Asymmetric CH₂ stretching mode of the methylene chains in membrane lipids	Lipids and small contribution of carbohydrates, nucleic acids and proteins with C-H bonds	All

2873	ν_s CH3, Stretching C–H and N–H, CH3 symmetric stretching, Symmetric stretching vibration of CH₃ of acyl chains	Lipids, peptides / proteins, and contribution from carbohydrates, nucleic acids with C-H and N-H bonds	All
2873
2927

	νC–H, ν_as CH₂and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Pregnant (control and pre-eclampsia)
2932	ν_as CH₂ and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Newborn (control and pre-eclampsia)
2932
2959	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	Newborn (pre-eclampsia)
2960	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	All
3011	ν =CH	Pregnant (control and pre-eclampsia)	Newborn (control and pre-eclampsia)
3013	ν =CH	Unsaturated lipids and cholesterol esters	All
3065	C^2c–H₂ aromatic stretching	Major contribution: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All

Talari ACS.; Martinez MAG.; Movasaghi Z.; Rehman S.; Rehman IU. (2017). Advances in Fourier transform infrared (FTIR) spectroscopy of biological tissues. Applied Spectroscopy Reviews, 2017, Vol52(5), page 456-506.
Movasaghi Z.; Rehman S.; Rehman DI. Fourier transform infrared (FTIR) spectroscopy of biological tissues. Applied Spectroscopy Reviews, 2008, Vol 43, page 134-79.

Check please the assignment the bands in the last three rows.

Response: Thank you for your comment. We understand that the assignment may not have been clear due to the lack of references used for the vibrational modes of such assignment. Therefore, we added the references in the header of Table 1 in the “3. Results” section. Also, we are aware that the high wavenumber infrared spectroscopy (HWIR) has been scarcely explored. With this in mind, we added more information about HWIR in the 6^th paragraph of the “1. Introduction” section.

Paper text modified:

Results (table 1 header)

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Introduction (6^th paragraph)

The use of the high wavenumber (HW) region, between 2600 and 3800 cm⁻¹, for the diagnosis of pathologies offers several advantages over the fingerprint region (400 to 1800 cm⁻¹). Recent studies have shown that the HW region contains diagnostic information similar to the fingerprint region, making it effective in discriminating tissue structures with different molecular compositions, as observed in brain and bladder tumors. The use of FTIR spectroscopy in the HW region can be defined as high-wavenumber infrared spectroscopy or HWIR spectroscopy. When investigating biological samples, HWIR spectroscopy captures molecular vibrations of the hydrophilic sites within biomolecules such as lipids, proteins, carbohydrates and nucleic acids. Given the maturity of FTIR as a technology, not only portable equipment can be produced, but also biofluid analysis and sample classification can be achieved in minutes, limited mostly by the sample drying time. Little to no professional training required to use FTIR equipment. In addition, clinical translation of bound water analysis can be facilitated by automatic analysis of the HW region of infrared spectra by using machine learning. Thus, HW spectroscopy presents a more practical and economical alternative for diagnostics that can complement biochemical analysis and potentially introduce new methods for time-sensitive clinical decision-making [12].

Response: Thank you for pointing out that. We have now harmonized the designations/notations of the vibrations in Table 1 and the rest of the manuscript. Our text modifications can be checked in the “4. Discussion” section.

Paper text modified:

In the high wavenumber region, the spectrum is primarily attributed to lipid absorption for bands at 2849, 2917, and 3008 cm^-1, and dominated by water absorption bands at 3350 cm⁻¹. The spectral region 3050–2800 cm^-1 corresponds to asymmetric CH₃ (ν_asCH₃), asymmetric CH₂ (ν_asCH₂), symmetric CH₃ (ν_sCH₃), and symmetric CH₂ (ν_sCH₂) distribution, observable in the control and preeclampsia sample results [14]. This factor is essential for fetal development, especially in the late gestation period, a critical time for preeclampsia patients [22].

