2.1.

Raman Microscope

A Renishaw inVia micro-Raman spectroscopy system (the United Kingdom) with the laser excitation wavelength of 785 nm was used to collect the Raman spectra of liquid blood specimens.^20,21 The incident light (785-nm diode laser) was focused into the sample and the backscattered Raman signal was collected with a Leica microscope with $50\times$ objective. The laser power incident on the sample was about 0.6 mW, which is safe for the blood samples. The wave number was calibrated with silicon whose Raman spectral peak is at $520{\mathrm{cm}}^{-1}$. The spectra were measured with an integration time of 10 s. The Raman spectra at the same sample position were accumulated seven times. The wave number range of the Raman spectra for blood samples was 300 to $1700{\mathrm{cm}}^{-1}$. 785 nm was chosen as the excitation wavelength because more characteristic Raman peaks can be acquired.²⁰

2.2.

Blood Samples

Human blood samples were obtained from the volunteers in Suzhou and animal blood samples obtained from dog, rat, and rabbit were provided by Laboratory Animal Center of Soochow University. All experiments were performed in compliance with the law. All the blood samples were contained in vacuum blood collection tubes with anticoagulants (ethylenediaminetetraacetic acid dipotassium salt dihydrate) and then preserved in icebox immediately after they were acquired. The spectra were obtained between 24 and 72 h after the samples were acquired. For each liquid blood sample, $\sim 20\mu \mathrm{L}$ was placed on an aluminum-covered microscope slide for detection. The aluminum cover was used to reduce the fluorescence background.

2.3.

Principle of Self-Reference Algorithm

The intensity of Raman scattering signal is decided by the intensity of excitation light of specific wavelength, the property and quantity of the sample, and the sensitivity of spectrometer. This relationship has the following form²²

The traditional single intensity identification method can be used to identify different materials with different Raman peaks. But for materials with similar Raman spectra, single intensity identification method is incapable of action, because the Raman intensity is related to many factors such as the intensity incident on the sample and the amount of sample. To solve this problem, we introduce a method defined as the self-reference algorithm. In this method, two Raman shifts are used to discriminate the human and nonhuman blood. The one named as reference Raman peak is used to eliminate the influence of the interference factors. The other named as identification Raman peak is used to demonstrate the essential difference of human and nonhuman blood. The ratio of the identification and reference peaks’ intensities is calculated as

3.1.

Necessity of Self-Reference Algorithm for Analysis of Raman Spectra of Human and Nonhuman Blood

The average Raman spectra of human and nonhuman’s blood are shown in Fig. 1. Even though there is a difference between the intensities of human and nonhuman blood, the Raman shifts of human and nonhuman blood are almost the same. Therefore, discrimination of human and nonhuman blood only according to the original Raman spectra is obviously not suitable.

Fig. 1

The average Raman spectra of human and nonhuman blood.

The Raman spectra were preprocessed with the following procedures: (1) smooth the signal with Savitzky–Golay algorithm and (2) remove the baseline of the smoothed signal with adaptive iteratively reweighted penalized least squares (airPLS).^23,24 Figure 2 shows the normalized average intensity peaks of nonhuman and human blood at $676{\mathrm{cm}}^{-1}$ (pyrrole symmetric bending), $1003{\mathrm{cm}}^{-1}$ (aromatic ring breathing), and $1449{\mathrm{cm}}^{-1}$ (${\mathrm{CH}}_2$ and ${\mathrm{CH}}_3$ bend) are approximately equal. The average intensity from $1310{\mathrm{cm}}^{-1}$ to $1342{\mathrm{cm}}^{-1}$ is obviously different. Therefore, we decided to make an evaluation on the discrimination of human and nonhuman blood using the intensity peaks of $1341{\mathrm{cm}}^{-1}$. Since the Raman shift at $1341{\mathrm{cm}}^{-1}$ is corresponding to C–H bend in tryptophan,^25,26 we could speculate that human blood contain larger amount of tryptophan than nonhuman blood.

Fig. 2

The normalized average spectra of human and nonhuman blood.

