Differential Diagnosis of the Etiologies of Bacterial and Viral Infections

Subscriber access provided by Kaohsiung Medical University

Article

Differential Diagnosis of the Etiologies of Bacterial and Viral Infections Using Infrared Microscopy of Peripheral Human Blood Samples and Multivariate Analysis Adam H. Agbaria, Guy Beck, Itshak Lapidot, Daniel H Rich, Mahmoud Huleihel, Shaul Mordechai, Ahmad Salman, and Joseph Kapelushnik Anal. Chem., Just Accepted Manuscript • DOI: 10.1021/acs.analchem.8b00017 • Publication Date (Web): 05 Jun 2018 Downloaded from http://pubs.acs.org on June 5, 2018

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Differential Diagnosis of the Etiologies of Bacterial and Viral Infections Using Infrared Microscopy of Peripheral Human Blood Samples and Multivariate Analysis Adam H. Agbariaa, Guy Beckb, Itshak Lapidotc, Daniel H. Richa, Mahmoud Huleihel#d, Shaul Mordechai#*a, Ahmad Salman#*e, and Joseph Kapelushnik#b

a

Department of Physics, Ben-Gurion University, Beer-Sheva 84105, Israel.

b

Department of Hematology, Soroka University Medical Center, Beer-Sheva, 84105, Israel.

c

Department of Electrical and Electronics Engineering, ACLP-Afeka Center for Language Processing, Afeka Tel-Aviv Academic College of Engineering, Tel-Aviv 69107, Israel

d Department of Microbiology, Immunology and Genetics, Faculty of Health Sciences, Ben-Gurion University of the Negev, Beer-Sheva 84105, Israel. e

Department of Physics, SCE-Sami Shamoon College of Engineering, Beer-Sheva 84100, Israel.

*Corresponding authors: Dr. Ahmad Salman

Prof. Shaul Mordechai

Tel: +972-8-6475794

Tel: +972-8-6461749

Fax: +972-8-851916

Fax: +972-8-6472903

e-mail: [email protected]

e-mail: [email protected]

# contributed equally

ACS Paragon Plus Environment

Analytical Chemistry 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Abstract Human viral and bacterial infections are responsible for a variety of diseases that are still the main causes of death and economic burden for society over the globe. Despite the different responses of the immune system to these infections, some of them have similar symptoms, such as fever, sneezing, inflammation, vomiting, diarrhea, and fatigue. Thus, physicians usually encounter difficulties in distinguishing between viral and bacterial infections based on these symptoms. Rapid identification of the etiology of infection is highly important for effective treatment and can save lives in some cases. The current methods used for the identification of the nature of the infection are mainly based on growing the infective agent in culture; a time consuming (over 24h) and usually expensive process. The main objective of this study was to evaluate the potential of the mid-infrared spectroscopic method for rapid and reliable identification of bacterial and viral infections based on simple peripheral blood samples. For this purpose, white blood cells (WBCs) and plasma were isolated from the peripheral blood samples of patients with confirmed viral or bacterial infections. The obtained spectra were analyzed by multivariate analysis: principle component analysis (PCA) followed by linear discriminant analysis (LDA), to identify the infectious agent type as bacterial or viral in a time span of about one hour after the collection of the blood sample. Our preliminary results showed that it is possible to determine the infectious agent with high success rates of 82% for sensitivity and 80% for specificity, based on the WBC data.

Keywords: Viral infection, bacterial infection, linear discriminant analysis, plasma, white blood cells, infrared spectroscopy.


Page 2 of 27

Page 3 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Introduction Bacterial infections differ from viral infections in several important aspects, including the response of the immune system and the response to various medications. While antibiotics are the main medication for bacterial infection, they have no effect on viruses. On the other hand, vaccines have been used to reduce or prevent many viral infections. Recently, medications have been developed against some viral infections like herpes, HIV/AIDS and influenza1-3. Unfortunately, the development of antibiotics has been associated with the evolution of drugresistant bacteria. A negative outcome of uncontrolled use of antibiotics is the development of new mutant strains of bacteria that are resistant to a wide range of antibiotics 4. Thus, it is advised now to use antibiotics for treatment only in cases of clear evidence of a bacterial infection. The human immune system defends our bodies by reacting against invading pathogens at two levels of immunity: innate (non-specific) and adaptive (specific) immunities. Through the innate immunity, the pathogens are attacked by phagocytes, which are macrophages and neutrophils cells that are considered as WBCs. In contrast, the adaptive immunity allows selective targeting of a specific pathogen by lymphocytes, which normally compose 20-35% of the WBCs, and are divided into two types: T- cells and B-cells. When T-cells recognize the antigens that appear in the macrophages membranes, they send messages via proteins to the B-cells. Both T and B lymphocytes start replicating rapidly and attack the specific invader pathogen. While T-cells attack the pathogens and cells infected with the pathogen and cause their degradation, B-cells start to produce millions of antibodies that are specific to the detected antigens that appeared on the macrophages membranes. These antibodies are small proteins, and are engineered to bind specifically to the invading pathogens, thereby neutralizing them 5-7. Moreover, the host cells release a group of signaling proteins in the blood called interferons (IFNs), which are divided into two common types: IFN-I is released in viral infection cases and activates



