View Article Online

View Journal

Analyst
Accepted Manuscript

This article can be cited before page numbers have been issued, to do this please use: D. R. Parachalil, C.
Bruno, F. Bonnier, H. Blasco, I. Chourpa, M. J. Baker, J. McIntyre and H. Byrne, Analyst, 2019, DOI:
10.1039/C9AN00125E.

Volume 141 Number 1 7 January 2016 Pages 1–354 This is an Accepted Manuscript, which has been through the
Royal Society of Chemistry peer review process and has been
accepted for publication.
Analyst Accepted Manuscripts are published online shortly after
www.rsc.org/analyst

acceptance, before technical editing, formatting and proof reading.

Using this free service, authors can make their results available
to the community, in citable form, before we publish the edited
article. We will replace this Accepted Manuscript with the edited
and formatted Advance Article as soon as it is available.

You can find more information about Accepted Manuscripts in the

author guidelines.

Please note that technical editing may introduce minor changes

to the text and/or graphics, which may alter content. The journal’s
ISSN 0003-2654 standard Terms & Conditions and the ethical guidelines, outlined
PAPER
in our author and reviewer resource centre, still apply. In no
Michele Zagnoni et al.
Emulsion technologies for multicellular tumour spheroid radiation assays
event shall the Royal Society of Chemistry be held responsible
for any errors or omissions in this Accepted Manuscript or any
consequences arising from the use of any information it contains.

rsc.li/analyst
 Page 1 of 31 Analyst

1
2
3 Analysis of bodily fluids using Vibrational Spectroscopy: A direct comparison ofView Article Online
DOI: 10.1039/C9AN00125E
4
5
6 Raman scattering and Infrared absorption techniques for the case of glucose in blood
7
8 serum
9
10
11 Drishya Rajan Parachalila,b*, Clément Brunoc,d;e, Franck Bonnierc , Hélène Blascod,e, Igor
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13 Chourpac, Matthew J. Bakerf,g, Jennifer McIntyrea, and Hugh J. Byrnea

14
15
16 a FOCAS Research Institute, Technological University Dublin, Kevin Street, Dublin 8, Ireland
17
18 bSchool of Physics and Optometric & Clinical Sciences, Technological University Dublin,

Analyst Accepted Manuscript

19 Kevin Street, Dublin 8, Ireland
20
21 cUniversité de Tours, UFR sciences pharmaceutiques, EA 6295 Nanomédicaments et
22 Nanosondes, 31 avenue Monge, 37200 Tours, France.
23
d CHRU de Tours, Laboratoire de Biochimie et Biologie Moléculaire, Tours, France.
24
25
e Université de Tours, iBrain, UMR INSERM U1253, 37032, France.
26
27 e
28 WestCHEM, Department of Pure & Applied Chemistry, Technology and Innovation Centre,
29 University of Strathclyde, Glasgow, G1 1RD, UK
30 f
31
ClinSpec Diagnostics Ltd (ClinSpec Dx), Level 7, Technology and Innovation Centre, 99
32 George Street, Glasgow, G1, 1RD, UK
33
34
35
36
37 *Corresponding Author: drishyarajan.parachalil@mydit.ie
38
39
40
41
42 Abstract
43
44
45
46 Analysis of biomarkers present in the blood stream can potentially deliver crucial information
47
48 on patient health and indicate the presence of numerous pathologies. The potential of
49
50 vibrational spectroscopic analysis of human serum for diagnostic purposes has been widely
51
52
53 investigated and, in recent times, infrared absorption spectroscopy, coupled with ultra-filtration
54
55 and multivariate analysis techniques, has attracted increasing attention, both clinical and
56
57 commercial. However, such methods commonly employ a drying step, which may hinder the
58
59
60
clinical work flow and thus hamper their clinical deployment. As an alternative, this study

1
 Analyst Page 2 of 31

1
2
3 explores the use of Raman spectroscopy, similarly coupled with ultra-filtration and DOI:
multivariate
View Article Online
10.1039/C9AN00125E
4
5
6 analysis techniques, to quantitatively monitor diagnostically relevant changes of glucose in
7
8 liquid serum samples, and compares the results with similar analysis protocols using infrared
9
10 spectroscopy of dried samples. The analysis protocols to detect the imbalances in glucose using
11
12
Raman spectroscopy are first demonstrated for aqueous solutions and spiked serum samples.
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 As in the case of infrared absorption studies, centrifugal filtration is utilised to deplete abundant
16
17 analytes and to reveal the spectral features of Low Molecular Weight Fraction analytes in order
18

