2.1.

Whole Blood and Plasma Samples

The study was approved by the Research Ethics Committee of Universidade Paulista—UNIP (Process CAAE No. 67895617.5.0000.5512). Human blood samples were obtained from the laboratory of clinical analyses of a reference hospital for oncohematological diseases in São José dos Campos. The diagnosis of leukemia in the samples enrolled in the study was done with qualitative and quantitative analysis of the peripheral blood cells through hemogram, which results in a suggestive hematological disease, progressing to detailed exams, such as myelogram, immunophenotyping, cytogenetics, and molecular biology³¹ when needed.

The blood samples were collected from peripheral veins of each subject by vacuum-closed method in tubes containing K3EDTA (7.2 mg) as anticoagulant (Sarstedt AG & Co., Nümbrecht, Germany). An aliquot of each blood sample was transferred to a tube for mechanical centrifugation at 3500 RPM for 10 min to obtain the plasma (model Combate, CELM Ltd., Barueri, SP, Brazil); the remnant blood was maintained in the original tube.

It was evaluated 25 samples of whole blood from healthy subjects and 17 samples from whole blood from acute leukemic subjects identified after conventional diagnostic techniques. The samples of whole blood and plasma were properly conditioned in thermal boxes (2°C to 8°C) in order to avoid changes in the biochemical constitution due to temperature. At the time of spectroscopy, these samples were separated into the two groups named healthy group and leukemic group.

2.2.

Raman Spectroscopy

Raman spectra were obtained in whole blood and plasma samples without any preparation, by pipetting an amount of 80 μL in an aluminum sample holder with holes, using a single-channel micropipette of variable volume (model P200, Bio-Rad, Hercules, California). The spectra were obtained in a near-infrared Raman spectrometer (model Dimension P1, Lambda Solutions Inc., Massachusetts), which uses a diode laser at 830 nm coupled to a Raman probe fiber optic cable for sample’s excitation, obtaining 280 mW of laser power at the excitation output of the probe. The scattering of the sample was collected by the Raman probe and coupled to the spectrometer for dispersion. The spectrometer disperses the scattered light onto a back thinned, deep-depleted charge-coupled device camera ($1340\times 100\text{pixels}$, cooled to $-75°\mathrm{C}$) in the spectral range between 400 and $1800{\mathrm{cm}}^{-1}$, providing $\sim 2{\mathrm{cm}}^{-1}$ of spectral resolution. The exposure time for obtaining each spectrum was 3 s with 10 accumulations (30 s of total exposure time), and each sample was analyzed between three and five replicates in order to increase the number of spectra in the discrimination model. Neither noticeable damage to the blood or serum nor spectral change was observed during the spectrum acquisition.

The collected Raman spectra were subjected to preprocessing to remove the Raman background (mainly fluorescence from blood components and cells) by fitting and subtracting a seventh-order polynomial over the entire spectral range of 400 to $1800{\mathrm{cm}}^{-1}$. Spikes from cosmic rays were removed manually and then the spectra were normalized by the “area under the curve” (one-norm).³² The mean spectra of each group were calculated for visual comparison and statistics. Table 1 presents an overview of the number of spectra collected in each of the groups. Four blood samples from the healthy group and two plasma samples from the leukemic group were excluded from the study due to the presence of hemolysis and low Raman signal-to-noise ratio identified after preprocessing.

Table 1

Number of samples and number of spectra in each group.

The mean spectra were plotted in order to identify visual differences in the intensities of the Raman peaks between the healthy and leukemic groups. The most intense peaks in both spectra were labeled and the Raman band positions were tabulated to determine the composition of whole blood and plasma based on the Raman features. Student’s $t$-test with significance level of 5% ($p<0.05$) was applied to the intensities of the peaks between healthy and leukemic groups to determine the peaks with significant differences, in a way to determine the differences in the biochemical composition between healthy and leukemic blood and plasma and to identify the differences in the biochemical profile of both groups. The $t$-test is used to compare the means when the samples follow normal distribution but with unknown variance, and the $p$-value is used to accept ($p$-value $>0.05$) or reject ($p$-value $<0.05$) the null hypothesis (equality in the mean between the two populations), being accepted the alternative hypothesis that the means come from different populations. In this particular study using the intensities of the Raman peaks, as the Raman peaks may present intensities higher or lower when comparing healthy and leukemic groups, the two-tailed $t$-test was used. Thus, the $t$-test can be used to make decisions based on statistical calculations with a high degree of confidence.³³

2.3.