Page 4 of 27

cells to prevent virus replication. IFN-II is released in case of infection, regardless of whether the infection is of viral or bacterial origin. Thus, a high concentration of IFN-Is in the blood may indicate a viral infection 8,9. By their experience, physicians relate a high concentration of neutrophils or C-reactive proteins in the blood, which is diagnosed by simple blood test, to bacterial infections

9-11

. In numerous cases

physicians diagnose the etiology of infection subjectively, based on their experience and on the symptoms, and therefore they may occasionally start unnecessary antibiotic treatments due to wrong diagnoses

12

. Therefore, for objective diagnoses, physicians must wait a long time (several

days) for the microbiological lab results, in order to determine objectively the etiology of the infection. There are several classical methods that are used in the microbiological lab for diagnosing the infection type, including urinalysis, urine culture

13

, blood culture

14-17

, and viral test (i.e.,

complement fixation or other similar tests ). All these methods are time consuming; some are expensive. FTIR spectroscopy is a simple and rapid biochemical and physical tool that has the ability to reveal information about the molecular composition of biological samples

18

. Previous studies reported

that FTIR spectroscopy of peripheral blood mononuclear cells and plasma is a potentially feasible and efficient tool for the early detection of breast neoplasms

19,20

, and in another study of

distinction, between benign lesions and malignant tumors21. Furthermore, the correlation of specific spectral changes with clinical parameters of cancer patients indicates an enhanced potential for possible contributions to diagnosis and prognosis 22-28. In our recent study, we used infrared microscopy to determine the susceptibility of E. coli bacteria to different antibiotics 29. Infrared spectroscopy was able to monitor small molecular changes in the bacterial genome that lead to acquiring resistivity. Thus, we hypothesize that the biochemical changes in the blood contents, which result from different responses of the human immune system due to viral and bacterial infections, can be monitored by infrared spectroscopy. Based on these


Page 5 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


spectral changes, multivariate analysis and machine learnings algorithms will be used to classify the samples into controls, bacterial and viral 30-33.

Methodology Patients and data collection The study was conducted in collaboration with Soroka University Medical Center and in accordance with the Institutional Helsinki Committee Approval. We collected peripheral blood samples from young patients with ages lower than18 years old, limited to those who were diagnosed with fever-producing illness. Table S-1 includes part of the patients included in this study, and their available information such as age, fever, gender and infection site whenever available at the time of arrival to the emergency room. The main steps of all the stages carried out in this study are described in Figure S-1. We studied 364 patients, of which 93 were controls, 126 had bacterial infections, and 145 had viral infections. The database of the infection cases was divided into two sub-categories according to the diagnosis methods used: physician-based diagnosis data (PDD), which included all the cases that were not diagnosed by microbiology tests (70 bacterial and 116 viral), and microbiology laboratory-based diagnosis (MLD), which included all the cases that were diagnosed using microbiology tests (56 bacterial and 29 viral). The combined MLD and PDD cases were abbreviated as CMPD. One ml of peripheral blood was collected in 3 ml ethylenediaminetetraacetic acid blood collection tubes (BD Vacutainer® Tubes, BD Vacutainer, Toronto) from each participant, using standard phlebotomy procedures from a peripheral vein 34. Sample preparation Samples were processed within 4 h after collection. The blood was transferred carefully to a new 3 ml tube that contains 1:1 in Histopague-1077 (Sigma Chemical Co., St. Louis, Missouri, USA) and then centrifuged at 133.6 g (1200 rpm) for 30 min, where the blood was separated into four coats, plasma, WBCs, histopaque, and red blood cells. Afterwards, the plasma was transferred to



Page 6 of 27

1.5 ml tubes and centrifuged at 835 g (3000 rpm) for 10 min. Then, the supernatant was transferred to a new 1.5 ml tube and diluted in 1:1 distilled water. Then 1 µl drop of plasma was deposited on a zinc selenide slide and air-dried for 15 min. The WBCs were transferred to 1.5 ml tube and washed three times by dilution in 1:1 isotonic saline (0.9% NaCl solution) and centrifuged at 59.4 g (800 rpm) for 5 min one drop 1 µl of WBCs was deposited on a zinc selenideslide after dilution and airdried for 15 min. FTIR microscopy The dried plasma and WBC samples were then subjected to micro-FTIR. We used an FTIR microscope Thermo Scientific™ Nicolet™ Continuµm™ Infrared Microscope with a liquid nitrogen-cooled mercury-cadmium-telluride detector, coupled to the “Nicolet iN10” FTIR spectrometer. The dried plasma/WBC samples were measured in transmission mode under the microscope with 128 scans for each measurement in the wavenumber region of 700 to 4000 cm-1 and spectral resolution of 4 cm-1. Spectral manipulation Each spectrum was initially manipulated using OPUSTM 7.0 software (BRUKER Germany). The spectra were smoothed (via the Savitzky–Golay technique) using 13 points

35

; cut into two

regions—800-1800 cm-1 (low region) and 2830-3010 cm-1 (high region). The low wavenumber region underwent baseline correction, using rubberband correction with 64 baseline points and five iterations. Vector normalization and offset correction were used after baseline correction for all the spectra before subsequent analysis 30.

