2.1.

Participants

Eighteen AD affected subjects (7 women, 11 men, and mean age $71\pm 10$ years) and seven MCI subjects (3 women, 4 men, and mean age $73\pm 9$ years) were included in this study. All met the core clinical criteria for AD and MCI of the National Institute on Aging–Alzheimer’s Association.^32,33 The control (Ctr) group was constituted by six healthy volunteers (3 women, 3 men, and mean age $71\pm 9$) without evidence of pathologies or of familial neurological diseases. All participants underwent a complete ophthalmological examination including the best-corrected visual acuity test. The demographic and clinical characteristics and the Mini-Mental State Exam (MMSE) scores of the recruited subjects are shown in Table 1. The study was approved by the Institutional Review Board of the University of Naples “Federico II” and all investigations adhered to the tenets of the Declaration of Helsinki (2013). Written informed consent was obtained from all the enrolled subjects.

Table 1

Demographic and clinical characteristics of healthy (Ctr), MCI, and AD subjects. Data expressed as mean±standard deviation. MMSE stands for minimental state examination.

2.2.

Tears Collection

Tear specimens were collected from the informed subjects and healthy volunteers. Smooth edge sterile capillary glass tubes, with internal diameter $d_{\mathrm{in}}=0.2\mathrm{mm}$, outside diameter $d_{\mathrm{out}}=1\mathrm{mm}$, and length $L=12.7\mathrm{mm}$, were used (ACCU-FILL 90 MICROPET by Becton, Dickinson and Co, Clay Adams, California, USA). During the collection, the subject’s lower eyelid was gently pulled down, and the tip of the open capillary tube was placed in contact with the tear meniscus without irritating the conjunctiva. At the least $1.0\mu \mathrm{l}$ tear was collected. The samples were stored at 4°C and tested by SERS within 7 days. Protein concentrations are not altered under these storage conditions.³⁴

2.3.

Surface-Enhanced Raman Spectroscopy

SERS measurements were performed with standard microscope glasses coated by a home-made colloid of gold nanoparticles (GNPs). Details of the fabrication and characterization of the GNPs are reported in Ref. 35. The average diameter of the GNP was 27 nm. About $1\mu \mathrm{l}$ of tears to be analyzed was dropped on the dried residual of the GNP colloid and measured a few minutes after drying. SERS signal spectra were acquired using a Jobin-Yvon instrument by Horiba Scientific ISA (Osaka, Japan) equipped with a TriAx 180 monochromator, a liquid-nitrogen cooled charge-coupled detector, and an 1800 grooves/mm grating (final spectral resolution: $2{\mathrm{cm}}^{-1}$). The spectra were recorded in the air at room temperature using a 17-mW-nominal power He–Ne laser source (wavelength 632.8 nm). The laser beam was focused on a $2\text{-}\mu {\mathrm{m}}^2$ spot area of the sample surface through an Olympus microscope equipped with $100\times$ optical objective. The laser power impinging on the sample is expected to be low enough to avoid significant heating effects, as demonstrated in preliminary measurements performed to optimize the experimental conditions. We experimentally verified that heating effects are negligible by measuring Stokes/anti-Stokes Raman response of silicon in similar acquisition conditions to the ones adopted in the present work (see, for example, Refs. 36 and 37). This is still true for tears even if their thermal conduction properties are quite different from those of silicon.³⁶ All SERS spectra were acquired using an accumulation time of 180 s. At least three SERS spectra were measured in different points of the sample area and averaged. For each sample, the average signal intensity dispersion for each wavenumber point was about $8\pm 5\%$. A representative spectrum for each class was obtained by averaging all SERS acquisitions from related samples.

2.4.

Data Analysis

3.1.

