2.1.

Sample Preparation

Blood samples from 60 suspicious malaria and dengue infected individuals were collected by the pathology department at PAEC General Hospital Islamabad, Pakistan and 10 from healthy volunteers at our institute. The patients were referred from the out patient department of the hospital on different days. For diagnosis of the malarial infection, whole blood is required, therefore, blood samples were collected in ETDA-treated tubes (Hebei Xinle, Sci&Tech Co. Ltd., China) in order to prevent clotting. For confirmation of malaria, smeared slides were prepared from whole blood to examine the hemozoin in the RBC through a white light-illuminated microscope. Similarly, serum is required for the diagnosis of dengue virus infection through ELISA; therefore, for this test, blood samples were collected in tubes which contained a clot activator gel (Hebei Xinle, Sci&Tech Co. Ltd., China). Blood samples were taken from the large vein in cubital fossa. The selected site was first cleaned and disinfected by using 5% alcohol swab. After complete aseptic measures, 3 ml of venous blood was drawn with a disposable syringe. Safety precautions recommended by National Institute of Health (NIH)⁴¹ were followed from sample collection to obtain the Raman spectra.

In order to proceed for Raman spectroscopic studies, plasma and serum were extracted with a centrifuge machine (D-7200, Hitachi, Japan) at 3500 rpm for 10 min. Both plasma and serum were extracted carefully from the blood samples as per standard procedure. Each sample of a $30\mu \mathrm{l}$ drop was placed on a glass slide (Sail brand microscope slide 7101, China) and dried for 2 h at room temperature under a cover lid before obtaining the Raman spectra. Wet plasma samples introduce an intense fluorescence background which sometimes buries the Raman signal of solutes. Therefore, dried samples were chosen over wet, which increased the volume fraction of analytes and thus enhanced the signal strength of the Raman spectra. The drying process for samples produced in-homogeneities at the edges of the sample either through surface tension or molecular weight differences. However, the central part of the dried sample was of uniform⁴² thickness; therefore, it was chosen to acquire the Raman spectra. From the central part of the spot, three Raman spectra were acquired from slightly shifted focusing positions from the same sample and their average was further used in the analysis.

2.2.

Acquiring Raman Spectra

Raman spectra from the dried samples were obtained using a Raman spectrometer ($\mu$Ramboss: Micro Raman measurement system, Dongwoo Optron Co. Ltd, Korea), which offered a resolution of $4{\mathrm{cm}}^{-1}$. A diode laser at 532 nm was focused on the surface of the dry sample with a $100\times$ objective in the backscattering configuration. The laser power irradiation over the samples was 40 mW. Laser power was carefully optimized to avoid photodegradation of the samples,^43,44 which was done through neutral density filters to record the Raman spectra from a healthy sample using different laser powers and spectral overlaying as a function of Raman shifts was found. Each Raman spectrum was recorded from 760 to $1770{\mathrm{cm}}^{-1}$ with an exposure time of 8 s. Raman spectra of diseased samples were acquired at different times and days depending on their collection. Calibration of the Raman spectrometer is performed each time through recording the Raman spectrum of a silicon crystal wafer at $520{\mathrm{cm}}^{-1}$. A glass slide produces fluorescence background while recording the Raman spectra of samples placed on it, however, its Raman spectrum was subtracted from all that were used in the development of the model and for the blind test.

2.3.

Processing Raman Spectra

All the Raman spectra were analyzed using MATLAB (Release 2009; The Mathworks Inc., USA). First of all, the Raman spectra were denoised using a built-in “wden” function in MATLAB which is based on the wavelet decomposition and reconstruction method.⁴⁵ Using wavelets, we decompose the spectrum into its component parts, called a set of wavelet coefficients, which represent the complete features of the spectrum. The spectrum can be completely reconstructed using all these wavelet coefficients. The denoising process proceeds in three steps: decomposition of spectrum in which an orthogonal wavelet of level 10 is selected, soft thresholding is used based on Stein’s principle of unbiased risk, and reconstruction of the spectrum using wavelet coefficients. Stein’s unbiased risk estimate is an unbiased estimator that provides the accuracy and the soft threshold is a preprocessing tool that reduces the background in the spectrum, so pixels with intensity values below the threshold are reduced. In the second step, Raman spectra were further smoothed using the “mssgolay” function in MATLAB, which applied a moving average Savitzky–Golay smoothing filter method^46,47 over a span of nine points with a polynomial fitting of degree 4. In the third step, the fluorescence baseline in the Raman spectra was corrected using the “msbackadj” function in MATLAB.^48,49 For each spectrum, the “msbackadj” function estimated a baseline within multiple shifted windows ($200{\mathrm{cm}}^{-1}$ width) and regressed the baseline using spline approximation before subtracting it from the spectrum. In this way, all Raman features narrower than $200{\mathrm{cm}}^{-1}$ were preserved, while those that were wider are considered a product of background fluorescence and are removed.