Calculating the second derivative spectra The second derivative spectra were calculated from the measured spectra after manipulation but excluding smoothing. These calculations were made with 13 smoothing points and used for further analysis.


Page 7 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Multivariate and statistical analysis Multivariate analysis provides both descriptive and inferential procedures to distinct between human bacterial or viral infections. Here we use it for classification of patterns

36

(bacterial verses

viral). In our study, principle components analysis (PCA) applied for dimensionality reduction, followed by Fisher linear discriminant analysis (FLDA) as a linear classifier in order to differentiate between bacterial and viral infections based on the second derivative spectra. Although PCA reduces the data dimension by projecting the data on the sub-space by preserving most of the variance, in practice in many applications it is also very useful for discrimination. FLDA is a linear classifier that assumes that each class can be described by a single Gaussian and all the classes share a common covariance matrix and differ by their expectation vectors. When the amount of the training data is small and the data is close to be linearly separable, this is a very simple and efficient approach.

Results Infrared Spectra The average infrared absorption spectra of the different categories: bacterial, viral and control for both plasma (a) and WBCs (b) are presented in Figure 1, in the wavenumber region 800-1800 cm-1 after manipulations. The centroids of the major absorption bands are displayed in the figure. The absorption bands centered at 1741 cm−1 are contributed mainly due to phospholipids. Amide I (C O) and amide II (C–N stretch and H–N–C bend) bands are contributed mainly by protein vibrations. Lipids are the main contributors to the band centered at 1456 cm−1, due to antisymmetric vibrations of CH3. The band centered at 1395 cm−1 is related mainly due to the COO− symmetric stretching of amino acids and the symmetric bending mode of the methyl group (CH3) in proteins. Proteins contribute to the amide III band at 1252 cm−1 37-40.The high wavenumber region was not included in



the analysis because the best classification results were acquired based on the 800-1800 cm-1 low wavenumber region, thus, we concentrated on the low region throughout the analysis. The spectra plotted in Figure 1 are the average spectra of the three investigated categories. Looking carefully at the individual spectra of these three investigated categories (Figure S-2) it can be seen that they are similar and overlap with minute spectral differences in shapes and relative intensities of the absorption bands. Thus, multivariate analysis PCA and LDA were used for the classification procedure. The spectral differences between controls, bacterial and viral infections are minute, thus, we need to measure high signal-to-noise ratios (SNRs) and high reproducibility spectra. As can be seen from Figure 2, which shows 16 spectra acquired from different sites of the same sample, the spectra are almost overlaid indicating the high reproducibility of the data. The average of theses 16 spectra was used to represent the sample (patient) spectrum. This procedure was used for each of the 364 samples included in this study. Our analysis was performed on the second derivative spectra, which were found to give the highest classification rates.

Evaluation procedure PCA calculations were applied on the second derivative spectra of WBCs and plasma components. This method was used for dimensionality reduction. In the present study, the dimensionality reduction is from 521 to a maximum of 21 dimensions. In all the different strategies and experiments, the LDA followed the PCA in order to differentiate between the various groups using second derivative spectra in the 800-1800 cm-1 region. We also tried the combined region (800-1800; 2800-3010 cm-1) but the classification results were not satisfying so all the analyses were performed based on the low region 800-1800 cm-1.


Page 8 of 27

Page 9 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


We used leave-one-out (LOO) cross validation to evaluate the performance of the classifier for different numbers of principal components and to choose the optimal number of principal components for good classification. According to this method, the classifier was trained on data from N-1 samples (the training set) and validated on one sample in each run (the validation sample) 41

.

In all experiments that included only MLD data, LOO was used for validation. When the experiments included the CMPD (combined MLD and PDD) the classifier was trained using the MLD only and the validation was performed using PDD. We tried to separately differentiate between samples, first with bacterial and controls, and then with viral and controls, using the WBCs and plasma (data not shown) of MLD and CMPD. The classification success rates in percentages versus PCs number of MLD and CMPD are plotted for the bacterial-controls and viral-controls categories, using WBC (Figure 3) and using plasma (data not shown). The performance of the LOO evaluation within the first 16 PCs for the MLD and CMPD were used to obtain the classification results summarized in Table 1, below. In the table, sensitivity (SE), is the probability of correctly detecting patients who do have bacterial or viral infections, specificity (SP) is the probability to correctly detect patients who do not have either viral or bacterial infections, positive predictive values (PPV), is the probability of diagnosis of bacterial or viral infections; and negative predictive values (NPV), is the probability of correct control diagnosis. The results were deduced based on LDA.