SERS Measurements

An example of SERS response of tears of a healthy subject is shown in Fig. 1(a) along with a representative spectrum of the same subject’s tears obtained by conventional Raman spectroscopy in the same experimental conditions [Fig. 1(b)]. The Raman spectrum does not provide any valuable information apart from very faint features. Conversely, SERS spectrum clearly evidences the various contributions of tear components. The main modes indicated in the Fig. 1(a) were individuated by a deconvolution of the spectrum in terms of Lorentzians. Raman features typical of proteins are observed in the SERS spectrum shown in Fig. 1(a), where the skeletal (primary protein structure) region, the amide I, amide II, and amide III bands can be recognized. The positions of the most relevant peaks are reported in the figure, and tentative assignments of the modes identified by the deconvolution process are reported in Table 2 together with related references. In particular, three strong peaks are found at 998, 1349, and $1170{\mathrm{cm}}^{-1}$ and assigned to phenylalanine (Phe), tryptophan (Trp), and tyrosine (Tyr), respectively.⁴⁴ The intense Trp mode may primarily be due to lysozyme (LZ), one of the major components of tears.^22,45

Fig. 1

(a) SERS spectrum of tears from a healthy subject. The acquisition time was 180 s. The experimental spectrum (black curve) was fitted by a convolution of Lorentzian functions. The gray curve is the resulting fit curve. The main modes found are indicated. (b) Spectrum of tear obtained by conventional Raman spectroscopy in similar acquisition conditions.

Table 2

SERS of human tears: assignment of main modes according to Refs. 23 and 44 (def, deformation; wag, wagging; str, stretching; bend, bending; sciss, scissoring).

SERS peaks at about 1256 and $1428{\mathrm{cm}}^{-1}$ are assigned to LZ or lactoferrin (LF) in agreement with conventional Raman spectroscopy results for single-component solutions of LF and LZ,^22,45 and with SERS response of LZ.⁴⁵ These tentative assignments are in agreement with previous studies showing that the major components of tears are immunoglobulins (IgA), lactoferrin, lysozyme, lipocalin, and albumin.^17,46 Among these components, lactoferrin and lysozyme have an important role in the eye functionality; indeed they provide defense mechanisms against infective agents.⁴⁷ Figure 2 shows the average SERS spectra for the three classes of subjects here considered (Ctr for control/healthy subjects; MCI for subjects suffering from MCI; AD subjects suffering from AD). The average relative dispersion of the signal intensity is $24\pm 17\%$ for the control set, $21\pm 14\%$ for the MCI data, and $37\pm 8\%$ for the AD data. The three spectra in Fig. 2(a) share some common features even if differences can be noticed. Spectra obtained by subtracting the control spectrum from the MCI and AD signals are reported in Fig. 2(b). In some spectral regions, the intensity variation is larger than the data dispersion. In particular, statistical deviations were determined by one-way ANOVA test and are indicated in Fig. 2(b) for AD and MCI signals. Positive deviations are found at 1007 and $1591{\mathrm{cm}}^{-1}$ and a negative deviation at $1459{\mathrm{cm}}^{-1}$ for the data related to the AD class. A positive deviations at $1563{\mathrm{cm}}^{-1}$ and negative deviations at 850 to 885, 996, 1244, 1343, and $1650{\mathrm{cm}}^{-1}$ are determined for the data of MCI class. Two large intensity deviations of MCI signal occur at 1063 and $1500{\mathrm{cm}}^{-1}$ but they were be not further considered being close to minima of the SERS spectrum.

Fig. 2

(a) Averaged SERS spectra of tear samples from healthy subjects (Ctr-green line), mild cognitive disease-affected subjects (MCI-blue line), and AD-affected subjects (AD-red line). The gray areas represent the standard deviation of the signal intensities within the considered data classes. Each of the reported spectra was obtained by averaging all SERS spectra of subjects belonging to a single class. SERS spectra were obtained by 180 s long acquisitions. The average relative dispersion of the signal intensity is $24\pm 17\%$ for the control set, $21\pm 14\%$ for the MCI data, and $37\pm 8\%$ for the AD data, respectively. (b) Signal differences concerning the control data of AD (red area) and MCI (blue area) spectrum. The green lines indicate the signal dispersion range (0.68 of the standard deviation). The statistically significant signal differences ($p$-value $<0.05$ in the one-way ANOVA statistics) are indicated by blue (MCI) and red (AD) marks, respectively.