Response: Thank you for pointing this out, as our conclusion is based on our results from our machine learning classification instead of interpreting anything directly from mean spectra, which were used primarily to understand the main contribution of biomolecular vibrational modes in our FTIR spectra. Instead, spectral changes between study groups (control and pre-eclampsia groups) are determined by how well machine learning algorithms are able to correctly classify FTIR spectra. We clarified this in the last paragraph of the “4. Discussion” section.

Paper text modified: Since our machine learning classification was more accurate for newborns than for pregnants, our results indicated that changes in the FTIR high wavenumber region are more prominent in newborn plasma compared to pregnant plasma. Hence, our results may suggest that newborns are highly affected at the time of birth.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

As shown in the attachment.

Comments for author File: Comments.pdf

Author Response

For the point-by-point responses including figures and indications of manuscript modifications, we suggest checking the PDF file attached.

Reviewer: 3

Comments to the Author:

Researchers combining high wavelength FTIR spectroscopy with other analytical methods to improve machine learning performance and thereby increase the accuracy of diagnosis of pre-eclampsia. The subject matter of the manuscript is of great importance to scientific and technical groups working in the field of developing methods for early detection of the disease.

In introduction, we suggest a comparison of FTIR spectroscopy with other spectroscopic techniques to show its advantages and limitations.

Response: Thank you for your suggestion regarding the inclusion of a comparison between FTIR spectroscopy and other spectroscopic techniques in the introduction. In response, we will incorporate a comparative analysis of FTIR with techniques such as Raman spectroscopy, mass spectrometry, and nuclear magnetic resonance (NMR) spectroscopy, highlighting how FTIR offers rapid, non-invasive, and high-throughput analysis with fewer issues related to fluorescence interference, which often complicates Raman spectra. Additionally, we will address the limitations of FTIR, such as its sensitivity to water, and emphasize how its ability to analyse complex biological fluids complements the more targeted yet time-consuming approaches of mass spectrometry and NMR. This comparison will further clarify the unique benefits of FTIR in the context of pre-eclampsia detection and other biomedical applications, so we incorporate this in the text.

Paper text modified: Fourier-transform infrared (FTIR) spectroscopy offers significant advantages compared to other spectroscopic techniques, such as mass spectrometry, nuclear magnetic resonance (NMR) spectroscopy, and UV-Vis spectroscopy. Mass spectrometry is highly sensitive and capable of identifying specific molecular compounds but requires complex sample preparation and may not be suitable for rapid analyses in clinical settings. Conversely, NMR spectroscopy provides detailed structural information about analysed molecules but is limited by high operational costs and the need for large sample volumes. UV-Vis spectroscopy, while useful for analyzing compounds that absorb ultraviolet or visible light, faces challenges when applied to complex biological fluids like blood plasma, where the presence of multiple components can interfere with readings. In contrast, FTIR stands out for its ability to perform rapid, non-invasive analyses, providing a comprehensive view of the biochemical composition of samples, making it a promising tool for early detection of pre-eclampsia and other pathological conditions.

J. D. S. Oliveira, et al. "Spectroscopic Techniques in the Diagnosis of Diseases: A Review." Journal of Biomedical Optics, vol. 25, no. 4, 2020, Article 040901. doi:10.1117/1.JBO.25.4.040901.

From Figure 1, it can be seen that the control and pre-eclampsia spectral data of newborns are quite close to each other, how to ensure the accuracy of the modeling of machine learning.

Response: Thank you for your observation regarding the similarity between the control and pre-eclampsia spectral data of newborns in Figure 1. We acknowledge that the spectral data from both groups may appear close to each other in average spectra. We emphasize that trying to observe differences in average spectra will not provide information about results of machine learning models. We employed several strategies to ensure the accuracy of the machine learning modeling was as high as possible, including optimizing our spectra processing and machine leaning classification described in our manuscript. Also, we evaluated our models by using one of the most robust validations of all models by training and validating our models in different sets by using 20 iterations of 5-fold cross-validation with random sampling. This validation is much more robust than other research works which only report the mean of the 5-fold cross validation without evaluating the standard deviation of the model with random sampling. In terms of spectral processing and machine learning classification, we clearly explained below the steps of our analysis.