Table 1: Performance of the classification method in terms of SE, SP, PPV, and NPV, calculated by LDA using WBC and plasma spectra in the 800-1800 cm-1 region for MLD and CMPD



Blood

SE (%)

SP (%)

PPV (%)

NPV (%)

MLD

CMPD

MLD

CMPD

MLD

CMPD

Control-

CMPD

component

MLD

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 27

WBC

97

99

93

90

95

96

96

96

Plasma

78

93

82

91

84

98

75

58

WBC

87

94

94

81

96

95

79

89

Plasma

83

90

75

89

83

97

75

67

Bacterial

ControlViral

As was seen in Figure 3, above, and based on the MLD using both WBCs and plasma (not shown), it is possible to differentiate between controls and each of the infection categories with high success rates. However, WBC samples enable the classifier to achieve overall better success rates. Moreover, the classifier performance based on CMPD and MLD were comparable, with only a slight preference for CMPD. Because the main point is to differentiate between cases of viral and bacterial infections, we updated the classifier in order to differentiate among all three classes simultaneously: control, viral, and bacterial samples. Figure 4a shows the classification success rate versus the PCs number based on the 800-1800 cm-1 region of WBC spectra using MLD and CMPD. Similar analysis was done using plasma spectra (Figure S-3). Figures 4a and S-3 indicate again that WBCs enable superior classification success rate compared with plasma. Moreover, MLD enables the classifier to achieve higher performance for both WBC and plasma spectra. Using the first 20 PCs for WBC spectra, 80% sensitivity, 64% specificity, 87% PPVs and 60% NPVs were achieved. In order to improve the performance of the LDA classifier, different strategies of two stages was used. In the first stage, the data were differentiated only into two categories: control and infection


Page 11 of 27

where the infection category includes both the viral and the bacterial cases. In the second stage, the infection dataset was separated into viral and bacterial categories. Figures 4b and 4c show the classification accuracy in percentages versus PC number for the first and second stages of the second strategy using WBCs with MLD and CMPD. Similar plots were generated for plasma (data not shown). The scores were calculated using PCA of the WBC spectra of the MLD data in the 800-1800 cm-1 region and plotted in Figure S-4 for a) control-infection groups and b) bacterial-viral groups. The variance of each PC was stated in parentheses along the axis. The performance of LDA within the first 16 PCs for the infection-controls classification and first 21 PCs, for the bacterial-viral classification using LOO approach with MLD and CMPD, were used to obtain the classification results summarized in Table 2.

Table 2: Performance of the method in terms of SE, SP, PPV, and NPV, calculated by LDA using WBC and plasma measurements for MLD and PDD in the 800-1800 cm-1 region.

Blood

SE (%)

SP (%)

PPV (%)

NPV (%)

MLD

CMPD

MLD

CMPD

MLD

CMPD

Control-

CMPD

component

MLD

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


WBC

98

95

90

71

95

96

96

63

Plasma

85

93

96

91

98

99

69

58

WBC

82

53

80

61

92

51

58

63

Infection

Bacterial-



Page 12 of 27

Viral Plasma

60

55

35

62

71

51

25

66

In Table 2 the infection-controls classification, SE is the probability to correctly detect patients who have either a bacterial or viral infection; SP is the probability to correctly detect patients who do not have either form of infection; PPV is the probability of correct diagnosis of infection; and NPV is the probability of correct diagnosis of controls. For bacterial-viral infections, SE is the probability of correct diagnosis of patients who do have bacterial infection; SP is the probability to correctly diagnose patients who do have viral infection (not bacterial); PPV is the probability of correct identification of bacterial infection; and NPV is the probability of correct viral infection diagnosis. The results were deduced based on LDA. As can be seen from Figure 4c, the best classification success rate of 82% was obtained for WBCs using MLD. In order to evaluate the confidence level of our classifier in classification of the bacterial and viral categories, a permutation test was performed. We applied 10,000 permutations on the LDA, based on the MLD data, using 8 PCs, and obtained 0.028 p-value using LOO. The variance of loading 1 is 90.7%. Employing this loading alone it was possible to achieve 74% success rate in differentiation between bacterial and viral samples by analyzing the MLD data of WBCs. Thus, analysis of loading 1 (PC1) may identify and annotate the main vibrational bands that contribute to the classification

42,43

. Figure 5 shows (a) one of the WBC spectra, (b) its second

derivative spectrum, and (c) the first loading of the analysis of the MLD data of WBCs.

Discussion Early diagnosis of the etiology of an infection as bacterial or viral is highly important in medicine. In this study, we used the WBC and plasma blood components in order to assist in the making of such diagnoses. The potential of FTIR microscopy combined with multivariate analysis was


Page 13 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


evaluated for objective determination of the source of human infection as bacterial or viral, and to accomplish that task in a time span of minutes, based on a simple blood test. As the measurements use only peripheral blood components, in this research the safety hazards or risks are very minimal. A recent study has reported a differential diagnosis of acute bacterial and viral infections 44 via biochemical techniques, which yielded a sensitivity of 90%, but the method is still in the research stage and is quite expensive. The development of new bioinformatic methods for spectral analysis and advanced instrumentation has enabled the examination of large numbers of samples in order to use blood components to detect different diseases

45

. This is considered as minimally invasive and could thus be used for

routine clinical tests 45,46. As the issue of determining the etiology of infection is extremely important in clinics, it is quite desirable to search for additional methods. Microbiology lab methods, which are routinely used for determining the etiology of infections in medical centers, are time consuming and expensive. Our method is based on routine blood examinations that could be performed almost in every clinic, using a simple blood test (2-4cc). The method is very simple and there is no need for highly trained technicians or complex and expensive instrumentation; thus, it may be suitable for routine clinic tests and screening of a large number of patients. Infrared microscopy is known for its simplicity and sensitivity to molecular and chemical changes in blood components. These attributes are due to secretion of amyloid beta peptides into the blood stream, as well as changes in the WBC and plasma structure (e.g., proteins and carbohydrates structures…) due to the infection type