Analyst Accepted Manuscript

19
to improve spectral sensitivity and detection limits. Improved Root Mean Square Error of Cross
20
21
22 Validation (RMSECV) was observed for Raman prediction models, whereas slightly higher R2
23
24 values were reported for infrared absorption prediction models. Summarising, it is
25
26 demonstrated that the Raman analysis protocol can yield accuracies which are comparable with
27
28
29 those reported using infrared absorption based measurements of dried serum, without the need
30
31 for additional drying steps.
32
33
34 Keywords
35
36 Raman Spectroscopy; Glucose; Background correction; Extended Multiplicative Signal
37 Correction (EMSC); Centrifugal filtration
38
39
40
41
42
43 1. Introduction
44
45
46 Human bodily fluids (e.g. blood serum/ plasma, urine, saliva, tears and cerebrospinal fluid) are
47
48
49 considered to be a rich reservoir of clinical biomarkers and are an interesting alternative to cells
50
51 and tissues in terms of disease diagnosis and prognosis, owing to advantages such as minimal
52
53 invasiveness, low cost, and rapid sample collection and processing (1–4). The biochemical
54
55
56
composition of human serum can provide crucial information on patient health and indicate the
57
58 presence of numerous pathologies, as it encompasses a vast range of proteins and biochemical
59
60 products accumulated while perfusing various organs (5). Moreover, alterations in the

2
 Page 3 of 31 Analyst

1
2
3 biochemical composition of the serum/plasma could reflect changes of physiological
DOI:states due
View Article Online
10.1039/C9AN00125E
4
5
6 to disease, enabling early disease diagnosis and treatment (6–8). For example, it is reported
7
8 that carcinoembryonic antigens, CA 15-3 and CA 27.29, can be considered as serum
9
10 biomarkers for breast cancer (9), KL-40 and MMP-9 have been found to be potential serum
11
12
biomarkers for high grade gliomas (10), prostate specific antigen was found in higher levels in
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 the serum of patients with prostate cancer (11,12), and carcinoembryonic antigen has been
16
17 reported to be serum biomarker for colorectal cancer (13). In recent decades, extensive studies
18

Analyst Accepted Manuscript

19
have explored the detection of serum biomarkers for various diseases, and quite naturally the
20
21
22 reference tools used are the conventional, complex analytical techniques such as
23
24 chromatography, electrophoresis or mass spectroscopy (14–18). More recently, the field of
25
26 serum proteomics has exploded in the literature, as this emerging field has gained world-wide
27
28
29 attention due to its potential to reveal important information regarding the pathogenesis or
30
31 progression of diseases (19–22).
32
33 Analysis of serum biomarkers based on protein content is inherently challenging, because of
34
35
36
the low concentrations and the vast variety and dynamic range of protein abundance. The
37
38 complex nature of the serum poses as a huge problem for the detection of small molecule
39
40 biomarkers (2,23). Human serum contains more than 10,000 different proteins, with an overall
41
42
concentration ranging from 60-80mg/mL. Furthermore, circulating species present in the
43
44
45 serum, such as metabolites, peptides, sugars, and lipids, add to its complexity. Conventional
46
47 proteomic methods struggle to handle large dynamic range of abundances of its constituent
48
49 components (17). The characteristics of serum are usually dominated by the high molecular
50
51
52 weight fraction (HMWF) of proteins, which includes albumin (57-71%) and globulin (8-26%)
53
54 and masks the features of low molecular weight analytes which are present in trace amounts
55
56 (<5%) (24). Chromatographic fractionation can significantly enhance the sensitivity and
57
58
59
specificity of the data recorded, and coupling with spectroscopic analysis techniques such as
60

3
 Analyst Page 4 of 31

1
2
3 IR absorption in hybrid techniques such as Liquid Chromatography (LC-IR) or View
GasArticle Online
DOI: 10.1039/C9AN00125E
4
5
6 Chromatography (GC-IR) (25) can improve the performance further. Such chromatographic
7
8 techniques, although extensively used in pharmaceutical, chemical or food science applications
9
10 are time consuming and costly, however (26–28), and commercially available centrifugal filters
11
12
have been employed to deplete the HMWF before analysis to facilitate analysis of the low
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 molecular weight fraction (LMWF), including proteins and molecular biomarkers (29).
16
17
18 Vibrational spectroscopic techniques, both Raman and infrared absorption, have emerged over

Analyst Accepted Manuscript

19
20 the past 20 years as increasingly routine analytical techniques for a wide range of applications,
21
22
as they reveal specific biochemical information without the use of extrinsic labels. Although
23
24
25 they are often considered complementary techniques (30) there are also specific considerations
26
27 for each, for specific applications such as measurement of bodily fluids, as described by
28
29 Bonnier et al. (31). They provide intrinsic vibrational signatures of the material of interest in
30
31
32 a non-destructive fashion, and the potential for diagnostic applications has been well
33
34 demonstrated, notably in human serum and plasma (7,32–34). However, although both Raman,
35
36 Fourier-Transform Infrared (FTIR) and Attenuated Total Reflectance-FTIR (ATR-FTIR)
37
38
39
spectroscopy have been widely explored to study bodily fluids over the last two decades, most
40
41 of these studies have been carried out on air dried samples, in order to avoid the water
42
43 contribution in the case of FTIR, and to increase the concentration of the analytes in the case
44
45
of Raman (35–40). Two major limiting factors in the use of dried samples are the drying time
46
47
48 (41) and also the so-called “coffee-ring” effect, or, specifically in terms of blood serum, the
49
50 Vroman effect (42–44), whereby different analytes precipitate from solution at different rates,
51
52 giving rise to variations in the spectral features due to chemical and physical inhomogeneity.
53
54
55 Previous studies have clearly shown that the IR spectrum of dried (aggregates of) molecular
56
57 species is not the same as that in solution (45). It has further been shown that, for dried deposits
58
59 from solutions of varying concentrations of analyte, the linearity of the Beer-Lambert law, and
60