Firstly, the spectral data underwent rigorous pre-processing, including smoothing with normalization, to minimize noise and enhance signal clarity. These steps are crucial in reducing the potential overlap in spectral features, especially when small differences exist between groups.

Secondly, the second derivative of the normalized spectra was utilized, which enhances subtle variations in the spectral data, particularly in cases where the original spectra show proximity. This method has been demonstrated to improve the discrimination between groups with similar spectral profiles by highlighting differential vibrational modes more effectively.

Furthermore, we employed several machines learning algorithms, including Support Vector Machine (SVM), k-Nearest Neighbors (k-NN), and Linear and Quadratic Discriminant Analyses, to classify the spectral data. The robustness of the classification was validated through 20 iterations of 5-fold cross-validation, ensuring that the models were not overfitted and that the performance metrics were stable across multiple random samples. The accuracy, sensitivity, and specificity metrics reported are the averaged results of these iterations, providing confidence in the models' ability to discriminate between the groups, despite the visual proximity of the data.

Lastly, while the visual closeness of the spectra is evident, the machine learning models were able to identify subtle biochemical differences between the control and pre-eclampsia groups, which are reflected in the classification performance.

We hope this explanation clarifies the steps taken to ensure the accuracy of the machine learning models. Should further details be required, we would be happy to provide additional information.

Numbers involved in functional groups should be subscripted, please check the full text.

Response: We appreciate your observation regarding the need for subscripting the numbers associated with functional groups throughout the text. We have carefully reviewed the entire manuscript and ensured that all relevant chemical formulas and functional group representations are correctly formatted with subscripts where applicable.

For instance, in the context of the FTIR analysis, we have corrected representations such as CH₂, CH₃, C–H, and N–H to maintain consistent scientific formatting. Our text modifications can be checked throughout Table 1 and the “4. Discussion” section.

Paper text modified:

Discussion:

“In the high wavenumber region, the spectrum is primarily attributed to lipid absorption for bands at 2849, 2917, and 3008 cm^-1, and dominated by water absorption bands at 3350 cm⁻¹. The spectral region 3050–2800 cm^-1 corresponds to asymmetric CH₃ (ν_asCH₃), asymmetric CH₂ (ν_asCH₂), symmetric CH₃ (ν_sCH₃), and symmetric CH₂ (ν_sCH₂) distribution, observable in the control and preeclampsia sample results [14]. This factor is essential for fetal development, especially in the late gestation period, a critical time for preeclampsia patients [22].”

“According to the absorbance range found for peptides in the high wavenumber region, the bands between 3280 cm^-1 (H – O – H stretch), 2957 cm^-1 (asymmetric CH₃ stretch), 2920 cm^-1 (asymmetric CH₂ stretch), 2872 cm^-1 (symmetric CH₃ stretch), 1536 cm^-1 (amide II of proteins), 1453 cm-1 (CH₂ scissor), 1394 cm^-1 (C=O stretch of COO–), 1242 cm^-1 (asymmetric PO₂ stretch), 1171 cm^-1 (asymmetric C –O ester stretch), and 1080 cm^-1 (C – O stretch) were observed [26].”

Results

Table 1. Assignment of the main vibrational modes and structural components to FTIR peaks evidenced in blood plasma spectra of our study groups [16, 17].