20,47

. Thus, infrared microscopy may suggest a new method

for diagnosing the etiology of human infection. The typical scenario for infection type diagnosis begins from the first impression of the physicians in the emergency room (ER). However, because bacterial and viral infections have similar symptoms in many cases, the type of infection diagnosed by physicians is mostly uncertain, with only a 55-70% likelihood of success


48,49

. Thus, the gold


standard is based on microbiology lab diagnoses, which were also used to train our machine learning classifier in all the experiments we were carried out to evaluate the performance of the classifier. Biochemical changes in blood contents resulting from different responses of the immune system to viral and bacterial infections lead to subtle spectral changes in the infrared spectra (Figure 1); these were monitored by infrared spectroscopy. In this research, several steps were carefully followed to optimize the quality of the measurements. First, we chose a sufficient concentration of WBC and plasma samples with optimum thickness for receiving a good SNR while the mercury-cadmiumtelluride detector is not saturated. Second, the representative spectrum of the sample is an average of 16 different spectra acquired from different sites of the sample spot. The SNR was high (more than 100), and a good reproducibility of the spectra was achieved, as shown in Figure 2. Although the manifested spectral differences between the average spectra corresponding to bacterial and viral infection of WBCs and plasma presented in Figure 1 are subtle, those changes are, nonetheless, sufficiently repeatable to yield a promising statistic for the classifications. Many studies have reported the advantages of using second derivative spectra, because they reveal hidden peaks on the shoulders of absorption bands 50,51. Moreover, using them as part of the procecedure could eliminate background differences between the spectra. We therefore adopted the use of second derivative spectral analysis in all of our experiments. First, we tried to differentiate between the bacterial infection cases and controls (healthy), and viral infection cases and controls. We used MLD and CMPD separately (Figure 3) in order to evaluate the potential of our classifier, as the classification is binary and is thus considered as a simple classification problem. Second, we tried to differentiate among bacterial infections, viral infections, and controls, simultaneously (Figure 4a for WBCs and Figure S-3 for plasma) as a first strategy. In the second strategy, we tried to differentiate between infected cases (bacterial and viral infections)


Page 14 of 27

Page 15 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


and controls (Figure 4b) as a first stage, and to classify the infection category into bacterial and viral infection sub-categories as a second stage (Figure 4c). Encouraging results were achieved based on using the LDA classifier; the performance of the classifier was better using WBC data compared to plasma in all the different experiments carried out in this study. This may be explained due to the heterogeneity (complexity) of the plasma. The spectral differences due to the type of infection are masked by the absorbance of other biomolecular contents in the plasma, which are inherent and not affected specifically by the infection type (viral or bacterial). The sensitivity performance of the classifier, based on the WBC spectra of CMPD, was better than the sensitivity using MLD for differentiation between the couples—bacterial-controls (Table 1), viral-controls (Table 1) and infection-controls (Table 2)—while for the specificity, performance based on MLD was higher than the specificity using the CMPD. These results may be explained as follows–some of the immune system responses due bacterial infections are similar to those due to viral infections. When a patient arrives at the ER with fever, sneezing, inflammation, vomiting, diarrhea, and fatigue that person almost definitely has either a bacterial or viral infection, so a physician’s general diagnosis of infection is true. As the CMPD (359) database is much larger than MLD (173), it yielded overall higher sensitivity. The higher specificity based on MLD, when compared to CMPD, may be explained due to a false diagnosis of the physician of patients that arrive at the ER and do not have any of the symptoms mentioned above as control (non-infected). The differences between the specificity based on MLD and CMPD are maximal when the infection is viral (Tables 1 and 2). This may be due to the larger number of patients arriving at the ER who are wrongly diagnosed as controls by the physician (Table 2) or vice versa. One reason for physicians to diagnose unknown infection cases as being viral occurs when there is no direct evidence of a bacterial infection according to the indicators known to them 52,53.



The analysis of loading 1 in Figure 5c shows that the spectral differences are widely spread over the entire spectrum, in the 800-1800 cm-1 region and it is difficult to relate these spectral differences to a specific biomolecule. In the second stage of the second strategy, when the differentiation is between viral and bacterial infections, the sensitivity and specificity of the classifier based on the MLD (82%,80%) are much higher than those derived based on CMPD (53%, 61%). This is probably due to the unreliable differentiation of the physician between viral and bacterial infections, as reported by previous studies 52-54. Here the first stage error (infection-control) was ignored because the sensitivity of this stage was higher than 98% (Table 2). The confidence level of our classification is high, based on the low p-value of 0.028, calculated using the permutation test. Although our preliminary results summarized in Table 2 are based on a relatively small number of patient samples, they are highly encouraging, with a promising potential to be improved by enlarging the database. Although the sensitivity and specificity obtained in our study are less than reported in the study of Eden et al

44

, their study is still limited as a routine diagnostic test for

determining the infection etiology because their determination of co-infections and false alarm of pathogenic infection resulted from natural flora. Thus, the gold standard database should be rebuilt and defined more carefully, to include only the most confident cases of infections or/and to determine a third case of co-infection (viral and bacterial). A larger database that includes more well-defined microbiology lab assessments will improve the statistics, especially when differentiating between viral and bacterial infection. This will make it more reliable, with confident conclusions 55,56 increasing the sensitivity and specificity of the determination of the infection etiology, by enabling the use of more sophisticated nonlinear classifiers.