Exploratory Analysis and Discrimination

3.1.

Whole Blood and Plasma Spectra

In this study, 312 spectra were collected from 42 individuals (25 healthy and 17 leukemic), being 142 spectra originated from whole blood (being 80 classified as healthy and 62 as leukemic) and 170 spectra originated from plasma (being 101 classified as healthy and 69 as leukemic), as shown in Table 1.

Figure 1 presents the mean Raman spectra of whole blood from healthy and leukemic groups. The spectrum from whole blood of the healthy group shows peaks in the positions of the blood constituents: cellular components (mainly leukocytes, erythrocytes, and platelets) and noncellular components (mainly plasma). Table 2 presents the positions of the main peaks of whole blood accompanied by their respective assignments as described in the literature^{24,25,30,42,43} and organized by biochemical groups. By visual inspection, some peaks showed small but perceptible differences in intensity and bandwidth between healthy and leukemic, suggesting some relevant variation in these tissue components among the samples. The $t$-test ($p<0.05$) applied to identify peaks with statistically significant differences in the healthy versus leukemic groups is presented in Table 2. The higher differences were observed in the peaks at the positions of 570, 678, $755{\mathrm{cm}}^{-1}$, peaks in the range between 820 and $920*{\mathrm{cm}}^{-1}$, and peaks at 1004, 1130, 1160, 1212, 1225, 1344, 1378, 1401, 1452*, 1552, 1567, 1586, 1623, 1640, and $1658*{\mathrm{cm}}^{-1}$ (the symbol * denotes peaks in which the leukemic group shows greater intensity than the healthy group). Peaks at 1212, 1401, and $1640{\mathrm{cm}}^{-1}$ did not present statistically significant differences between the two groups ($p>0.05$), suggesting that they do not aid in the differentiation between healthy and leukemic. As noted in Fig. 1, all the peaks in the spectrum of the healthy group are more intense when compared to the leukemic group, with the exception of the peaks at 820 to 920, 1452, and $1658{\mathrm{cm}}^{-1}$, which present greater intensity in the leukemic group.

Fig. 1

Mean Raman spectra of whole blood from subjects in the healthy and leukemic groups and the spectrum of the difference between leukemic and healthy. The symbol * represents the peaks that are more intense in the spectra of leukemic than in the spectra of healthy.

Table 2

Grouping of the main Raman peaks of whole blood and plasma according to the biochemical constitution, assignments according to the published literature,24,25,30,42,43 and statistically significance (p-value) of the peak’s intensities between healthy and leukemic.

Figure 2 presents the mean Raman spectra of blood plasma from healthy and leukemic groups. The spectrum from plasma of the healthy group shows peaks in the positions of plasma constituents, mainly proteins (albumin, globulins, and amino acids), lipid fractions, carbohydrates (glucose), carotenoids, and metabolites. Table 2 presents the positions of the main peaks of plasma peaks accompanied by their respective assignments, as described in the literature^{24,25,30,42,43} and organized by biochemical groups. By visual inspection, there were no relevant differences in the intensities of the peaks of the healthy versus leukemic groups; however, significant differences were observed in the peaks at 510, 721, 760, 837, 947, 1004, 1132, 1160, 1210, 1269, 1334, 1344, 1407, 1448, 1455, 1525, 1630, 1659, and $1666{\mathrm{cm}}^{-1}$ ($p<0.05$). Due to the lower composition complexity of the plasma compared to the whole blood, the amount of peaks with statistically significant differences is lower in plasma than in whole blood, since in this specimen, the peaks referring to leukocytes, erythrocytes, and platelets are absent.

Fig. 2

Mean Raman spectra of plasma from subjects in the healthy and leukemic groups and the spectrum of the difference between leukemic and healthy.

3.2.