Bands (cm-1)	Vibrational modes	Biomolecular components	Study groups where vibrational modes can be observed
2854
	ν_s CH₂ of lipids, νC–H, CH₂ symmetric stretching, Asymmetric CH₂ stretching mode of the methylene chains in membrane lipids	Lipids and small contribution of carbohydrates, nucleic acids and proteins with C-H bonds	All

2873	ν_s CH3, Stretching C–H and N–H, CH3 symmetric stretching, Symmetric stretching vibration of CH₃ of acyl chains	Lipids, peptides / proteins, and contribution from carbohydrates, nucleic acids with C-H and N-H bonds	All
2873
2927

	νC–H, ν_as CH₂and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Pregnant (control and pre-eclampsia)
2932	ν_as CH₂ and CH₃, Stretching C–H	Lipids, proteins, carbohydrates and nucleic acids	Newborn (control and pre-eclampsia)
2932
2959	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	Newborn (pre-eclampsia)
2960	CH₃ asymmetric stretching, CH stretching, ν_as CH₃, asymmetric stretching mode of the methyl groups	Lipids, proteins, carbohydrates and nucleic acids	All
3011	ν =CH	Pregnant (control and pre-eclampsia)	Newborn (control and pre-eclampsia)
3013	ν =CH	Unsaturated lipids and cholesterol esters	All
3065	C^2c–H₂ aromatic stretching	Major contribution: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All

3083	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3114	C–H ring	Major: Steroids and carbohydrates, Small contribution: side chains of lipids and proteins	All
3192	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Newborn (control and pre-eclampsia)
3194	N–H stretching bands of mainly cis-ordered substructures, Stretching N–H symmetric	Peptides / proteins	Pregnant (control and pre-eclampsia)
3280	νO–H(water), νN–H (protein)	Water and proteins	All
3288	Stretching OH symmetric	Water	All
3456	Stretching OH symmetric	Water	All
3483	OH bonds	Water	All

To help the readers have a more comprehensive understanding of the new research on terahertz spectra, I suggest supplementing some latest works about it [Photonix. 2021, 2.;

Photonix. 2020, 1.; Photonics. 2024, 11.; Advanced Photonics Nexus. 2023,044002.].

Response: Thank you for your valuable suggestion regarding the inclusion of recent works on terahertz spectra. Without any clear explanation from the reviewer, we were not sure how “the new research on terahertz spectra” is relevant to our work in infrared spectroscopy. Therefore, we still did not add the recommended references to our manuscript. Our decision relies on the fact that it would be unethical for a reviewer to recommend references without clear explanation of how the quality of the manuscript will be improved by adding such references. Please check the section 7.4 Review reports in the webpage https://www.mdpi.com/reviewers (text copied below):

Reviewers must not recommend excessive citation of their work (self-citations), another author’s work (honorary citations) or articles from the journal where the manuscript was submitted as a means of increasing the citations of the reviewer/authors/journal. You can provide references as needed, but they must clearly improve the quality of the manuscript under review.

The formatting of references need to be checked in full.

Response: Thank you for your comment. We reformatted the references and included the new citations requested by the reviewers.

Paper text modified:

References

Bunaciu, A.; Fleschin, S.; Aboul-Enein, H. Infrared Microspectroscopy Applications - Review. Current Analytical Chemistry, 2013, Vol.10(1), pages 132–139. doi:10.2174/1573411011410010011
Vidaeff AC.; Saade GR.; Sibai BM. Preeclampsia: The Need for a Biological Definition and Diagnosis. American Journal of Perinatology, 2020.doi:10.1055/s-0039-1701023
Rana S.; Lemoine E.; Granger JP.; Karumanchi SA. Preeclampsia: Pathophysiology, Challenges, and perspectives. Circ Res. 2019, Vol 7, page 1094-1112.
Cameron J.M.; Rinaldi C.; Butler H.J.; Hegarty M.G.; Brennan P.M.; Jenkinson M.D.; Syed K.; Ashton K.M.; Dawson T.P.; Palmer, D.S.; et al. Stratifying Brain Tumour Histological Sub-Types: The Application of ATR-FTIR Serum Spectroscopy in Secondary Care. Cancers, 2020,12, 1710. https://doi.org/10.3390/cancers12071710
Magalhães S.; Goodfellow BJ.; Nunes A. FTIR spectroscopy in biomedical research: how to get the most out of its potential. Applied Spectroscopy Reviews, 2021. Page 1–39. doi:10.1080/05704928.2021.1946822
Baker MJ.; Trevisan J.; Bassan P.; Bhargava R.; Butler HJ.; Dorling KM., et al. Using Fourier transform IR spectroscopy to analyze biological materials. Nature Protocols, 2014. Vol 9(8), page 1771–1791. doi:10.1038/nprot.2014.110
Kumar S.; Srinivasan A.; Nikolajeff, F. Role of Infrared Spectroscopy and Imaging in Cancer Diagnosis. Current Medicinal Chemistry, 2018, Vol. 25(9), page 1055–1072. doi:10.2174/092986732466617052312
Al-Kelani M.; Buthelezi N.; Advancements in medical research: Exploring Fourier Transform Infrared (FTIR) spectroscopy for tissue, cell, and hair sample analysis. Skin Res Technol, 2024, Vol 30. Doi: 10.1111/srt.13733.
Theakstone A. G.; Rinaldi C.; Butler H.J. et al. Fourier‐transform infrared spectroscopy of biofluids: A practical approach. Translational Biophotonics, 2021, Vol 3.
Movasaghi Z.; Rehman S.; Rehman DI. Fourier Transform Infrared (FTIR) Spectroscopy of Biological Tissues. Applied Spectroscopy Reviews, 2008, Vol 2, page 134–179.
das Chagas e Silva de Carvalho LF.; Sato ÉT.; Almeida JD. et al.Diagnosis of inflammatory lesions by high-wavenumber FT-Raman spectroscopy. Theor Chem Acc, 2011, Vol 130, 1221–1229. Doi: https://doi.org/10.1007/s00214-011-0972-2
Korb E.; Bağcıoğlu M.; Garner-Spitzer E.; Wiedermann U.; Ehling-Schulz M.; Schabussova I. Machine Learning-Empowered FTIR Spectroscopy Serum Analysis Stratifies Healthy, Allergic, and SIT-Treated Mice and Humans. Biomolecules, 2020. Vol. 10. doi: 10.3390/biom10071058.
Kenny L.C.; Dunnb W.B.; Ellis, DI. et al.Novel biomarkers for pre-eclampsia detected using metabolomics and machine learning. Metabolomics, 2005. Vol. 1, page 227–234. Doi: https://doi.org/10.1007/s11306-005-0003-1
Mukherjee R.; Ray CD.; Ray S.; Dasgupta S.; Chaudhury K. Altered metabolic profile in early and late onset preeclampsia: An FTIR spectroscopic study. Pregnancy Hypertension: An International Journal of Women’s Cardiovascular Health, 2014, Vol 4(1), 70–80.
Kulkarni A.; Chavan-Gautam P.; Mehendale S.; Yadav H.; Joshi S. Global DNA methylation patterns in placenta and its association with maternal hypertension in pre-eclampsia. DNA Cell Biol, 2011. Vol. 30, 79–84.
Talari ACS.; Martinez MAG.; Movasaghi Z.; Rehman S.; Rehman IU. (2017). Advances in Fourier transform infrared (FTIR) spectroscopy of biological tissues. Applied Spectroscopy Reviews, 2017, Vol52(5), page 456-506.
Movasaghi Z.; Rehman S.; Rehman DI. Fourier transform infrared (FTIR) spectroscopy of biological tissues. Applied Spectroscopy Reviews, 2008, Vol 43, page 134-79.
Silasi, M., Cohen, B., Karumanchi, S.A., Rana, S. Abnormal placentation, angiogenic factors, and the pathogenesis of preeclampsia. Obstet. Gynecol. North, 2010, Vol 37, pages 239–253.
Ferreira MCC.; Monteiro GR.; Peralta F.; Castro PAA.; Zezell D.; Nogueira MS.; Carvalho LFC. Assessment of bound water of saliva samples by using FT-IR spectroscopy. Latin America Optics and Photonics (LAOP) Conference, 2022. paper M4B.1.
Sibai, B., Dekker, G., Kupferminc, M. Pre-eclampsia. The Lancet, 2005, Vol 9461, page 785–799.
Qian, Y., Zhang, L., Rui, C., Ding, H., Mao, P., Ruan, H., & Jia, R. (2017). Peptidome analysis of amniotic fluid from pregnancies with preeclampsia. Molecular Medicine Reports, 2017, Vol 16, 7337-7344.
Mukherjee R.; Ray CD.; Chakraborty C.; Dasgupta S.; Chaudhury K. Clinical biomarker for predicting preeclampsia in women with abnormal lipid profile: statistical pattern classification approach. In: International conference on systems in medicine and biology, 2010.
Nobakht BF. Application of metabolomics to preeclampsia diagnosis. Systems Biology in Reproductive Medicine, 2018, Vol 64(5), 324–339.