4
 Page 5 of 31 Analyst

1
2
3 therefore the quantitative nature of the measurement is compromised as the concentration of
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 the analyte is increased (45). In the case of the “coffee ring effect”, the thickness, and therefore
7
8 the quantitative accuracy of the measurement is spatially inhomogeneous, and the technique is
9
10 not ideally suited for quantitative measurements, such as those considered here. Notably,
11
12
Spalding et al. have used a serum dilution technique to improve the reliability of the technique
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 for quantitative measurement (46). The issues with inhomogeneity can potentially be overcome
16
17 by micropipetting and sampling the whole drop, and there have been studies that show excellent
18

Analyst Accepted Manuscript

19
specificity/sensitivity of classification of diseased state using dried samples (40,47,48).
20
21
22 Nevertheless, the requirement of a drying step adds considerably to the sampling time and
23
24 complexity of the workflow (41, 49).
25
26
27
28
29 Protocols for monitoring changes in HMWF serum proteins in their native liquid form using
30
31
32 Raman spectroscopy have recently been demonstrated, however, without the need for drying
33
34 (50). Analysing in the native liquid state, the chemical composition is averaged out by
35
36 molecular motion over the measurement time, greatly reducing the variability of the
37
38
39
measurement. Bonnier et al. compared FTIR and Raman for measuring gelatin in solution and
40
41 described protocols for isolation of LMWF from serum (31). This study indicated that Raman
42
43 in the liquid form could deliver similar sensitivities to ATR-FTIR for measurement of LMWF
44
45
species.
46
47
48 The aim of this study is to investigate the sensitivity and accuracy of Raman spectroscopy in
49
50 the liquid state, coupled with centrifugal filtration, as a biochemical tool to detect clinically
51
52 relevant changes in the biochemical composition of serum and compare to the ATR-FTIR
53
54
55 technique, using the specific example of glucose. The study is specifically designed according
56
57 to the ATR-FTIR study of glucose in serum by Bonnier et al., using identical parameters and
58
59 protocols of ultra-filtration and multivariate regression analysis, such that a direct comparison
60

5
 Analyst Page 6 of 31

1
2
3 of the two techniques can be made (29). The protocol is first demonstrated using aqueous
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 solutions and human serum spiked with systematically varied concentrations of glucose, before
7
8 exploring the sensitivity and accuracy of the technique in patient samples.
9
10
11
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13 2. Materials and Methods

14
15 2.1 Preparation of varying concentration of glucose in distilled water model
16
17
18 D-glucose (G8769) was purchased from Sigma Aldrich, Ireland and 6 glucose solutions were

Analyst Accepted Manuscript

19
20 prepared over the concentration range 100mg/dL to 1000mg/dL. Amicon Ultra 0.5mL
21
22
23 centrifugal filter devices (Millipore- Merck, Germany), with 10kDa cut off point, were
24
25 employed to concentrate and fractionate the serum samples. The centrifugation procedure
26
27 previously reported by Bonnier et al. was followed (31). A further study published by Bonnier
28
29
30 et al. reported 100% recovery of LMWF using 10kDa, hence only 10kDa cut off filtration has
31
32 been used in the present study (45). Pre-rinsing of the filter devices with 0.1M NaOH prior to
33
34 plasma analysis is essential to avoid glycerine interference in the analysis (32). The optimised
35
36
37
washing and rinsing procedure includes spinning 0.5mL 0.1M NaOH at 14000×g for 30
38
39 minutes followed by three rinses with distilled water by spinning 0.5mL distilled water for 30
40
41 minutes at 14000×g. Every 30 minute wash and rinse must be followed by spinning the device
42
43 in the inverted position at 1000×g for 2 minutes, to remove the residual solution contained in
44
45
46 the filter. After washing, 0.5mL of glucose solution is transferred to the 10kDa filter and
47
48 centrifuged at 14000×g for 30 minutes. The solution that passes through the 10kDa filter is the
49
50 filtrate, which contains mostly water and molecules smaller than 10kDa. The remainder of the
51
52
53 serum, known as the concentrate, is collected by placing the filter device upside down and
54
55 spinning for 1000×g for 2 minutes. The resultant concentrate, ~50µL, contains molecules with
56
57 molecular weight larger than 10kDa, concentrated by a factor of ~10, and can be employed for
58
59
60
study of the HMWF (50). All the filtrate solutions were analysed using Raman spectroscopy

6
 Page 7 of 31 Analyst

1
2
3 and five replicate measurements from different positions have been performed for each sample
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 of ~50µL. In subsequent analysis, each patient is represented by all the spectra recorded from
7
8 that patient, rather than the mean.
9
10
11 2.2 Preparation of glucose spiked in serum model
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14 Sterile, human serum (H6194) and D-glucose (G8769) were purchased from Sigma Aldrich,
15
16 Ireland, for the preparation of in vitro spiked models. The commercial human serum was spiked
17
18