Page 16 of 27

Page 17 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Conclusions (a) FTIR microcopy with multivariate analysis of peripheral blood tests enable objective determination of the etiology of infection as being viral or bacterial, with reasonable success rates. The analysis and diagnostic procedures require only a few minutes. (b) The present preliminary study suggests that FTIR spectroscopy of WBCs is potentially a feasible and efficient tool for the diagnosis of infection etiology in humans. (c) Increasing the number of laboratory diagnosis gold standard results (which is time consuming), may improve the results and increase the classification accuracy rates.

Conflict of Interest Disclosure There is no competing financial interest.

Figure captions Figure 1: Average spectra of controls, bacterial and viral infections. (a) plasma (b) WBC. The insets show the details of some of the spectral features for WBC and plasma that contribute to the classification Figure 2: 16 typical infrared absorption spectra measured from the same WBC sample in the region 1800-800 cm-1 after spectral manipulation. Figure 3: Classification success rates in percentage versus PCs number, calculated using LDA based on the MLD and CMPD of the WBC spectra in the low wavenumber 800-1800 cm-1 region. The differentiation was performed between the pairs: (a) bacterial-controls and (b) viral-controls categories. Figure 4: Classification accuracy in percentage versus PCs number calculated using LDA based on the low wavenumber 800-1800 cm-1 region for WBC spectra with MLD and CMPD. The



differentiation was performed between pairs: (a) bacterial-viral-controls (b) controlinfection, and (c) bacterial-viral categories. Figure 5: (a) Typical spectrum of WBC sample; (b) second derivative spectrum of 5(a); (c) Loading 1 calculated for the MLD data for the classification between bacterial and viral infection. The main peaks appearing in the loading 1 spectrum were labeled and correlated with the IR and the second derivative spectra.

Supporting Information Figure S-1: Schematic diagram illustrating the main steps carried out throughout the present study. In all the experiments, MLD was used to train the classifier. The validation was done using MLD, PDD and the combined MLD and PDD (abbreviated as CMPD) separately. Figure S-2: Individual spectra of controls, bacterial and viral infections of WBC. The spectra are similar and overlap with minute spectral differences in shapes and relative intensities of the absorption bands features. Figure S-3: Classification success rates in percentage versus PCs number calculated using LDA based on the low wavenumber 800-1800 cm-1 region of the spectra. The differentiation was performed among controls, bacterial and viral categories for plasma using MLD and CMPD. Figure S-4: PCA score plots of a) infection-control groups and b) viral-bacterial groups. All the scores were calculated based on the WBC spectra in the region 800-1800 cm-1 of the MLD data. The infection group includes the combined viral and bacterial data.


