Label‐free atherosclerosis diagnosis through a blood drop of apolipoprotein E knockout mouse model using surface‐enhanced Raman spectroscopy validated by machine learning algorithm

Abstract The direct preventative detection of flow‐induced atherosclerosis remains a significant challenge, impeding the development of early treatments and prevention measures. This study proposes a method for diagnosing atherosclerosis in the carotid artery using nanometer biomarker measurements through surface‐enhanced Raman spectroscopy (SERS) from single‐drop blood samples. Atherosclerotic acceleration is induced in apolipoprotein E knockout mice which underwent a partial carotid ligation and were fed a high‐fat diet to rapidly induce disturbed flow‐induced atherosclerosis in the left common carotid artery while using the unligated, contralateral right carotid artery as control. The progressive atherosclerosis development of the left carotid artery was verified by micro‐magnetic resonance imaging (micro‐MRI) and histology in comparison to the right carotid artery. Single‐drop blood samples are deposited on chips of gold‐coated ZnO nanorods grown on silicon wafers that filter the nanometer markers and provide strong SERS signals. A diagnostic classifier was established based on principal component analysis (PCA), which separates the resultant spectra into the atherosclerotic and control groups. Scoring based on the principal components enabled the classification of samples into control, mild, and severe atherosclerotic disease. The PCA‐based analysis was validated against an independent test sample and compared against the PCA‐PLS‐DA machine learning algorithm which is known for applicability to Raman diagnosis. The accuracy of the PCA modification‐based diagnostic criteria was 94.5%, and that of the machine learning algorithm 97.5%. Using a mouse model, this study demonstrates that diagnosing and classifying the severity of atherosclerosis is possible using a single blood drop, SERS technology, and machine learning algorithm, indicating the detectability of biomarkers and vascular factors in the blood which correlate with the early stages of atherosclerosis development.

sample and compared against the PCA-PLS-DA machine learning algorithm which is known for applicability to Raman diagnosis. The accuracy of the PCA modificationbased diagnostic criteria was 94.5%, and that of the machine learning algorithm 97.5%. Using a mouse model, this study demonstrates that diagnosing and classifying the severity of atherosclerosis is possible using a single blood drop, SERS technology, and machine learning algorithm, indicating the detectability of biomarkers and vascular factors in the blood which correlate with the early stages of atherosclerosis development.

K E Y W O R D S
ApoE KO mouse, atherosclerosis, nano-sized biomarker, principal component analysis, surfaceenhanced Raman spectroscopy