In Table 3, the first column has already included the percentage symbol (%), so there is no need to add the percentage again to the specific numerical values.

Response: We agree with the reviewer and removed the percentage symbol from the values of table 3. For consistency when reporting our classification performance metrics in Table 2 and 3, we also applied the same notation to Table 2.

Paper text modified:

Classification performance metric	FTIR spectra	Second derivative of FTIR spectra
Accuracy (%)	71.8 ± 2.5	63.8 ± 2.0
AUC_{Control newborn}	0.918 ± 0.013	0.923 ± 0.008
AUC_{Control pregnant}	0.901 ± 0.009	0.886 ± 0.009
AUC_{Preclampsia newborn}	0.933 ± 0.016	0.864 ± 0.015
AUC_{Preclampsia pregnant}	0.837 ± 0.020	0.810± 0.013

AUC: area under curve

	Pregnant		Newborn
Spectral type	FTIR spectra	Second derivative of FTIR spectra	FTIR spectra	Second derivative of FTIR spectra
Classification performance metric
Sensitivity (%)	76.3 ± 3.5	63.7 ± 5.9	79.0 ± 3.5	74.5 ± 4.2
Specificity (%)	56.1 ± 4.4	54.1 ± 5.1	76.9 ± 6.2	68.8 ± 4.0
Accuracy (%)	66.3 ± 2.9	58.9 ± 3.5	78.0 ± 3.8	71.6 ± 2.8
	0.692 ± 0.036	0.613 ± 0.032	0.83 ± 0.04	0.792 ± 0.025
	0.692 ± 0.036	0.613 ± 0.032	0.83 ± 0.04	0.792 ± 0.025

Although the paper provides metrics for classification performance, there is insufficient explanation of the model's predictive results. Specifically, it is unclear which biochemical changes are associated with pre-eclampsia. Additionally, it is recommended to perform external validation or use an independent dataset to test the model's predictive capability. Interpretability of the model is crucial, especially in medical diagnostics.

Response: Thank you for your comment. The objective of our work is not to identifying the most relevant biomarkers for classification of control and pre-eclampsia samples based on FTIR spectra. Instead of the two aims of our current work are: 1. optimizing the machine-learning classification of control and pre-eclampsia samples, and 2. reporting the most prominent biochemical components of blood plasma of pregnants and newborns in the control and pre-eclampsia groups). These two aims make our work already extremely novel and impactful, given how scarcely explored the high-wavenumber infrared region is.

While we understand that the reviewer wants to have interpretability of models, using Principal component Analysis (PCA), Partial Least-Squares (PLS) and similar/derived methods for feature selection and dimensionality reduction is out of the scope of this work. We already planned to do this type of analysis as the next step of our future work, when we include a “new analysis with increased number of patients and association with 24-hour urine sample analysis for further correlation with plasma biochemical changes and understanding the origin of such changes”, as already mentioned in the ”4. Discussion” section. We clarified this in the end of the “4. Discussion” section.