Figure 3 shows the intensity at $1341{\mathrm{cm}}^{-1}$ of each blood sample. Samples from 1 to 35 are human blood and the ones from 36 to 79 are nonhuman blood, respectively. The human blood’s mean value of intensity at $1341{\mathrm{cm}}^{-1}$ is 1536.8 and the root mean square (RMS) is 70.5. The nonhuman blood’s mean value of intensities at $1341{\mathrm{cm}}^{-1}$ is 1604.1 and the RMS is 282.8, which is relatively large. Therefore, even though the average intensities at $1341{\mathrm{cm}}^{-1}$ of human and nonhuman blood are different, the discrimination of human and nonhuman blood according to the single intensity peak is still difficult. The intensity peak of nonhuman blood varies large. This large RMS induces that the intensity peak at $1341{\mathrm{cm}}^{-1}$ alone is not suitable for the discrimination of blood specimens. As mentioned in Sec. 2.3, the Raman intensity is influenced by many factors such as the intensity incident on the sample and the amount of sample even for the same spectrometer. Therefore, we introduce the self-reference algorithm. Two intensity peaks are used: one is named reference intensity peak, such as $1449{\mathrm{cm}}^{-1}$ and the other is named identification intensity peak, such as $1341{\mathrm{cm}}^{-1}$. By calculation the ratio of identification intensity peak to reference intensity peak, discrimination of human and nonhuman blood can be realized.

Fig. 3

The intensity of the spectra at $1341{\mathrm{cm}}^{-1}$.

3.2.

First Example of Blood Discrimination Using Self-Reference Algorithm and its Cross Validation

Figure 4 shows the results obtained by self-reference algorithm. The ratio of the intensity peak at $1341{\mathrm{cm}}^{-1}$ to that at $1132{\mathrm{cm}}^{-1}$ for human and nonhuman blood is listed. The mean value of human blood is 2.3118 and its RMS is 0.3457. The mean value of nonhuman blood is 1.1868 and its RMS is 0.1936. The threshold for discriminating human blood and nonhuman blood is set as the average of the minimal ratio of human blood and maximal ratio of nonhuman blood, 1.7 in this case as shown in Fig. 4. The value above this line is human blood and below this line is nonhuman blood.

Fig. 4

The ratio of the intensity peak at $1341{\mathrm{cm}}^{-1}$ to that at $1132{\mathrm{cm}}^{-1}$.

Leave-one-out cross-validation method^27,28 was used to validate the self-reference algorithm. The cross-validation results were summarized in Table 1. This table shows the number of positive identified samples, along with a total percentage of true positive classifications for each individual species.

Table 1

Internal cross-validation results for self-reference algorithm.

To avoid genetic disclosure and protect biodiversity, the performance of self-reference algorithm whose internal true positive rate (TPR) is 100% is significant for customs inspection. These results demonstrate that self-reference algorithm works well in discriminating human and nonhuman blood, which has real-world applicability in customs inspection, forensic science, and veterinary setting. To demonstrate the ratio used to discriminate human and nonhuman blood is not coincidental, different peak choices were performed: (1) the same reference peak and all the identification peaks; (2) another reference peak and all the identification peaks.

3.3.

Study of Influence of Using Other Identification Peaks by Self-Reference Algorithm

From Fig. 2, we can see several wave numbers can be used as the identification intensity peak. The range of identification wave number is from 1310 to $1342{\mathrm{cm}}^{-1}$. To evaluate the self-reference algorithm’s generality for all identification wave numbers, we calculated the ratio every five wave numbers from the wave number $1311{\mathrm{cm}}^{-1}$. The results are given in Table 2. The parameter named with 1 represents human blood and 2 represents nonhuman blood. The threshold value can be set as 1.7 to discriminate human and nonhuman blood for all the other wave numbers expect for two wave numbers of 1331 and $1341{\mathrm{cm}}^{-1}$, whose threshold is 1.63 and 1.68, respectively. The TPR, the false positive rate (FPR), the true negative rate (TNR), and the false negative rate (FNR) were then analyzed and listed in Table 3.