Analyst Accepted Manuscript

19 with glucose in the concentration range that matches of the study of Bonnier et al. (29). Since
20
21 glucose is already present in normal human serum at concentrations of 70-110mg/dl (51), the
22
23 final concentration of the spiked samples covers the physiologically relevant ranges of normal
24
25
26
(80-120mg/dL) (52), and hyperglycaemia (>120mg/dL). The centrifugal processing step
27
28 described in Section 2.1 was performed on the spiked serum samples to obtain the filtrate. Note,
29
30 in the analysis of Figure 3, the regression is performed over spiked, rather than total glucose
31
32
concentration, consistent with the approach of Bonnier et al. (29).
33
34
35
36
37 2.3 Glucose levels in patient serum samples
38
39
Patient serum samples were donated by the University Hospital (CHU) Bretonneau de Tours
40
41 (France) and the ethical procedures were followed. The blood samples were collected from the
42
43 individuals as routine blood check-ups and 1 mL per patient was provided for spectroscopic
44
45
analysis. A total of 25 patient samples were included in the present study. Samples were
46
47
48 collected by personnel of CHU, under standard clinical protocols and approved ethical
49
50 procedures (Comité de Protection des Personnes, Tours Region Central Oest 1- PP/ANSM-
51
52 PHAO15-HB-METABOMU, registered internationally: ClinicalTrials.gov ID:
53
54
55 NCT02670226). The samples were serologically profiled, for other purposes, and the
56
57 anonymised, residual discard samples, along with their serological profiles were donated to the
58
59 Université François-Rabelais de Tours, for further study. No further specific ethical approval
60

7
 Analyst Page 8 of 31

1
2
3 or patient consent is required. Glucose concentrations were obtained by routine biochemical
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 analysis using a COBAS analyser, following the in house guidelines for routine biochemical
7
8 analysis. The principle of the test is based on the enzymatic reference method with hexokinase,
9
10 which catalyses the phosphorylation of glucose to glucose-6-phosphate by ATP methods
11
12
(53,54). Subsequently, glucose-6-phosphate is oxidised by glucose-6-phosphate
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 dehydrogenase, in the presence of NADP, to gluconate-6-phosphate. This reaction is specific,
16
17 with no other carbohydrate being oxidised. The rate of NADPH formation during the reaction
18

Analyst Accepted Manuscript

19
is directly proportional to the glucose concentration and is measured photometrically in the
20
21
22 UV. Measured glucose levels in the patient samples are listed in Table 1, and can be seen to
23
24 cover the range 55-435 mg dL-1. Note, that the distribution of glucose levels covers a broader
25
26 range than that of the study of similar (n=15) patients by Bonnier et al. (61-208 mg dL-1) (29).
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

8
 Page 9 of 31 Analyst

1
2
3 Table 1: List of measured glucose levels in patient samples. Glucose blood levelsDOI:
are10.1039/C9AN00125E
quoted
View Article Online

4
5
in terms of the SI unit of mmol L-1 as well as mg dL-1, commonly used in serology literature
6 and in the study of Bonnier et al.(29)
7
8
Sample number Glucose blood levels
9 mmol L-1 mg dL-1
10 1 2.9 52.25
11 2 3.1 55.85
12 3 3.9 70.27
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
4 3.9 70.27
14
15 5 4.2 75.67
16 6 4.2 75.67
17 7 4.3 77.47
18 8 4.5 81.08

Analyst Accepted Manuscript

19
9 5.1 91.89
20
21 10 5.1 91.89
22 11 5.2 93.69
23 12 5.7 102.70
24 13 6 109.10
25 14 6.1 109.90
26
27 15 6.4 115.31
28 16 6.4 115.31
29 17 7.2 129.72
30 18 8.8 158.55
31 19 9.9 178.37
32
33
20 11.5 207.20
34 21 11.7 210.81
35 22 13.4 241.44
36 23 15.7 282.88
37 24 15.8 284.68
38
39
25 24.1 434.23
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

9
 Analyst Page 10 of 31

1
2
3 2.4 Data collection using Raman spectrophotometer View Article Online
DOI: 10.1039/C9AN00125E
4
5 The measurement conditions used for screening HMWF proteins in solution have recently been
6
7
8
detailed (50). Raman spectra of all the liquid samples were recorded at stabilised room
9
10 temperature (18ºC) using a Horiba Jobin-Yvon LabRam HR800 spectrometer with a 16-bit
11
12 dynamic range Peltier cooled CCD detector. The spectrometer was coupled to an Olympus
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
1X71 inverted microscope and a x60 water immersion objective (LUMPlanF1, Olympus) was
15
16
17 employed. The substrate used was a Lab-Tek plate (154534) with a 0.16-0.19mm thick glass
18