Exploratory Analysis

In the exploratory analysis, the PCA technique has been used to identify the spectral features that presented differences between the groups, through the PCs and SCs. Interpretation consists of identifying the peaks present in the first PCs, and correlating these peaks with the biochemical compounds present in each group according to the literature. The scores are then used to “quantify” these compounds in each of the healthy and leukemic groups. Positive peaks with positive scores, as well as negative peaks with negative scores, show that the specific biochemical is at a high concentration in that particular group, while positive peaks with negative scores and vice-versa show that the biochemical components attributed to that PC are presented in lower concentrations.

Figure 3 shows the plot of the PCs and SCs of the whole blood samples. The PC1 has characteristic peaks of whole blood, but the leukemic group presents a lower score of this loading (SC1) ($p<0.001$), suggesting that despite the same constitution in terms of Raman features, the leukemic group presents these constituents in lower concentrations. PC2 shows peaks with negative intensities in the leukemic group and positive peaks in the healthy group, with a statistically significant difference in the SC2 ($p<0.01$). This loading has positive peaks at 752, 1213, 1524, 1547, and $1619{\mathrm{cm}}^{-1}$ and negative peaks at 1007, 1381, 1403, and $1643{\mathrm{cm}}^{-1}$, being the positive peaks assigned to proteins, amino acids and carotenoids, and the negative peaks related to amino acids, carbohydrates, and lipids. The positive features suggest the highest protein/amino acid concentration for the healthy group, mainly by observing peaks at 752, 1213, and $1547{\mathrm{cm}}^{-1}$, which coincide with the peaks of the whole blood, especially erythrocytes.²⁷ Also, the presence of highest carotenoid concentration in the healthy has been evidenced. Still, PC3 presents peaks with negative intensities in the leukemic group and positive peaks in the healthy group, with significant difference for SC3 ($p<0.001$). This component has positive peaks at 570, 678, 756, 1142, 1226, 1381, 1404, 1501, 1569, 1624, and $1643{\mathrm{cm}}^{-1}$ and negative peaks at 1003, 1212, 1451, 1544, and $1664{\mathrm{cm}}^{-1}$, being the positive peaks assigned to proteins, amino acids, and carbohydrates, and negative peaks referred to amino acids and lipids. The SC3 indicates that the healthy group has higher concentrations of proteins (amide bands I and III and glutathione), amino acids (Trp and Tyr), and glucosamine, assigned to the biochemical components of whole blood. The leukemic group, with negative SC3, shows negative peaks related to amino acids (Phe and Trp) and lipids (cellular phospholipids), which indicates the presence of these compounds in higher concentrations in leukemia and suggests hypercellularity caused by blasts (white cells, particularly granulocytes).³⁰ Although PC4 presented peaks in positions related to proteins and amino acids, SC4 did not present significant difference of healthy versus leukemic groups ($p>0.05$). Therefore, PC2 and PC3 indicated the major differences, mainly higher concentration of red blood cells in healthy group and higher concentration of white cells in the leukemic group.

Fig. 3

Plots of the PCs and SCs calculated for the whole blood samples for exploratory analysis. The peaks marked with “*” represent peaks found in this study which do not present known assignment.

The hypercellularity due to the presence of blasts in the leukemias causes an increase in the metabolic rate of the cells, resulting in greater consumption of amino acids and proteins that participate in specific enzymatic actions, such as the tyrosine kinase. This enzyme has the function of phosphorylation of protein substrates, such as the myeloperoxidase, an enzyme present in some types of acute myeloid leukemias.^44–46 This may explain the fact that PC2 and PC3 exhibit some peaks related to amino acids with higher intensities in the healthy compared to leukemic groups, especially Tyr and Trp. The peaks of proteins and glutathione (linked to erythrocytes) are also found in lower intensities in the leukemic group, which was also observed by Sanches et al.,⁴⁷ who carried out a study comparing the biochemical profile between leukemic patients and healthy individuals. The peak assigned to the amino acid Phe was highlighted in the first three loadings. Although in low amounts, the literature describes that this amino acid is one of the activators of the BCR-ABL gene, which present in a class of chronic myeloid leukemia and other myeloproliferative disorders.⁴⁴ Its presence in the first loadings would suggest that Phe may play an important role in acute leukemias.