3.1.

Spectral Recognition of Disease Signatures

Malaria and dengue virus infectious diseases appear almost at the same time of year and come out with fever symptoms. Clinically, both fever types cannot be differentiated in early days, which might result in a false diagnosis; therefore, a reliable, efficient, fast and early laboratory diagnostic test is very important for disease management. This article reports a comparison of Raman spectroscopic signatures of malaria and dengue virus infected samples¹⁹ with healthy ones. During the centrifugation of blood in the ETDA-treated tubes, fibrinogen rests inside plasma, and in the gel activated tubes, it is completely absent in the serum. In order to compare the similarities and differences in both samples, a Raman spectrum from each the healthy volunteers is recorded and is presented in Fig. 1. Both spectra show a similar trend except for a few Raman lines that appear due to fibrinogen, but these are of very low intensity. The spectral range from 760 to $1770{\mathrm{cm}}^{-1}$ that was used in the development of the multivariate model contained only one Raman line at $829{\mathrm{cm}}^{-1}$ in the spectrum of the plasma. This line represents fibrinogen^50,51 and is of very low intensity.

Fig. 1

Normalized Raman spectrum of a serum and plasma of a healthy volunteer illustrates their mutual differences and similarities. These Raman spectra are recorded from 760 to $1750{\mathrm{cm}}^{-1}$ using a laser at 532 nm as the excitation source.

Out of 60 blood samples collected from the hospital, 20 were clinically confirmed for malaria infection and 10 for non-malaria but the patients were suffering from fever. Raman spectra of all the samples were acquired, processed, and plotted on the vertical axis in Fig. 2. The Raman spectra of the dengue-infected samples from six individuals from another location are also included in Fig. 2.¹⁹ This depicted a nice comparison between the biochemical changes associated with malaria infection (red/cyan color) from 20 samples, non-malaria but suffering from fever (pink color) and healthy (green color) ones each from 10 cases, and dengue infection (black color) from 06 samples. The overlapping of individual categories Raman spectra in all the cases shows the precision in the measurement. These Raman spectral peaks are listed in Table 1 to present a quantitative comparison between the spectral signatures of both diseases with healthy ones.

Fig. 2

Normalized Raman spectra of 20 malaria (plasmodium vivax) infected (red/cyan color), 10 non-malarial having fever (pink color), 10 healthy (green color), and 6 dengue-infected¹⁹ samples (black color) using a laser at 532 nm as an excitation source, it extends from 760 to $1750{\mathrm{cm}}^{-1}$ spectral shift, which illustrated specific prominent peaks in each case.

Table 1

This shows a quantitative comparison of Raman spectral peaks of malaria, dengue virus infected, non-malaria, and healthy plasma/sera samples (Raman lines are taken from the spectra shown in Fig. 2).