Analyst Accepted Manuscript

19 bottom, 1.0 borosilicate cover glass, and was purchased from Thermo Fischer Scientific,
20
21 Ireland. Since the scattering efficiency is inversely proportional to the fourth power of the
22
23
24 wavelength, it is always desirable to use a short wavelength Raman source. In the filtered
25
26 serum, there are no molecular species which are resonant with 532nm, and therefore neither
27
28 fluorescence nor photodamage are limiting factors.
29
30
31
32
2.5 Data pre-processing and analysis
33
34 The raw spectra were subjected to pre-processing techniques in Matlab before further analysis,
35
36
37 to remove the background signal and reduce the noise. Smoothing of the raw data was done
38
39 using the Savitzky–Golay method, with a polynomial order of 5 and window 13 and the
40
41 Extended Multiplicative Signal Correction (EMSC) algorithm was applied to the smoothed
42
43
44 spectra to remove the underlying water spectrum, whose OH bending feature at 1640 cm-1 can
45
46 interfere with the protein spectra, particularly at low concentrations (55).
47
48 The principle of EMSC for subtraction of a specific measureable background spectrum and the
49
50
associated Matlab codes have previously been published by Kerr and Hennelly, 2016 (55), and
51
52
53 their description has been adapted in the following. In the case of measurement in aqueous
54
55 solution, the raw sample spectrum, S, consists of Raman spectrum of the analyte of interest, R,
56
57 a baseline signal, B, and the water signal, W.
58
59
60 S = R + B + W (55) (1)

10
 Page 11 of 31 Analyst

1
2
3 The Raman spectrum of interest can be represented by a reference spectrum of the analyte of
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 interest, r, and it can be assumed that R is the product of this reference spectrum and a certain
7
8 scalar weight, Cr, which describes the concentration dependence (56,57).
9
10 R ~ Cr x r (55) (2)
11
12
Similarly, a spectrum, w, is recorded from water directly in order to represent the spectral
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 contribution of water in W, as the product of pure water spectrum and a certain scalar weight.
16
17 W = Cw x w (55) (3)
18

Analyst Accepted Manuscript

19
The baseline, B, is now represented by a polynomial:
20
21
22 BN = C0 + C1X + C2X +……+ CNXN (55) (4)
23
24 where N is the order of polynomial and Cm for m = 0  N represents various coefficients of
25
26 polynomial (58). The EMSC algorithm is used to obtain estimates of the scalar values Cr, Cm
27
28
29 and Cw. These estimates are obtained from an optimal fit of the various vectors in Equation 5.
30
𝑁
31 S~ [𝐶𝑟 × 𝑟] + [𝐶𝑤 × 𝑤] +[∑𝑚 = 0𝐶𝑚𝑋𝑚 ] (55) (5)
32
33
34 The background corrected, concentration dependent analyte spectra, T, can be represented as:
35
𝑁
36 S ― [𝐶𝑤 × 𝑤] ― [∑𝑚 = 0𝐶𝑚𝑋𝑚]
37 T= 𝐶𝑤
(55) (6)
38
39
40 Note, that division by Cw has the effect of scaling the analyte spectra, assuming a constant
41
42 water contribution to all sample spectra. The Raman spectrum of the stock glucose solution
43
44 (450g/L) is used as the reference for EMSC, and polynomial of order 5 was used in all cases.
45
46
47
The glucose reference spectrum and water spectrum used in EMSC were recorded using the
48
49 532nm laser line and Lab-Tek plate as substrate. The x60 objective brings the laser to a focus
50
51 within the liquid, beyond the thin (0.16mm-0.19mm) glass substrate. No significant
52
53
contribution from the glass to the recorded spectrum was apparent, and therefore no correction
54
55
56 was required. The Rubberband method (59) was used to baseline correct the glucose reference
57
58 spectrum after smoothing it using the Savitzky–Golay method.
59
60

11
 Analyst Page 12 of 31

1
2
3 2.5.2 Partial Least Squares Regression View Article Online
DOI: 10.1039/C9AN00125E
4
5
6
7 Partial Least Squares Regression (PLSR) was employed to establish a model that relates the
8
9 variations of the spectral data to a series of concentrations. This regression model can be used
10
11 to improve the limit of detection of Raman bio-sensing (60). Constructed based on the spectra
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
of samples of known glucose content, either solutions of varying concentrations of glucose (in
15
16 water or in commercial serum), or those of the patient serum, the model is then validated using
17
18 a rigorous cross validation procedure which evaluates its performance in accurately predicting

Analyst Accepted Manuscript

19
20 glucose concentrations. A 20 fold cross validation approach has been employed to validate the
21
22
23 robustness of the method. This approach involves randomly dividing the set of observations
24
25 into approximately equal size, 50% of the spectral data were randomly selected as test set,
26
27 while the remaining 50% is used as the training set (61). In the current case, (5 x 25 spectra)
28
29
30 were divided into two groups of 65 (test) and 60 (training) spectra. The cross-validation process
31
32 is then repeated 20 times (the folds), with all observations are used for both training and testing,
33
34 and each observation is used for testing exactly once. The results from the folds can then be
35
36
averaged to produce a single estimation. The Root Mean Square Error of Cross Validation
37
38
39 (RMSECV) is calculated from the 20 iterations to measure the performance of the model for
40
41 the unknown cases within the calibration set. The correlation between the concentration and
42
43 spectral intensity is given by the R2 value. The standard deviation was calculated to find the
44
45
46 variation between each spectrum calculated from the same sample. The number of latent
47
48 variables used for building the PLSR model is optimised by finding the value that is equivalent
49
50 to the minimum of the RMSECV
51
52
53
54
55
56 3. Results
57
58
59
60