The occurrence of an intensity increase or decrease of the SERS signal of AD and MCI samples with respect to the Ctr sample-related spectrum is indicated in Table 2 by $\uparrow$ and $\downarrow$ symbols, respectively. It is noteworthy to note that SERS spectra from MCI-affected subjects exhibits a signal intensity lower than the reference spectrum (from healthy subjects) for modes at 1343 and $1243{\mathrm{cm}}^{-1}$. These modes are assigned to the C-H deformation and amide III $\beta$-sheet, respectively.²⁴ Their intensity change indicates a modification of protein secondary structures that can result from altered physiological conditions. Kim et al.²⁴ reported a correlation between the $I_{1342}/I_{1243}$ ratio of SERS modes at 1243 and $1342{\mathrm{cm}}^{-1}$ and eye disease state in human tears. They found an increase of the $I_{1342}/I_{1243}$ ratio in subjects affected by adenovirus ($I_{1342}/I_{1243}=1.13$) and herpes simplex ($I_{1342}/I_{1243}=3.73$) versus a control value $I_{1342}/I_{1243}=0.8$. In the present case, the mean value estimated for $I_{1342}/I_{1243}$ is $1.1\pm 0.4$ for healthy subjects, $1.5\pm 0.7$ for samples from AD class, and $1.8\pm 0.7$ for tear samples from MCI class. These average ratio values are larger in AD and MCI cases than in control; however, the data dispersion is too large to find a significant correlation.

Figure 3 shows a comparison between the average SERS spectra of healthy (Ctr), MCI- and AD-affected subjects for the amide III (1100 to $1400{\mathrm{cm}}^{-1}$) region together with their deconvolution in terms of Lorentzian functions. The three main components in the Ctr spectrum occur at 1173, 1248, and $1346{\mathrm{cm}}^{-1}$. These modes are compared with the equivalent components of MCI and AD spectra. The mode found at $1173{\mathrm{cm}}^{-1}$ in the Ctr spectrum and assigned to tyrosine is centered at a lower wavenumber value (about $1165{\mathrm{cm}}^{-1}$) in the MCI and AD spectra. The level of tyrosine is expected to decrease in intensity during the process of amyloid fibril formation,⁴⁸ and the observed weakening of the mode energy indicates a degradation of protein structure. The average SERS intensity of the MCI spectrum in the 1244 to $1350{\mathrm{cm}}^{-1}$ range is lower than the intensity in the other two cases. The AD spectrum deconvolution is characterized by relatively broader modes, which indicate an increased disorder of the protein secondary structure. A shift to lower energies of SERS modes is also observed in the MCI and AD spectra of Fig. 4, in which the amide I and amide II band regions are reported with the result of the signal deconvolution in terms of Lorentzians. The amide I band of the spectrum (Ctr) from healthy subjects is centered at $1639{\mathrm{cm}}^{-1}$. This band is shifted to 1625 and $1628{\mathrm{cm}}^{-1}$ in MCI and AD cases, respectively. These changes could be due to an increase contribution of the random coil component and a decrease of the intensity of the $\alpha$-helix-related mode. The overall effect of the combination of these two intensity changes is to move the center of the amide I band toward lower wavenumber shift values. It is noteworthy that also hydration causes a faint change of $\alpha$-helix contribution as reported in a Raman study on lysozyme.⁴⁹ It is interesting to note that the intensity of the amide I band is lower than the control data only in the MCI case. All the observed changes suggest a modification of the secondary structure of the proteins stronger in MCI samples than in AD samples. Notwithstanding the evidence that the changes in the tear composition induced by different physiological conditions influence their SERS response, it is hard to obtain a clear indication on the occurrence of AD and MCI diseases due to the presence in the samples of confounder proteins having strong spectral similarities with disease markers as amyloids and tau protein.⁵⁰