There are three prominent Raman lines at 1005, 1156, and $1514{\mathrm{cm}}^{-1}$ that appeared in the healthy samples and which represent $\beta$-carotenoids;⁵³ however, the line at $1005{\mathrm{cm}}^{-1}$ is also indicative for phenylalanine⁵³ and IgG.⁵² This line is sharp in the case of healthy/non-malaria samples as compared with the malarial infected ones. It can be inferred that for healthy samples, its occurrence is mainly due to $\beta$-carotenoids and Phenylalanine, however, its broadening in malaria samples is due to the rise of hemoglobin because a Raman line at $1001{\mathrm{cm}}^{-1}$ representing hemoglobin^20,21 is expected to be buried inside $1005{\mathrm{cm}}^{-1}$. The resolution of the Raman spectrometer used in the present studies cannot resolve the Raman line at $1001{\mathrm{cm}}^{-1}$ from $1005{\mathrm{cm}}^{-1}$. Therefore, the presence of the Raman line at $1001{\mathrm{cm}}^{-1}$ cannot be confirmed; however, if present, it will contribute to the broadening of the Raman line at $1005{\mathrm{cm}}^{-1}$ in malaria samples. The appearance of $1005{\mathrm{cm}}^{-1}$ in the dengue infected Raman spectra (top in Fig. 2) might be due to IgG,⁵² and is comparatively of low intensity which means that this line in malaria infection is mainly due to hemoglobin. Similarly, the Raman line at $1156{\mathrm{cm}}^{-1}$, appeared in all categories as shown in Fig. 2; however, its signal intensity is higher in healthy/non-malaria samples, and it reduces in strength in malaria and to a minimum level in dengue virus infected cases. The Raman line at $1514{\mathrm{cm}}^{-1}$ is present in healthy/non-malaria at the highest level, whereas its intensity level decreases many times in the case of malaria and completely disappears in dengue-infected samples. $\beta$-carotenoids are important antioxidants and play an important role in protecting the organism and immune system^19,54 from free radicals and disappear or are reduced in the case of dengue virus infection.¹⁹ It is suggested that the monitoring of $\beta$-carotenoids may help in better disease management.

In Fig. 2, Raman lines at 835, 837 and $845{\mathrm{cm}}^{-1}$, representing IgM⁵² appeared in the Raman spectra of non-malaria (pink color), malaria, and dengue infected samples, respectively. Both antibodies, IgM and IgG, are long chain molecular structures and different functional groups are attached to them. The above cited Raman peaks that represent IgM at three different shifted wavenumbers appear due to the vibrations of these functional groups in response to different infections. These shifts in the position of the Raman peaks play a significant role in the development of the PLS prediction model.

The Raman line at $968{\mathrm{cm}}^{-1}$, representing lipids,⁵³ is present only in malaria and dengue infected samples, but it is more pronounced in the malarial cases. Similarly, the Raman line at $1066{\mathrm{cm}}^{-1}$ is labeled for IgM⁵² and Porline⁵³ (collagen assignment), and appeared only in the dengue infected samples. The Raman line at $1082{\mathrm{cm}}^{-1}$, representing hemoglobin,^20,21 is present in malaria and non-malaria samples, but is more pronounced in malaria cases. The Raman line at $1123{\mathrm{cm}}^{-1}$, representing IgG,⁴⁷ appeared only in the dengue infected samples, whereas the line at $1131{\mathrm{cm}}^{-1}$, representing hemoglobin^20,21 and Palmitic/fatty acid,⁵³ appeared in non-malaria and malaria cases; however, its level increases many times in malaria samples. In the non-malaria samples, this line may appear due to Palmitic/fatty acid. Raman lines at 1172 and $1226{\mathrm{cm}}^{-1}$, representing hemoglobin,^20–23 appeared only in malaria infected samples. Hemoglobin comes out in the plasma due to the rupturing of the RBC wall during the plasmodium multiplication process. The Raman line at $1303{\mathrm{cm}}^{-1}$, representing IgM,⁵² is present both in malaria and dengue infected samples; however, its signal intensity is higher in malarial spectra. The Raman line at $1339{\mathrm{cm}}^{-1}$, representing tryptophan,⁵³ IgG⁵² and hemoglobin,^20–23 and lines at 1357 and $1393{\mathrm{cm}}^{-1}$, assigned for hemoglobin,^19–22 are present only in the malaria infected samples. The Raman line at $1452{\mathrm{cm}}^{-1}$, assigned for IgG,⁵² is present in dengue, malaria and non-malaria samples and is absent in healthy ones. However, its intensity is higher in the case of the dengue virus infection as compared with malaria and non-malaria samples.