12
 Page 13 of 31 Analyst

1
2
3 In this study, Raman spectra of the samples were recorded in the inverted geometry using a
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 x60 water immersion objective with a 532nm laser and the Lab-Tek plate was used as the
7
8 substrate. The 532nm laser was chosen as it is compatible with the thin glass bottomed Lab-
9
10 Tek plate and provides a strong Raman signal of the sample with minimal background
11
12
interference. This set-up was previously reported by Bonnier et al. (31) to yield better analysis
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 of serum using Raman spectroscopy when the sample was analysed in the inverted geometry
16
17 using a water immersion objective with a 785nm laser and CaF2 substrate. The added advantage
18

Analyst Accepted Manuscript

19
of this setup is that it provides high quality, consistent Raman spectra from sample volumes as
20
21
22 low as 1μL. The protocol using 532nm was more recently further explored for analysis of
23
24 HMWF serum proteins (50).
25
26
27 Figure 1 presents the spectra of the fingerprint region of the pure glucose solution recorded in
28
29 the inverted geometry. The raw spectra of the glucose were baseline corrected using the
30
31
32 rubberband method and smoothed using the Savitzky–Golay algorithm (polynomial 5, window
33
34 13). Example signature peaks of glucose (indicated by asterisks) appear at ~450cm-1, associated
35
36 with an endocyclic δ(C-C-O) ring mode, δ(C1-H1) vibration at 911cm-1, a peak at 1060cm-1
37
38
39
due to ν(C1-OH) stretching, a relatively sharp peak at 1125cm-1 which can be assigned to the
40
41 δ(C-O-C) angle-bending mode and the sharp peaks at 1340cm-1 and 1460cm-1, related to the
42
43 δ(C-C-H) vibration and pure CH2 group vibration, respectively(62).
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

13
 Analyst Page 14 of 31

1
2
3 View Article Online
10 4 DOI: 10.1039/C9AN00125E
4 3.5
5 *
6 3
7
8 2.5
*
9

Intensity (a.u)
10 2

11 *
1.5
12
* * *
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
1
14
15 0.5
16
17 0
18

Analyst Accepted Manuscript

19 -0.5
20 400 600 800 1000 1200 1400 1600 1800

21 -1
Raman shift(cm )
22
23
24 Fig.1. Raman spectrum of an aqueous glucose solution, concentration 450g/L. Example
25
26 signature peaks at 450cm-1, 911cm-1, 1125cm-1, 1340cm-1 and 1460cm-1 are labelled in the
27
28
29 figure.
30
31
32 3.1 Monitoring the concentration dependence of glucose in distilled water
33
34
35
In order to establish the analysis protocol for the patient samples, a PLSR prediction model
36
37
38 was first built and applied to the set of varying concentrations of glucose in distilled water, as
39
40 well as spiked into serum. The first step in this study was to optimise the measurement protocol
41
42 and also to evaluate the efficacy of the centrifugal filtration technique in separating and
43
44
45 concentrating glucose from the HMWF proteins. For this, different amounts of pure glucose
46
47 were spiked into the distilled water. The normal glucose concentration and the concentration
48
49 and hyper-glycaemia were deliberately included to simulate physiologically relevant
50
51
52
concentrations. Figure 2 A displays examples of the EMSC (polynomial order 5) corrected low
53
54 (5 x100mg/dL), medium (5 x 450mg/dL) and high (5 x 1000mg/dL) datasets, showing glucose
55
56 features (indicated by asterisks) of greater intensity at higher concentration, whereas the lower
57
58
concentrations shows weaker features of glucose. The groups of spectra are off set for clarity.
59
60

14
 Page 15 of 31 Analyst

1
2
3 The concentration 1000mg/dL was deliberately included in the dataset to evaluate the
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 consistency of the glucose spectral features as the concentration increases from 100mg/dL to
7
8 1000mg/dL. In order to analyse the spectral variations and the glucose concentrations, the
9
10 PLSR algorithm was applied.
11
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15
16
17
18

Analyst Accepted Manuscript

19 1800 80

20
AA *
1600
75 B
21
22
1400
* 70

*
RMSECV (mg/dL)
1200
23
* * * 65
Intensity (a.u)

24 1000

25 800
60

26 600 55
27
28
400
50

29 200
45
30 0

31 -200
40
0 2 4 6 8 10 12 14 16 18 20
32 400 600 800 1000 1200 1400 1600 1800
Number of components
33 Raman shift(cm -1)
0.16
34 1000

35 0.14
C 900 D
36 0.12 800
37
Predicted (mg/dL)

0.1 700
38
Intensity (a.u)

39 0.08 600

40 0.06
500

41 400
0.04
42 300
43 0.02
200
44 0

45 100
-0.02
46 400 600 800 1000 1200 1400 1600 1800 0
0 100 200 300 400 500 600 700 800 900 1000
47 Raman shift(cm -1)
Measured (mg/dL)
48
49
50 Fig. 2. (A): EMSC corrected Raman spectra of filtrate obtained after centrifugal filtration with
51 10kDa filters of varying concentrations of glucose (5 x100mg/dL, 5 x 450mg/dL and 5 x
52
53 1000mg/dL, offset for clarity), in distilled water and signature peaks of glucose are highlighted
54 with asterisks, (B): Evolution of the RMSECV on the validation model, (C): plot of PLSR
55 coefficient with glucose features, (D): Predictive model built from the PLSR analysis. The
56 value displayed in the PLSR model is an average of the concentration predicted with the
57
58 corresponding standard deviation calculated from the 20 iterations of the cross validation. The
59 RMSECV and R2 values were calculated as 10.93mg/dL and 0.9705 respectively.
60