Page 18 of 27

Page 19 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


References (1) Balfour, L.; Tasca, G. A.; Kowal, J.; Corace, K.; Cooper, C. L.; Angel, J. B.; Garber, G.; Macpherson, P. A.; Cameron, D. W. Assessment 2007, 14, 408-416. (2) Perfect, M. M.; Bourne, N.; Ebel, C.; Rosenthal, S. L. Herpes : the journal of the IHMF 2005, 12, 38-41. (3) Banning, M. British journal of nursing (Mark Allen Publishing) 2005, 14, 1192-1197. (4) Miles, T. D.; McLaughlin, W.; Brown, P. D. BMC veterinary research 2006, 2, 7. (5) Abbas, A. K., Andrew H. Lichtman, and Shiv Pillai Cellular and Molecular Immunology, 9 ed.; Saunders/Elsevier: Philadelphia, 2010. (6) Helbert, M. Immunology for Medical Students, 3 ed.; Elsevier: UK, 2016. (7) Fox, D. A. In Microbiology and Immunology On-line, Hunt, R. C., Ed., 2016. (8) Perry, A. K.; Chen, G.; Zheng, D.; Tang, H.; Cheng, G. Cell research 2005, 15, 407-422. (9) Gendrel, D.; Raymond, J.; Coste, J.; Moulin, F.; Lorrot, M.; Guerin, S.; Ravilly, S.; Lefevre, H.; Royer, C.; Lacombe, C.; Palmer, P.; Bohuon, C. Pediatr Infect Dis J 1999, 18, 875-881. (10) Haran, J. P.; Beaudoin, F. L.; Suner, S.; Lu, S. The American journal of emergency medicine 2013, 31, 137-144. (11) Ahlers, A. A.; Schonheyder, H. C. Ugeskrift for laeger 1990, 153, 13-16. (12) Hamilton, J. L.; John, S. P. Am Fam Physician 2013, 87, 254-260. (13) Finnell, S. M.; Carroll, A. E.; Downs, S. M.; Subcommittee on Urinary Tract, I. Pediatrics 2011, 128, e749-770. (14) Rudinsky, S. L.; Carstairs, K. L.; Reardon, J. M.; Simon, L. V.; Riffenburgh, R. H.; Tanen, D. A. Acad Emerg Med 2009, 16, 585-590. (15) Melendez, E.; Harper, M. B. Acad Emerg Med 2010, 17, 163-167. (16) Van den Bruel, A.; Thompson, M. J.; Haj-Hassan, T.; Stevens, R.; Moll, H.; Lakhanpaul, M.; Mant, D. BMJ 2011, 342, d3082. (17) Stoll, M. L.; Rubin, L. G. Arch Pediatr Adolesc Med 2004, 158, 671-675. (18) Diem M; Griffiths PR; JM, C. Vibrational spectroscopy for medical diagnosis, 40 ed.; Wiley: Chichester, 2008. (19) Ostrovsky, E.; Zelig, U.; Gusakova, I.; Ariad, S.; Mordechai, S.; Nisky, I.; Kapilushnik, J. IEEE transactions on bio-medical engineering 2013, 60, 343-353. (20) Zelig, U.; Mordechai, S.; Shubinsky, G.; Sahu, R. K.; Huleihel, M.; Leibovitz, E.; Nathan, I.; Kapelushnik, J. Biochimica et biophysica acta 2011, 1810, 827-835. (21) Zelig, U.; Barlev, E.; Bar, O.; Gross, I.; Flomen, F.; Mordechai, S.; Kapelushnik, J.; Nathan, I.; Kashtan, H.; Wasserberg, N.; Madhala-Givon, O. BMC Cancer 2015, 15, 408. (22) Gajjar, K.; Trevisan, J.; Owens, G.; Keating, P. J.; Wood, N. J.; Stringfellow, H. F.; MartinHirsch, P. L.; Martin, F. L. The Analyst 2013, 138, 3917-3926. (23) Nabers, A.; Ollesch, J.; Schartner, J.; Kötting, C.; Genius, J.; Haußmann, U.; Klafki, H.; Wiltfang, J.; Gerwert, K. Journal of Biophotonics 2016, 9, 224-234. (24) Carmona, P.; Molina, M.; Lopez-Tobar, E.; Toledano, A. Anal Bioanal Chem 2015, 407, 77477756. (25) Carmona, P.; Molina, M.; Calero, M.; Bermejo-Pareja, F.; Martínez-Martín, P.; Toledano, A. Journal of Alzheimer's Disease 2013, 34, 911-920. (26) Erukhimovitch, V.; Talyshinsky, M.; Souprun, Y.; Huleihel, M. Journal of biomedical optics 2006, 11. (27) Njoroge, E.; Alty, S. R.; Gani, M. R.; Alkatib, M. Conference proceedings : ... Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual Conference 2006, 1, 5338-5341.