There are two malaria-infected samples (cyan color), which show a Raman line at $1543{\mathrm{cm}}^{-1}$, different from the other malarial infected ones. It is assigned for ${\nu }_{11}$ (${\mathrm{C}}_{\beta }{\mathrm{C}}_{\beta }$), observed in the spectra of human RBC,^18–55 which were from the samples whose color was reddish rather than straw-yellow, the color of pure plasma; it means that it is mixed with RBC. However, this Raman line does not affect the prediction of malaria infection, which is due to distinct Raman signatures of malaria infection. It emphasizes the potential of Raman spectroscopy for predictive diagnosis of infectious diseases even if patients have more than one disease. The Raman line at $1552{\mathrm{cm}}^{-1}$, representing hemoglobin^31,32 and tryptophan,⁵³ appeared only in malaria infected samples, whereas at $1583{\mathrm{cm}}^{-1}$, assigned for phenylalanine,⁵³ it showed its weak signatures in the non-malaria cases. Raman lines at 1586, 1603, and $1673{\mathrm{cm}}^{-1}$, assigned for hemoglobin,^31,32 phenylalanine,⁵³ and amide-I group,⁵³ respectively, appeared only in malaria infected samples, however, at $1732{\mathrm{cm}}^{-1}$, representing cortisone, is responsible for stress in the body¹⁹ and is present only in the dengue virus infected cases.

Figure 2 showed Raman spectra of three amino acids: tryptophan, tyrosine, and phenylalanine,⁵⁶ their concentrations in healthy samples are low, it increases many times in malaria-infected samples, and they are completely absent in dengue-infected cases. These play an important role for the quality of health and the rise of their concentrations, particularly tryptophan,^57–59 causes brain fatigue. Tryptophan is the precursor for neurotransmitter 5-hydroxytryptamine (5-HT), which is involved in fatigue and sleep.⁵⁸ An increased level of tryptophan in plasma leads to a higher level of 5-HT, which is reported for central brain fatigue.⁵⁹ In the case of malarial infection, patients usually feel a headache, stress, and exhaustion in their body, which is due to the increase of amino acids in the plasma.

3.2.

Disease Prediction Model Based on Raman Spectra

Raman spectra of body fluids are complex due to multiple background sources, which causes fluorescence and spectral overlapping; therefore, straightforward spectral analysis cannot be effectively used for accurate diagnosis. Multivariate analysis can facilitate this in such a way that it eliminates background interferences to make possible the prediction of viral infectious diseases like dengue.^19,60 In this study, a multivariate model has been developed with a PLS regression engine NIPALS algorithm, a built-in routine written in the MATLAB environment (Release 2009; The Mathworks Inc., USA), with the Raman spectra of malaria, dengue virus infected, and healthy samples as the trainee data set. PLS regression offers the advantage of predicting the unknown by developing a proper model based on the known database.⁶⁰ The benefit of PLS-based multivariate regression is that it takes the whole Raman spectra rather than prominent peaks and includes minute spectral variations for the prediction of diseases. This methodology supports the development of a multivariate database; it matches the spectral pattern of the unknown samples with the stored ones for disease prediction in the presence of multianalyte background sources.^61,62 PLS uses loading vectors for principal component analysis that reflect the intrinsic features of the pure component spectra and we used four loading vectors for accurate prediction. PLS, a strong regression engine, enables the prediction of the analyte of interest without any information about interfering constituents. It achieves a prediction accuracy comparable to that of more explicit models requiring more information.⁶⁰ In order to develop the multivariate model, Raman spectra of 20 malaria infected (red/cyan color), 10 non-malaria but suffering from fever (pink color), 10 healthy (green color) and 06 dengue infected samples (black color) were used as trainees to teach the model. In Sec. 3.1, it can be seen that there are 10 distinctive Raman peaks at 837, 1172, 1226, 1339, 1357, 1393, 1552, 1586, 1603 and $1673{\mathrm{cm}}^{-1}$, which are found only in the malaria infected plasma samples. These strong malaria Raman signatures played a significant part in the development of the statistical PLS model for the differentiation of malaria from dengue and healthy ones. In addition, blood samples from 30 suspicious cases were kept hidden for a blind test of the multivariate model by the pathology department of the hospital.

During the development of the multivariate model, a leave-one-sample-out (LOO) cross validation method⁶³ was employed for authenticating the model. The LOO cross validation process was repeated for all the samples used for training the model and then the left out sample was predicted. The process was repeated as for many times as s there were samples. The predicted values of the trainee spectra through the LOO cross validation process are shown in Fig. 3 with hollow red circles as a function of clinical results. It also showed a calibration curve that was produced by the multivariate model which fitted well with the LOO cross validated predicted values. The correlation coefficient $r^2$ between the LOO cross validated predicted and clinical results was found to be 0.981, which showed that the validation process is accurate.^63,64 In addition, the root mean square error in cross validation (RMSECV) for the entire LOO cross validated predicted values was calculated to be 0.093, which showed that the multivariate model is authenticated.⁶⁵

Fig. 3

This shows the calibration curve that fit best on the predicted values of the trainee samples through the LOO cross validated process (red hollow circles) verses known clinical results. It also illustrated the predicted values of 22 unknown suspected samples (blue solid circles) verses clinical results that were kept hidden during a blind test of the model.