15
 Analyst Page 16 of 31

1
2
3 View Article Online
DOI: 10.1039/C9AN00125E
4
5 Based on the percent variance explained by the latent variables and the minimum value of
6
7
8
RMSECV (Figure 2B), the optimum number of latent variables to reach the best performance
9
10 is determined to be 4. The PLSR coefficient plot displayed in Figure 2C confirms the
11
12 correlation of the data in Figure 2D is based on glucose features, such as the peaks at ~1060cm1,
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
~1125cm1, 1450cm-1 and ~1340cm-1. Finally, after selecting the optimum number of
15
16
17 components for the data set analysed, a predictive model is built from the PLSR analysis
18

Analyst Accepted Manuscript

19 (Figure 2D), to compare the observations to the known concentrations of glucose in the samples
20
21 with the estimated concentrations from the spectral data sets. Figure 2D indicates that a
22
23
24 satisfactory linear model could be obtained with the raw data set and that the concentration
25
26 dependence of the sample set is conserved by centrifugal filtration. From the results shown in
27
28 figure 2, it is evident that the predicted values are in good agreement with the reference
29
30
31
concentrations and the corresponding correlation coefficient (R2) is calculated as 0.9705. Note,
32
33 each concentration point has five independent measurements, and the mean standard deviation
34
35 of each measurement is 4.8mg/dL. The RMSECV calculated from the 20 iteration of cross
36
37
validation is 10.93mg/dL,, thereby indicating that PLSR provides accurate predictions for the
38
39
40 glucose Raman data over the entire concentration range of 100mg/dL to 1000mg/dL.
41
42
43
44
45 3.2 Monitoring the glucose concentration in spiked serum
46
47
48 In an attempt to extend the optimised protocol to a more complex environment, PLSR analysis
49
50 was performed on the EMSC corrected data set recorded from the filtrate of the serum samples
51
52 with spiked glucose concentrations varying from 0mg/dL to 220mg/dL. When the entire
53
54
55
fingerprint region was selected for PLSR analysis, the resultant PLSR coefficient displayed a
56
57 negative peak at ~1000cm-1 which could potentially derive from other LMWF species such as
58
59 urea (63) (Figure S1C in Supplemental). Therefore, the spectral region from 1030cm-1 to
60

16
 Page 17 of 31 Analyst

1
2
3 1400cm-1, which contains strong glucose features at ~1050cm-1 and at~1340cm-1 but minimal
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 interference from urea, has been chosen for PLSR analysis (62).
7
8
9
10
11
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14 1400 60
15
A * * B
16 1200
50
17 *
1000
18

RMSECV (mg/dL)

Analyst Accepted Manuscript

40
19
Intensity (a.u)

800

20 600 30
21
22 400
20
23
200
24 10
25 0

26 0
-200
27 1000 1050 1100 1150 1200 1250 1300 1350 1400 0 2 4 6 8 10 12 14 16 18 20

28 Raman shift(cm -1) Number of components

29 0.14 250
30
C D
31 0.12
200
32
0.1
Predicted (mg/dL)

33
PLSR coefficient

34 0.08
150

35
36 0.06
100
37
38 0.04

39 0.02
50

40
41 0 0
1000 1050 1100 1150 1200 1250 1300 1350 1400
42 -1
0 50 100 150 200 250

43 Raman shift(cm ) Measured (mg/dL)

44
45 Fig. 3. (A): EMSC corrected Raman spectra of filtrate obtained after centrifugal filtration with
46 10kDa filters of glucose spiked in serum (spiked concentrations 5 x 0mg/dL, 5 x 120mg/dL
47 and 5 x 220mg/dL, offset for clarity) and the signature peaks of glucose are highlighted by
48 asterisks, (B): Evolution of RMSECV of the data set (C): plot of PLSR coefficient with glucose
49
50
features, (D): Predictive model built from the PLSR analysis. The value displayed in the PLSR
51 model is an average of the concentration predicted with the corresponding standard deviation
52 calculated from the 20 iterations of the cross validation. The RMSECV and R2 values were
53 calculated as 1.66mg/dL and 0.9914
54
55
56
57
58
59
60

17
 Analyst Page 18 of 31

1
2
3 Figure 3A shows the glucose data set after background correction using the EMSC
DOI:algorithm
View Article Online
10.1039/C9AN00125E
4
5
6 (offset for clarity by 5.0 units). The optimum number of latent variables were chosen by
7
8 calculating the lowest value of RMSECV (Figure 3B). Six latent variables were chosen for this
9
10 model and the resultant PLSR coefficient exhibits strong glucose features, as shown in Figure
11
12
3C. A linear predictive model can be defined from the EMSC corrected data set of varying
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 concentration of glucose in serum Figure 3D. The RMSECV was found to be 1.66mg/dL and
16
17 R2 value was calculated as 0.9914. The mean standard deviation of each measurement from 20
18