(28) Li, Q. B.; Li, X.; Zhang, G. J.; Xu, Y. Z.; Wu, J. G.; Sun, X. J. Guang pu xue yu guang pu fen xi = Guang pu 2009, 29, 1553-1557. (29) Sharaha, U.; Rodriguez-Diaz, E.; Riesenberg, K.; Bigio, I. J.; Huleihel, M.; Salman, A. Anal Chem 2017, 89, 8782-8790. (30) Salman, A.; Sharaha, U.; Rodriguez-Diaz, E.; Shufan, E.; Riesenberg, K.; Bigio, I. J.; Huleihel, M. Analyst 2017, 142, 2136-2144. (31) Mordechai, S.; Shufan, E.; Porat Katz, B. S.; Salman, A. Analyst 2017, 142, 1276-1284. (32) Salman, A.; Shufan, E.; Zeiri, L.; Huleihel, M. Biochim Biophys Acta 2013, 1830, 2720-2727. (33) Salman, A.; Shufan, E.; Lapidot, I.; Tsror, L.; Moreh, R.; Mordechai, S.; Huleihel, M. Analyst 2015, 140, 3098-3106. (34) Banfi, G.; Salvagno, G. L.; Lippi, G. Clinical chemistry and laboratory medicine 2007, 45, 565-576. (35) Whitaker, S.; Pigford, R. L. Industrial & Engineering Chemistry 1960, 52, 185-187. (36) Alvin C. Rencher; Christensen, W. F. Methods of Multivariate Analysis, 3 ed.; A JOHN WILEY & SONS: Hoboken New-Jersy, 2012. (37) Petibois, C.; Déléris, G. Vibrational Spectroscopy 2003, 32, 117-128. (38) Naumann, D.; Helm, D.; Labischinski, H. Nature 1991, 351, 81-82. (39) Movasaghi, Z.; Rehman, S.; ur Rehman, D. I. Applied Spectroscopy Reviews 2008, 43, 134179. (40) Mantsch, H.; Chapman, D. Infrared Spectroscopy of Biomolecules: Wiley–Liss, New York, 1996. (41) Duda, R. O.; Hart, P. E.; Stork, D. G. Pattern Classification (2nd Edition); Wiley-Interscience: New York, NY, USA, 2000. (42) Moor, K.; Ohtani, K.; Myrzakozha, D.; Zhanserkenova, O.; Andriana, B. B.; Sato, H. J Biomed Opt 2014, 19, 067003. (43) Salman, A.; Shufan, E.; Sahu, R. K.; Mordechai, S.; Sebbag, G. Vibrational Spectroscopy 2016, 83, 17-25. (44) Eden, E.; Srugo, I.; Gottlieb, T.; Navon, R.; Boico, O.; Cohen, A.; Bamberger, E.; Klein, A.; Oved, K. The Journal of infection 2016, 73, 177-180. (45) Caine, S.; Heraud, P.; Tobin, M. J.; McNaughton, D.; Bernard, C. C. A. NeuroImage 2012, 59, 3624-3640. (46) Baker, M. J.; Hussain, S. R.; Lovergne, L.; Untereiner, V.; Hughes, C.; Lukaszewski, R. A.; Thiefin, G.; Sockalingum, G. D. Chemical Society Reviews 2016, 45, 1803-1818. (47) Sahu, R. K.; Mordechai, S.; Salman, A.; Manor, E. In Imaging and Applied Optics; Optical Society of America: Arlington, Virginia, 2013, p FTu2D.5. (48) Christakis, N. A.; Lamont, E. B. West J Med 2000, 172, 310-313. (49) Fleming-Dutra, K. E.; Hersh, A. L.; Shapiro, D. J.; Bartoces, M.; Enns, E. A.; File, T. M., Jr.; Finkelstein, J. A.; Gerber, J. S.; Hyun, D. Y.; Linder, J. A., et al. JAMA 2016, 315, 1864-1873. (50) Baker, M. J.; Trevisan, J.; Bassan, P.; Bhargava, R.; Butler, H. J.; Dorling, K. M.; Fielden, P. R.; Fogarty, S. W.; Fullwood, N. J.; Heys, K. A., et al. Nat. Protocols 2014, 9, 1771-1791. (51) Rieppo, L.; Saarakkala, S.; Närhi, T.; Helminen, H. J.; Jurvelin, J. S.; Rieppo, J. Osteoarthritis and Cartilage 2012, 20, 451-459. (52) Craig, J. C.; Williams, G. J.; Jones, M.; Codarini, M.; Macaskill, P.; Hayen, A.; Irwig, L.; Fitzgerald, D. A.; Isaacs, D.; McCaskill, M. Bmj 2010, 340, c1594. (53) Caliendo, A. M.; Gilbert, D. N.; Ginocchio, C. C.; Hanson, K. E.; May, L.; Quinn, T. C.; Tenover, F. C.; Alland, D.; Blaschke, A. J.; Bonomo, R. A.; Carroll, K. C.; Ferraro, M. J.; Hirschhorn, L. R.; Joseph, W. P.; Karchmer, T.; MacIntyre, A. T.; Reller, L. B.; Jackson, A. F.; Infectious Diseases Society of, A. Clin Infect Dis 2013, 57 Suppl 3, S139-170. (54) Gonzales, R.; Malone, D. C.; Maselli, J. H.; Sande, M. A. Clinical Infectious Diseases 2001, 33, 757-762.


Page 20 of 27

Page 21 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


(55) Beleites, C.; Neugebauer, U.; Bocklitz, T.; Krafft, C.; Popp, J. Analytica Chimica Acta 2013, 760, 25-33. (56) Mu, X.; Kon, M.; Ergin, A.; Remiszewski, S.; Akalin, A.; Thompson, C. M.; Diem, M. Analyst 2015, 140, 2449-2464.



Page 22 of 27

for TOC only

Blood separation

FTIR spectroscopy

Objective Classification

Multivariate Analysis: PCA-LDA


Page 23 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 1: Average spectra of controls, bacterial and viral infections. (a) plasma (b) WBC. The insets show the details of some of the spectral features for WBC and plasma that contribute to the classification 153x86mm (300 x 300 DPI)



Figure 2: 16 typical infrared absorption spectra measured from the same WBC sample in the region 1800800 cm-1 after spectral manipulation. 153x86mm (300 x 300 DPI)


Page 24 of 27

Page 25 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 3: Classification success rates in percentage versus PCs number, calculated using LDA based on the MLD and CMPD of the WBC spectra in the low wavenumber 800-1800 cm-1 region. The differentiation was performed between the pairs: (a) bacterial-controls and (b) viral-controls categories. 153x86mm (300 x 300 DPI)



Figure 4: Classification accuracy in percentage versus PCs number calculated using LDA based on the low wavenumber 800-1800 cm-1 region for WBC spectra with MLD and CMPD. The differentiation was performed between pairs: (a) bacterial-viral-controls (b) control-infection, and (c) bacterial-viral categories. 153x86mm (300 x 300 DPI)


Page 26 of 27

Page 27 of 27 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 5: (a) Typical spectrum of WBC sample; (b) second derivative spectrum of 5(a); (c) Loading 1 calculated for the MLD data for the classification between bacterial and viral infection. The main peaks appearing in the loading 1 spectrum were labeled and correlated with the IR and the second derivative spectra. 152x85mm (300 x 300 DPI)


Differential Diagnosis of the Etiologies of Bacterial and Viral Infections

Recommend Documents