Paper text modified: This future work may include feature selection or dimensionality reduction techniques to identify the most relevant biomarkers for classification of control and pre-eclampsia samples based on FTIR spectra (instead of the two aims of this work: 1. optimizing the machine-learning classification of control and pre-eclampsia samples, and 2. reporting the most prominent biochemical components of blood plasma of pregnants and newborns in the control and pre-eclampsia groups).

The paper does not explicitly mention the application of feature selection or dimensionality reduction techniques. In high wavenumber FTIR spectral analysis, there may be a large number of irrelevant or redundant features.

Response: Thank you for your comment. Feature selection and dimensionality reduction techniques have not been used in our work because SVM already selects the specific supporting vectors to compute the margin that best separate groups. Since the separation is based on the margin, there is no need for feature selection and/or dimensionality reduction. In fact, the classification performance metrics are even reduced when feature selection and/or dimensionality reduction are applied prior to using SVM. Therefore, not only talking about feature selection and dimensionality reduction is out of the scope of our work, but also would degrade the quality of our work.

Our future work will explore feature selection and dimensionality reduction techniques to look at the most relevant biomarkers within FTIR spectra. This analysis is not to be confused with our peak assignment, and is not in the scope of our current manuscript, as it has a totally different purpose (not associated with optimizing sample classification or describing the most prominent biochemical components in blood plasma of pregnants and newborns in the control and pre-eclampsia groups). We clarified our future work in the end of the “4. Discussion” section.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear Authors,

Thank you for submitting your revised manuscript to Photonics. I have carefully reviewed the updated version and am pleased to report that all the points raised in my initial review have been thoroughly addressed.

Your manuscript presents an innovative and scientifically sound approach, and the revisions have successfully improved the clarity and completeness of the manuscript. The corrections and added details are well-incorporated, and I appreciate the thoughtful responses to my comments.

I believe the manuscript is suitable for publication and will contribute to the field. Congratulations on your work!

Best regards,

Author Response

Thank you for accepting our study.

Reviewer 2 Report

Comments and Suggestions for Authors

Paper text modified:

“Abstract: We evaluated the diagnostic potential of high wavenumber FTIR spectroscopy (2800-3600 cm^-1) and captured critical biochemical information associated with the interaction between bound water and other metabolites within blood plasma metabolites.”

“1. Introduction

By providing details on biomarkers and the diagnostic relevance associated with high-wavenumber FTIR spectroscopy (2800-3600 cm^-1), we demonstrated its potential to be translated into clinical practice and/or to complement other clinical and laboratory tests.”

First of all I apologize for suggesting wrong term "Middle Infrared Spectroscopy". The right one is "mid-infrared spectroscopy". It is often used for designation of the region of 4000 – 400 cm⁻¹. (https://www.sciencedirect.com/topics/chemistry/mid-ir-spectroscopy, https://en.wikipedia.org/wiki/Infrared_spectroscopy). However if authors will indicate what they mean by their term in Abstract and in Introduction, they, of course, may use it. Unfortunately, in the final version of Abstract the mention of the range disappeared. The text in Introduction also missed in the final version, however the range is mentioned.

The other my comments was either taken into account or clarified. So, I believed that the text can be published after minor correction.

Author Response

Thank you for accepting our study.

Concernt the comment, we appreciate and add the sentences in the abstract and introduction, this could be seen in RED in the new version of the manuscript.

"This study investigates the potential of high-wavenumber FTIR spectroscopy (region between 2800-3600 cm^-1) as an innovative"

" The use of the high wavenumber (HW) region, between 2600 and 3800 cm⁻¹, for the diagnosis of pathologies offers several advantages over the fingerprint region (400 to 1800 cm⁻¹). This spectral region mostly comprises vibrational modes related to CH2, CH3 and OH, which are biochemically within certain lipids, proteins and confined water."

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for your revision.

Author Response

Thank you for accepting our study.

Article Menu

High-Wavenumber Infrared Spectroscopy of Blood Plasma for Pre-Eclampsia Detection with Machine Learning

Further Information

Guidelines

MDPI Initiatives

Follow MDPI