| INTRODUCTION
Atherosclerosis is a chronic inflammatory vascular disease that is initiated by the accumulation, oxidation, and glycation of lowdensity lipoprotein (LDL) cholesterol in the endothelium and progresses with the expression of pro-inflammatory adhesion molecules and release of chemo-attractants. 1,2 Most clinical atherosclerotic cardiovascular diseases (ASCVDs) are characterized by stable or unstable coronary disease, ischemic stroke, and peripheral arterial disease; ASCVD is the final step of a vascular inflammatory process. 3,4 Advanced atherosclerosis induces chronic luminal narrowing and ischemia of peripheral tissues. 5 Moreover, unstable plaques with a high content of necrotic core and thin fibrous cap are prone to rupture, which may cause acute atherothrombotic events and tissue necrosis. 6 Unfortunately, the early identification of vulnerable atherosclerotic lesions using serum biomarkers remains challenging. 2,7 Traditional risk factors such as old age, hypertension, smoking, obesity, and diabetes show a considerably low specificity in ASCVD detection.
Although hyperlipidemia is related to the initiation and progression of atherosclerosis, no direct correlation has been established between the serum level of LDL cholesterol and lesion severity. The elevated serum level of C-reactive protein, an inflammatory mediator, is associated with the presence of ASCVD and occurrence of ischemic events.
However, the role of C-reactive protein in inflammatory mechanisms and its causality in atherosclerosis has not yet been fully elucidated. 8 Moreover, the level of C-reactive protein poorly reflects the degree of stenosis and morphological plaque characteristics. 9 Although cardiac troponin and creatine kinase-MB (CK-MB) are the key markers for the diagnosis of acute myocardial infarction, these post-event biomarkers cannot detect stable coronary artery disease or reduce cardiovascular risks. 10,11 While angiography is the gold standard tool used to assess the severity of stenosis, the contrast luminography cannot provide information on plaque morphology or arterial wall pathology. With a high resolution, intravascular ultrasound is useful in quantifying atheroma and characterizing tissue in the catheterization laboratory. However, the extensive use of intravascular imaging in the real world is restricted by its invasiveness and high cost. 12,13 Biological samples for liquid biopsy can be derived from any secretion containing metabolism byproducts, such as tears, blood, urine, and saliva. Biomarkers and reagents present in these materials can provide a basis for diagnosis, enabling an early measurement of drug treatment outcomes. 7,14,15 Biomarkers can be grouped into cells (approx. size of tens of μm), red blood cells (approx. 8 μm), bacteria (approx. 1 μm), viruses (approx. 400 nm), exosomes (several nm to tens of nm), and proteins, among others, with each group having a typical size range. The type of biomarker most appropriate for diagnosis depends upon the disease and its occurrence mechanism. In general, smaller target marker molecules in liquid samples can be detected from smaller sample quantities; however, a higher sensitivity detection technology is required.
Surface-enhanced Raman spectroscopy (SERS) is a useful candidate diagnostic technology for amplifying and measuring the signal of a small quantity of nanometer-sized markers. The magnitude of the enhanced Raman signal greatly varies with the type of metal, roughness, curvature, and shape of the nanostructure. Research is being conducted on which of these structured substrates efficiently enhances surface Raman scattering. Recently, the medical application of surface-enhanced Raman analysis has established the groundwork for diagnosis following a label-free approach, and optimizing the acquired signal has been shown to be possible by controlling the spacing and porosity of nanorod structures when fabricating a SERS chip. 16,17 Previous studies have demonstrated that animal models of interstitial cystitis and kidney injury were diagnosed by selectively filtering nanometer-sized biomarkers using a ZnO nanostructure-based SERS chip. 18,19 The nanorod-based surface-enhanced Raman chip with nanometer spacing proposed in this study has a structure suitable for measuring nano-sized biological particles such as proteins, lipids, nucleic acids, exosomes, and metabolic substances. If the nanometer marker is used as a diagnostic target, it can be measured in small quantities because it is evenly distributed in the liquid sample, and high reproducibility can be expected. However, as blood consists of a liquid-based serum and large amount of various cells such as red blood cells, white blood cells, and platelets, selective nanometer-scale biomarker detection using techniques such as centrifugation entails significant technical time and cost. Because the SERS chip used in this study exhibits nano-porosity, the chip itself may diffusely separate nanometer-scale biomarkers from the blood drop without additional separation steps. In addition, as the Raman signal is selectively enhanced by surface plasmons, the excitation is localized to only the separated and trapped region owing to the characteristics of the nanostructure, and processes such as enrichment become unnecessary. 20 In addition to high-sensitivity Raman signal acquisition nanotechnology, a technology capable of analyzing these spectral data and securing the basis for diagnosis is required because Raman signals in biological materials exhibit a variety of spectral peaks of various origins. Efforts have been made to secure the basis for diagnosis from data sets consisting of multiple overlapping peaks through dimensionality reduction using techniques such as principal component analysis (PCA). In particular, the Raman signal from a biomaterial has a greater data deviation than that of a crystalline sample, and because each peak of the spectrum changes dynamically, data analysis using automated pattern recognition technology is required. The application of artificial intelligence (AI)-based automation algorithms to SERS diagnosis technology has attracted considerable attention, as it enables the improvement and optimization of diagnosis accuracy. 21,22 One method of constructing a mouse model for the implementation of atherosclerosis is by administering a high-fat diet. However, in wild-type mice, the generation of plaques seldom proceeds past the early stages of lesion development. 23,24 In addition, wild-type mice are unsuitable for generating a model for atherosclerosis alone because other high-cholesterol-induced vascular diseases may be involved. 25 Apolipoprotein E (ApoE) knockout (KO) mice are host to a defective gene involved in cholesterol metabolism that accelerates atherosclerotic plaque formation under a high-fat diet. In addition, the acute development of atherosclerotic plaques can be triggered by partial carotid ligation surgery, which induces the pro-atherogenic disturbed blood flow characterized by low and oscillating shear stress. 26,27 The accelerated rate of atherosclerosis progression in ApoE KO mice minimizes other complications that may accompany atherosclerosis, making them ideal for confirming inflammatory biomarkers or Raman signals caused only by atherosclerosis.
Here, we applied a nanostructured SERS assay to predict early or advanced carotid atherosclerosis development status in individual mice, as shown in Figure 1. A mouse model of disturbed blood flow- This enabled the development of an early diagnostic criterion based on the severity of atherosclerosis progression, which was based on analyzing statistical information from the Raman signal to determine relevant factors. The criterion's predictive ability was validated by diagnosing an additional group, whose data had not been used to establish the criterion. In addition, a performance review and diagnosis optimization were performed by comparing the performance of F I G U R E 1 Schematic for nanomarker Raman-based diagnosis of arteriosclerosis. Atherosclerosis animal model preparation, sample acquisition, histopathology and immunofluorescence staining, pretreatment-and label-free SERS measurement. The red arrows track the process of obtaining a Raman signal based on a nanometer marker for inducing the severity of atherosclerosis, and the blue arrows track the verification process using histopathology. KO, knockout; LCA, left carotid artery; PCA, principal component analysis; RCA, right carotid artery; SERS, surfaceenhanced Raman spectroscopy.
PCA to PCA-PLS-DA, an additional machine learning algorithm.
Altogether, the results indicate that the combination of machine learning and SERS-based nano biomarker technologies using blood samples merits further application as a means for the early diagnosis of atherosclerosis.