Figure 4 shows the plot of the PCs and SCs of plasma samples. The PC1 has characteristic peaks of plasma, but there was no statistically significant difference in SC1 for healthy versus leukemic groups ($p>0.05$); therefore, such peaks are not relevant for the differentiation of the groups studied. PC2 shows negative peaks at $1630{\mathrm{cm}}^{-1}$ in the leukemic group and positive peaks at 947, 1004, 1159, 1348, 1407, 1453 and $1520{\mathrm{cm}}^{-1}$ in the healthy group, with a significant difference for SC2 ($p<0.01$). These peaks have assignments related to proteins, amino acids, lipids, and carotenoids. The low intensity of SC2 suggests that the difference in the concentration of these constituents is small, but significant in the groups ($p<0.01$), being that the healthy group has higher concentrations of proteins, Trp, Phe, glutathione, phospholipids (free), and carotenoids than the leukemic group. PC3 shows peaks with negative intensities in the leukemic group and positive in the healthy group, with a significant difference for SC3 ($p<0.05$). This component has positive peaks at 1003 and $1433{\mathrm{cm}}^{-1}$ and negative at 1007, 1159, 1344, and $1527{\mathrm{cm}}^{-1}$, assigned to proteins, amino acids, and carotenoids. The low intensity of SC3 suggests that the biochemical elements of these peaks are not useful for the differentiation between the healthy group and leukemic group. PC4 showed a statistically significant difference in the intensities of SC4 ($p<0.001$), with intense negative peaks at 897, 959, 1004, 1160, 1447, and $1525{\mathrm{cm}}^{-1}$, referring to the group of proteins, amino acids, and carotenoids, suggesting a higher concentration of these components in the healthy group, and a positive peak at $1375{\mathrm{cm}}^{-1}$ attributed to glucosamine, which presents in a low concentration in the healthy group compared to the leukemic group.

Fig. 4

Plots of the PCs and SCs calculated for the plasma samples for exploratory analysis.

Regarding the carotenoid peaks in plasma, the exploratory analysis by PCA showed that both PC2 and PC4 presented peaks of this component (1004, 1159/1160, and $1520/1525{\mathrm{cm}}^{-1}$),^24,30 and SC2 and SC4 indicated significant differences, indicating the presence of these components in high concentrations in the healthy group, as demonstrated by González-Solís et al.,²⁴ when analyzing Raman spectra of blood in leukemic subjects. The literature has shown that healthy individuals have a high plasma carotenoid concentration and that carotenoids act as protectors against neoplasms (antioncogenic), such as acute leukemias.^48,49 Another study showed that carotenoid concentration increased in leukemic individuals in remission.²⁴

3.3.

Discriminant Analysis (PLS-DA and PCA-DA)

Discrimination techniques based on PLS and PCA (PLS-DA and PCA-DA, respectively) applied to spectral data have been used with relative success as predictors of discrimination or differentiation between healthy and diseased tissues, especially in oncology.^37,50,51 The DA using PLS and PCA was applied to the normalized spectra of the healthy and leukemic groups using the entire spectral range (400 to $1800{\mathrm{cm}}^{-1}$) for both whole blood and plasma samples. The “leave-one-out” cross-validation method defined an initial condition of 10 latent variables (PLS-DA) and 10 PCs (PCA-DA) to be modeled.³⁹

The results of the discrimination models using the spectra of whole blood and plasma from each sample group were tabulated (confusion table, Table 3), presenting sensitivity, specificity, and overall accuracy values.⁵² For the whole blood, using the PLS-DA, the maximum accuracy occurred by using the first three latent variables, with a value of 96.5% success classification and 91.9% sensitivity, while the maximum accuracy using PCA-DA occurred by using the first four PCs, with a 93.0% success classification and 83.9% sensitivity. Both discriminant models reached 100% specificity, showing that healthy individuals presented a spectral profile capable of allocating them to the healthy group independently on the model used. For the plasma, using the PLS-DA, the maximum accuracy occurred using the first four latent variables, with a 97.1% of success classification, 95.7% sensitivity, and 98.0% specificity, while the accuracy using PCA-DA occurred using the first PC, with 64.1% of success classification, 11.6% sensitivity, and 100% specificity.