Fig. 3

Comparison of averaged SERS spectra of tears from AD-affected (red curve), MCI-affected (blue curve), and Ctr healthy (green curve) subjects in the amide III region. All SERS spectra were obtained by 180 s long acquisitions. The mode peaks resulting from the fitting spectrum deconvolution are reported together with the experimental data. Thick curves indicate selected modes featuring larger differences with respect to the control spectrum (see text).

Fig. 4

Comparison of averaged SERS spectra of tears from AD-affected (red curve), MCI-affected (blue curve), and Ctr healthy (green curve) subjects in the 1500 to $1700{\mathrm{cm}}^{-1}$ range including amide II and amide I bands. All SERS spectra were obtained by 180 s long acquisitions. The mode peaks resulting from the fitting spectrum deconvolution are reported together with the experimental data. Thick curves indicate selected modes featuring larger differences with respect to the control spectrum (see text).

3.2.

Patient Clinical State Recall Using i-PCA Results

SERS spectra of tears reported herein are very complex as expected for these body fluids. In addition, the differences among SERS spectra of tears from subjects with different clinical states are sometimes very subtle and are mainly localized in some specific spectral ranges. Notably, we found that PCA performed on the whole spectral range did not differentiate the subtle changes that occurred in the examined spectra. To better evidence these differences and to profit from them to classify the spectra, we performed the PCA in the intervals reported in Sec. 2.4.2 (i-PCA). The scores of the first seven PCs obtained for all the intervals were examined. The score plots for the single intervals do not suggest simple grouping related to the state of dementia. As examples, the i-PCA results for some of the intervals are reported in the Supplemental Material in terms of score plots and loadings of the first two components (see Fig. S2 in the Supplemental Material). This finding suggested the need for a complex model for classification. Accordingly, the score coefficients obtained by i-PCA are used to build models enabling them to obtain the correct classification of the spectrum to one of the dementia disease classes of the corresponding subject. The classification models were built using the naïve Bayes classification algorithm,⁴³ according to the procedure described in Sec. 2.4. The ability of the various models to obtain the correct classification of the average spectrum to one of the disease classes was barely evaluated by considering the overall accuracy in order to prove that the model itself is able to recall the correct disease class for all the considered spectra. To obtain a 100% accuracy in recalling the disease class of the spectra dataset used to build the model by using the SERS spectra of tears, we built a model using the scores of PC3, PC4, PC5, and PC6 calculated in the intervals 456 to 553, 707 to 755, 1202 to 1307, 1308 to 1353, 1354 to 1383, 1552 to 1582, 1583 to $1639{\mathrm{cm}}^{-1}$, which enabled us to obtain the correct recall of all the spectra (i.e., 100% of overall accuracy and 0% of overall misclassification rate) (representative outcomes of the i-PCA in terms of score higher-order score components are reported in Fig. S3 in the Supplemental Material). Since the problem is a multiclass classification (i.e., a classification procedure involving more than two classes) (see Ref. 51 and references therein) because three clinical states are considered and the dataset is small and unbalanced (6, 7, and 18 patients belong to the Ctr, MCI-affected, and AD-affected disease classes, respectively), it was not possible to perform further studies using standard validation procedures for the model. Nevertheless, these results are encouraging since they confirm the ability of correctly recovering the clinical state of patients by analyzing tears SERS spectra, thus offering proof of the potential of the method that could be usefully applied for classifying spectra acquired from new enrolled patients. These results boost further studies involving more and properly selected patients, and SERS spectra that will allow to test the efficacy of the use of different intervals for i-PCA, and/or the use of different classification methods including neural network algorithms.