| Selection of atherosclerotic animal groups for diagnostic criteria
To confirm the performance of SERS, atherosclerotic mice were produced by the partial ligation of the LCA, which can generate disturbed blood flow with low endothelial shear stress, as described in previous studies. [26][27][28] In C57BL/6J ApoE KO mice, three out of four branches of the LCA (left external carotid, internal carotid, and occipital artery) were surgically ligated near the carotid artery, and a high-fat diet was administered for either 2 or 4 weeks. These mice were measured via MRI to monitor blood flow in the LCA and right carotid artery (RCA).
Before extracting carotids from the 2-and 4-week groups, blood was collected from the outlet of the aortic arch and measured using Raman spectroscopy. Figure 2 shows histopathology tissue cross-section and micro-MRI blood flow images of the RCA and LCA. In particular, Figure 2b clearly shows that the atherosclerosis is accelerated in the LCA, that the blood vessels are blocked, and that the reduced blood flow of the LCA is visible on the MRI. Based on the presence of LCA and RCA blood flow in the MR image just before carotid extraction, transparency of the extracted carotid, and cross-sectional histopathology of the LCA, mice were classified into three groups: wild-type C57BL/6J mice as controls (no LCA ligation), mice without abnormal findings (Figure 2a), and mice with significantly advanced atherosclerosis ( Figure 2b). Raman signals were analyzed to identify a diagnostic classifier between these groups. been diagnosed by measuring the plaque burden using intravascular ultrasound. 31,32 Figure 3a,c is Movat's pentachrome and hematoxylin/ eosin (H&E) stained images of the RCA and LCA cross-sections, respectively. The LCA plaque-induced turbulence grows internally, resulting in a plaque burden of >78%, as measured according to the plaque area. These plaques in the LCA primarily consist of foam cells (red arrow in Figure 3d) and macrophage-like cells, and the atherosclerosis is clearly accelerated compared with the RCA. In addition, Figure 3d shows the higher distribution of NF-kB in plaques formed through ligation-induced shear stress-based atherosclerosis induction. 33 For diagnostic criteria, mice with a plaque burden of 70% or more of the LCA section were selected for the atherosclerosis group, after which Raman spectroscopy was performed. In addition, for where plaque formation was not observed in the RCA but confirmed in the LCA, samples of shear stress or low flow-induced atherosclerosis through partial ligation were selected as the disease group. Therefore, to obtain diagnostic criteria, the Raman data were divided into "normal" and "mild disease," which constitute a carotid plaque burden of less than 40%, and "severe disease," which corresponds to a plaque burden of more than 70%. For mild disease data, almost no plaques were observed to form onto the internal elastic membrane, and the plaque burden was not distinguished by less than 40%, as in the controls. However, when Movat staining was performed, as shown in Figure 3e, a slight increase in fibrin/fibrinoid tissues (reddish) and collagen accumulation (yellowish) were observed in the LCA media area.