The multivariate model generated regression curve based on the Raman spectra of the trainee sample, called the beta vector, is shown in Fig. 4. It showed the spectral line shape of the beta vector (blue color) and the Raman spectra of trainee samples are used in the development of the model. When an unknown suspected Raman spectrum is loaded in the multivariate model, it is multiplied with the beta curve and produces a correlation factor $r^2$, from which the disease is predicted. The PLS regression vector (blue color) takes positive values for unique Raman peaks that appeared only in the malaria infected samples (it overlaps the malaria infected Raman signatures) and changes its sign where non-malarial/healthy or dengue infected Raman peaks or both existed. The beta vector showed strong correlation/overlapping with the Raman peaks at 1131, 1172, 1226, 1339 and $1357{\mathrm{cm}}^{-1}$, which appear only in malaria infected samples.

Fig. 4

This shows the PLS generated regression vector (blue color), called a beta vector, along with the Raman spectra of trainee samples used in the multivariate model. The PLS regression vector (blue color) overlaps the unique Raman peaks that appeared only in malaria-infected samples and changes its sign where non-malaria/healthy Raman peaks or dengue virus infected Raman peaks or both are existed.

For further evaluation of the model, a blind test was performed using Raman spectra of 30 unknown suspected samples, which include malaria, dengue infected, non-malarial having fever, and healthy ones. These spectra were loaded by the model one by one to predict malaria infection. These predicted values are shown in Fig. 3 with solid circles (blue color) versus clinical results for a nice comparison. For these predicted values, root mean square error in prediction (RMSEP) was calculated to be 0.149, which showed accurate prediction of the multivariate model.^63–65 Another important parameter, the standard deviation (SD) in error for 30 unknown predicted samples, was calculated to be 0.150, which is awarded the zero point score as per the scoring system of quality control for molecular diagnostics,⁶⁶ therefore, these results are clinically accepted.

It is evident from Fig. 3 that the prediction of malaria infection in unknown plasma/sera samples depends upon PLS prediction values. We fixed the lower limit as 0.2 for non-malaria and the upper limit as 0.8 for malaria infection. These limits were defined based on the predicted values of the trainee samples through the LOO cross validation procedure for the healthy and diseased cases. The region between both the limits is defined as a grey region. If the model predicted value of a sample lies in the grey region, it will be asked for a fresh test.

The PLS predicted values between the cut off limits may have some relation with the development of disease and for that, PCR studies of disease evolution from initial to advance levels are imperative to develop a strong correlation between the plasmodium load and Raman signatures, which may be the topic of future research. The prediction accuracy of the multivariate model for unknown samples was 86% compared with the clinical results. The inaccuracy was due to the prediction of three samples in the grey region, i.e., their predicted values were found inside the limits from 0.2 to 0.8 that is defined for disease prediction. This may be due to the quality of samples and the method of obtaining the Raman spectra. Accuracy in the prediction can be increased by developing the model with a large number of samples, careful screening of blood samples through microscope for malaria confirmation, and detailed history of patients regarding their other diseases and medicines in use. Nevertheless, a lot of work is needed to put the technique into clinical practice. Comparison of the Raman spectrum of serum and plasma in Fig. 1 showed that both spectra differ for the peak at $829{\mathrm{cm}}^{-1}$ (of much lower intensity) for the spectral range of 760 to $1770{\mathrm{cm}}^{-1}$ used in the development of the model, but the strong Raman signatures of malaria infection did not affect the predicted accuracy when using serum or plasma samples. Practically, we did not observe any dependence of plasma/sera samples on the prediction of malaria infection in the loaded spectrum.

In the present studies, plasmodium vivax infected samples were used; therefore, the model will predict only this infection in patients. Other types of infections may induce some different biochemical changes; therefore, further studies are needed for the development of relevant model.