Table 2

The results obtained by the ratio of identification peak at different wave numbers to reference peak at 1132  cm−1.

Table 3

The identification rates when the reference peak is 1132  cm−1.

3.4.

Study of Influence of Using Other Reference Peaks by Self-Reference Algorithm

To evaluate the influence of choosing the reference wave number, $1003{\mathrm{cm}}^{-1}$ was chosen as the reference wave number to compare with the results using peak at $1132{\mathrm{cm}}^{-1}$ as reference. The same identification wave numbers were used. The results were shown in Table 4. Compared with Table 2, even though the threshold value is different, the discrimination of blood samples is still effective. The threshold value varies because of the different intensity peak at 1003 and $1132{\mathrm{cm}}^{-1}$. The accuracy of the discriminations based on the data in Table 4 is 100%.

Table 4

The results obtained by the ratio of reference peak at 1003  cm−1 to identification peak at different wave numbers.

Figure 5 shows the results obtained by self-reference algorithm and normalization. The identification peaks of both are $1341{\mathrm{cm}}^{-1}$. The reference peak in self-reference algorithm is $1003{\mathrm{cm}}^{-1}$. In normalization process, the intensity at $1341{\mathrm{cm}}^{-1}$ is divided by the maximal value of the Raman spectrum. The results shown in Fig. 5 indicate that the maximal value for each blood sample is not at the same wavenumber. The data overlapping of self-reference algorithm and normalization means the maximal value is just at the reference Raman peak $1003{\mathrm{cm}}^{-1}$. The change of denominator in normalization is equivalent to the change of reference peak. Therefore, the result obtained by normalization is lack of stability theoretically, even though the discrimination could be realized occasionally. The peak intensity at $1341{\mathrm{cm}}^{-1}$ in normalized Raman spectrum is not a suitable parameter that can be used to discriminate human and nonhuman blood.

Fig. 5

The discrimination of human and nonhuman blood with self-reference algorithm and normalization.

3.5.

Validation of the Self-Reference Algorithm by Blind Test

In total, 10 human blood samples and 16 animal blood samples (four dog blood samples, four rabbit blood samples, four rat blood samples, and four horse blood samples) were used to validate the self-reference algorithm. $1003{\mathrm{cm}}^{-1}$ was chosen as reference peak. $1341{\mathrm{cm}}^{-1}$ was chosen as the identification peak. The results were displayed in Table 5 and Fig. 6. Therefore, the self-reference algorithm is proved to be accurate and stable by blind test.

Table 5

The validation of the self-reference algorithm.

Fig. 6

The validation of self-reference algorithm with 10 human blood samples and 16 animal blood samples.

3.6.

Comparison Between Self-Reference Algorithm and Partial Least Square

The PLS^29,30 discrimination analysis was applied and the results were shown in Fig. 7. Figure 7(a) shows the 2-D score distributions of PC1 and PC2. Figure 7(b) shows the 3-D score distributions of PC1, PC2, and PC3. The results indicate that the discrimination of human and nonhuman blood could be realized with two principle components.

Fig. 7

The score plots of PLS model (a) 2-D score plot with two principal components and (b) 3-D score plot with three principal components.

The performance of PLS model was also evaluated with leave-one-out cross-validation method. The results were displayed in Table 6. Both self-reference algorithm and PLS can achieve satisfactory results. However, the model construction by self-reference algorithm is much easier. Moreover, the difference of chemical composition between human blood and animal blood can be well analyzed by self-reference algorithm. We find the C–H bend in tryptophan in blood is one of the main factors that induce the difference of the Raman spectra between human and nonhuman blood. For forensic science and veterinary purposes, we want to know not only the blood species, but also the relative concentration of some chemical components. Self-reference algorithm can provide the relative concentration information. In comparison, for PLS model, the information of the difference in chemical composition and the relative concentration cannot be achieved, although the results are enough for discriminating human and nonhuman blood.

Table 6

Internal cross-validation results for self-reference algorithm.