| Microscopic histopathology and immunochemistry of atherosclerotic mouse group
This parallels the analyses noted by previous studies in atherosclerotic samples, 34,35 and Figure 3f shows that this increase appears accelerated for severe disease. The samples selected as a mild disease in this study are representative of the early status of atherosclerosis. The mouse groups thus separated were physically induced to produce plaques from blood shear stress through ligation while feeding ApoE KO mice of the same lineage the same high-fat diet. Chemical additions were controlled so that chemical changes in the blood and changes in biomarkers occurring as plaques were formed could be monitored by Raman. To evaluate the signal difference between the filtered area and the original droplet area, Raman spectroscopy signals were measured using whole blood and centrifuged blood plasma drop. Figure 4g shows the comparison of the Raman signal inside and outside the droplet boundary, and it is shown that the signal rises dramatically in the outside area (nano biomarker filtrated area). In addition, because there seems to be a rare difference in signal between whole blood and centrifuged plasma, the SERS sensing chip used in this study does not require sample pretreatment such as centrifugation and heparin treatment. Because signal acquisition is possible in one drop, only a very small amount of blood is required compared to the amount required for centrifugation, so there is rare animal kill or damage, even in mouse experiments. Figure 4h shows the difference of the Raman signal by distance at the droplet boundary and shows no variation The difference between the inside and outside of the droplet boundary was plotted as the average and standard deviation for every 20 points, and the difference according to distance was the result for every 5 points. On the other hand, when the Raman signal was acquired in the whole blood-filtered area using a 532-nm laser, the Raman peak pattern disappeared due to strong noise caused by auto-fluorescence. Diagnostic criteria were prepared according to the projection in the PC1 direction, with a score of À10 or less denoting the region with mostly control samples, between À10 and 0, the region of mild disease samples, and 0 or more indicating severe disease. As shown in Figure 6d, the total accuracy according to these criteria was calculated to be 97.5%. The significance for each of the mild and severe disease F I G U R E 5 Raman signals by animal group and labeling by major peaks. Raman spectra of blood sampled from control group C57BL/6 wild-type mice (black), mice with mild atherosclerotic disease (blue), and mice showing severe atherosclerosis (red). Each solid-colored line shows the average value of the measured spectra, with lighter shades showing the standard deviation. Spectra were normalized based on the value at 1000 cm À1 (black dotted line). Black dots identify peaks suppressed as atherosclerosis progresses. Green bands indicate peaks that emerge as atherosclerosis progresses.