Table 3

Confusion matrix with the results of sensitivity, specificity, and correct classification for the discrimination model using the Raman spectra of whole blood and plasma samples.

Figure 5 shows the confusion plot with the resulting PLS-DA and PCA-DA discrimination for whole blood and plasma, respectively, as shown in Table 3. The sensitivity, specificity, and accuracy values for the PLS-DA were effective in both types of samples, being the results for the correct classification for both whole blood and plasma samples close to each other (around 97%). In general, the results for both PLS-DA and PCA-DA models were more favorable to the whole blood samples (96.5% and 93.0%, respectively), defining this sample as a good option for the differentiation of the leukemic from the healthy group, with spectral variables related to the red blood cells and white cells which allowed to obtain a reduced set of predictors with greater potential for success in the discrimination.

Fig. 5

Confusion plot with the classifications by healthy group and leukemic group, through the DA, being (a) PLS-DA (whole blood); (b) PCA-DA (whole blood); (c) PLS-DA (plasma); and (d) PCA-DA (plasma).

In the PCA-DA, the results obtained in the whole blood were high, demonstrating that the presence of blood cells had relevance in the description of the differences in the groups by the PCs, while the lower biochemical constituents of the plasma reduced the sensitivity of the PCA model, resulting in an accuracy of 64.1%. Gonzáles-Solís et al.²⁴ used RS in the serum to monitor patients with acute leukemia under chemotherapeutic treatment, and through PCA and LDA, successfully identified sample groups belonging to healthy and leukemic individuals, with significant differences between biochemical components, such as proteins, amino acids, and carotenoids, which resemble the results of the research presented here.

The use of multivariate analysis with the objective of identifying and classifying sample groups in the most diverse areas of knowledge is well known.^{24,34,35,37,42,51} The PCA aims to perform the segregation of groups based on the variances that, maximized, support the classification of samples according to the differences between the groups, and has shown to be an important tool when the main information capable of differentiating the sample groups is just the variability between groups, which needs to be greater than the intragroup variability.^38,39 However, the PLS-DA stands out from the PCA-DA because, in addition to the information of the differences between the groups, the variances obtained within each group are also recognized, and these variances are associated to the groups when modeling the regression curve;^38–40 therefore, the PLS-DA leads to better performance in the classification of samples when compared to PCA-DA.^38,39

Compared to the techniques currently available for the diagnosis of acute leukemias based on cytomorphology and immunophenotyping,^9–11,13,14 RS model presented accuracy (97.1%) close to phenotyping (94%¹³) and sensitivity and specificity (95.7% and 98.0%) close or overmatch the flow cytometry technique (97% and 88%¹⁴), despite the difficulty in obtaining the sensitivity and specificity reference values for the standard diagnostic techniques of myelogram and phenotyping. Thus, Raman-based analysis could significantly differentiate between healthy and leukemic individuals based on differences in the spectral pattern related to the chemical composition (amino acids and carotenoids) of the blood plasma, and cellularity (red and white cells) and chemical composition (proteins) in the whole blood using a small volume of peripheral blood. This infers a minimally invasive character when obtaining the samples to be diagnosed.

RS does not require the use of reagents or special preparation of the sample, spectral analysis can be done in reduced time and mathematical and statistical algorithms can lead to a simplification in the analysis and interpretation of the results. The RS technique can be included in the list of diagnostic screening tests, since its high specificity makes it possible to exclude the suspicion of acute leukemia, allowing physicians to extend the diagnostic investigation to other diseases, defining the therapeutic conduction on their patients early, since the time between diagnosis and treatment is crucial for a good prognosis. The technique can benefit from compact Raman systems that resemble bedside instrumentation for the point-of-care use,^53,54 where blood and serum samples recently withdrawn can be evaluated rapidly and nondestructively, with advantages when compared to established screening techniques, such as blood count and myelogram. The Raman technique may establish a new technology for rapid and accurate diagnosis of complex diseases related to human blood including leukemias.