| Raman spectra and assignments in atherosclerotic blood
groups was expressed as a p value, and each p < 0.0001 was confirmed. In addition, the diagnostic validity according to PC1 was reviewed using the machine learning technique used in the application of diagnostic statistics for the recognition of Raman signals.   the parts of the Raman spectra that have significant correlations with diagnostic meaning. In Figure 6c, the eigenvectors for PC1 and PC2 are plotted as coefficient values according to the Raman energy shift.
In addition, the direction parallel to PC2 and positive is observed to be independent of the sample group boundary in Figure 6b. In Figure 6c, the spectral region that contributes the main value to the PC2 positive direction is indicated by translucent gray bars. This region extends up to 425 cm À1 and assumes a high value in the range of 710-980 cm À1 . While this range corresponds to a wide region of high spectral intensity in Figure 5, this area is a dead space with minimal diagnostic meaning, as shown by its low (and mostly orthogonal) value in the PC1 eigenvector.
In Figure 6b, the PC1 scale was normalized such that the normal group is divided at a value of À10, and the mildly and severely atherosclerotic disease groups are divided at 0, in the PC1 direction.
Therefore, the stronger the PC1 eigenvector and spectrum values in the positive direction, the closer the sample is to (severe) atherosclerosis. Table 1  (1305-1350 cm À1 ). The negative value of the eigenvector means the weight of the corresponding substance (peak intensity) for atherosclerosis diagnosis and is far from the increase or decrease of the absolute value of the substance. In addition, because the analysis was based on data normalized to 1000 cm À1 , the difference in relative values becomes the criterion for diagnosis.

| Validation of classification criteria using data from additional animal groups
Machine learning algorithms may be used to create diagnostic models.
However, owing to the risk of overtraining, these models cannot be considered meaningful until tested on an independent data set. The validity of the atherosclerosis diagnosis secured based on the eigenvectors and Raman assignments in Figure 6b was assessed using data from additional animal groups. Figure 8 shows physiological mechanisms allowing the endothelial cells to react to hemodynamic forces. 58,59 Partial carotid ligation is a novel method that enables the collection of endothelial-enriched RNAs and evaluation of the molecular mechanisms underlying flow-dependent regulation in vascular biology.
A previous study associated the disturbed flow site generated by partial carotid ligation with endothelial dysfunction, the regulation of pro-and anti-atherogenic genes, and rapid atherosclerosis progression. 28 In this study, atherosclerosis was selectively induced in the Committee (IACUC). Two mice groups were established: a control group of male C57BL/6 wild-type mice and treatment group of male C57BL/6 ApoE À/À mice, both purchased from Jackson Laboratory. In the high-fat diet of C57BL/6 mice, there were differences in HDL levels according to male or female. 23 In this study, it was fixed as male to minimize variation in blood biomarkers by sex. The generation and evaluation of mice carrying the inactivated mutant apolipoprotein E gene were confirmed. 61 At 8 weeks of age, the carotid artery was ligated in the treatment group to produce a disturbed blood flow (d-flow), which is known to lead to consistent and representative generation of arteriosclerosis in these mice. 26 at a beam voltage of 10 kV.

| Statistical analysis and machine learning validation for Raman spectral data
Blood sampled from the inferior vena cava of a random sampling of mice in each group (n = 5) was dropped in 5 μL quantities on SERS chips and left until spreading was complete. A Raman spectrometer (NOST) equipped with a Â40 objective lens (LUCPLFLN40X, Olympus, NA = 0.6, WD = 2.7-4.0) and 785-nm laser source passed through a 1% neutral density filter was used to illuminate and measure Raman spectra from the samples. Raman spectra were measured five times for 4 s each from 400 to 2400 cm À1 via a 600 groove grating. The spectral data pitch was 2 cm À1 for 1000 data points. Sample measurements were accumulated for 20 s, and five-polynomial fitting was then applied to remove background noise. To evaluate chip performance through whole blood and centrifugal plasma, 1 mL of mouse blood was placed in PST Tubes with Lithium Heparin (Becton and Dickinson), one drop was supplied to the SERS chip, and one drop of plasma was also dropped onto the SERS chip after centrifugation.
Blood plasma separation was performed in a centrifugation system (hanil M15R) at 2500Âg for 10 min at 4 C.
PCA was employed to identify differences in the Raman spectra of the control and atherosclerosis groups. The entire spectral range was used as a variable, and the analysis was conducted using XLSTAT

PEER REVIEW
The peer review history for this article is available at https://www. webofscience.com/api/gateway/wos/peer-review/10.1002/btm2.

DATA AVAILABILITY STATEMENT
The data has been included in the manuscript or Supporting Information. Additional generated or analyzed data are available from the corresponding author upon reasonable request.