Analyst Accepted Manuscript

19
iterations of cross validation was calculated to be 3.2mg/dL. The results suggest that this
20
21
22 optimised protocol can be applied to the patient samples to build a quantitative model.
23
24
25 Bonnier et al. have previously illustrated the strategy of centrifugal filtration to fractionate the
26
27 samples to eliminate the influence of HMW in order to screen potential LMWF biomarkers
28
29 using glucose as a model analyte, measured by ATR-FTIR [33]. The results of the analysis of
30
31
32 similarly centrifugally filtered, glucose spiked samples of commercial serum are reproduced in
33
34 Table 2, and directly compared to the results of the current study using Raman spectroscopic
35
36 analysis in the native liquid state. Although each method yields similar R2 values, the RMSECV
37
38
39
of Raman spectroscopic analysis is significantly lower, which indicates an increased accuracy
40
41 of the measurement protocol in the liquid state.
42
43
44 Table 2. Comparison of the results of ATR-FTIR and Raman spectroscopic analysis of human
45 serum spiked with varying concentrations of glucose.
46
47
48
49
50 Measurement Concentration RMSECV(mg/dL) Standard R2
51 type range (mg/dL) deviation
52
53 FTIR (Min- 0-220 2.199 0.250 0.995
54
Max
55
56
normalised)
57
58 Raman 0-220 1.665 3.2 0.991
59 spectroscopy
60

18
 Page 19 of 31 Analyst

1
2
3 View Article Online
DOI: 10.1039/C9AN00125E
4
5
6
7
8 3.3 Monitoring the glucose concentration in patient samples
9
10
11 Figure 4A displays the Raman spectra from patient samples after performing background
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
correction using the EMSC algorithm, in the reduced spectral range of 1030-1400 cm-1. The
14
15
16 spectra are offset for clarity by 5.0 units. The glucose bands at 1340cm-1 and 1460cm-1, related
17
18 to the δ(C-C-H) vibration and pure CH2 group vibration, and the peak at 1060cm-1 due to ν(C1-

Analyst Accepted Manuscript

19
20 OH) stretching can be clearly seen. Based on the minimum RMSECV value (Figure 4B), 12
21
22
23 latent variables were found to be optimal for constructing a PLSR based model. The PLSR
24
25 coefficient clearly shows glucose features (Figure 4 C), indicating that the prediction is based
26
27 on the variation in the glucose peak intensities. In the prediction model of Figure 4D, the
28
29
30
minimum value of RMSECV is 1.84mg/dL and the R2 value is calculated as 0.82, indicating a
31
32 high prediction capacity. The mean standard deviation was determined to be 3.48mg/dL,
33
34 indicating acceptable repeatability between the cross validation iterations.
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

19
 Analyst Page 20 of 31

1
2
3 View Article Online
DOI: 10.1039/C9AN00125E
4
5 1400
*
57

6 A * 56 B
*
1200
7
55
8 1000

RMSECV (mg/dL)
9 54
Intensity (a.u)

800
10 53

11 600
52
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

400
13 51

14 200 50

15 49
0
16
48
17 -200 0 2 4 6 8 10 12 14 16 18 20
1000 1050 1100 1150 1200 1250 1300 1350 1400
18 Number of components

Analyst Accepted Manuscript

-1
10 -3 Raman shift(cm )
19 6

C
500
20
D
Predicted concentrations(mg/dL)
450
21 5

400
22
23 350
PLSR Coefficient

24 300

25 3
250

26 200
27 2
150
28
100
29 1

30 50

31 0
1000 1050 1100 1150 1200 1250 1300 1350 1400
0
0 50 100 150 200 250 300 350 400 450 500
32 Raman shift(cm -1) Observations (mg/dL)
33
34 Fig. 4. (A): EMSC corrected Raman spectra of filtrate obtained after centrifugal filtration with
35 10kDa filters of patient samples (5 x 52.25mg/dL, 5 x 75.67mg/dL, 5 x 93.69mg/dL, 5 x
36
210.81mg/dL and 5 x 434.35mg, offset for clarity) and the signature peaks are marked by
37
38 asterisks, (B): Evolution of RMSECV of the data set, (C): plot of PLSR coefficient with glucose
39 features, (D): Predictive model built from the PLSR analysis. The value displayed in the PLSR
40 model is an average of the concentration predicted with the corresponding standard deviation
41
calculated from the 20 iterations of the cross validation The RMSECV and R2 values were
42
43 calculated as 1.84mg/dL and 0.84 respectively.
44
45
46
47 Table 3. Comparison of the results of ATR-FTIR (29) and Raman spectroscopic analysis of
48 patient sample set for monitoring the glucose levels. FTIR results are normalised.
49
50
51
Measurement Concentration RMSECV(mg/dL) Standard R2
52 type range (mg/dL) deviation(mg/dL)
53
54
55
56 ATR-FTIR 61.25-210 3.1 1.90 0.9957
57 (Min-Max
58 normalised)
59
60

20
 Page 21 of 31 Analyst

1
2
3 Raman 52.25-210 1.6 2.31 0.91DOI: 10.1039/C9AN00125E
View Article Online

4
5
Spectroscopy
6
7
8 52.25-440 1.84 3.48 0.84
9
10
11
12
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14 Table 3 directly compares the results of analysis of a similar patient sample-set, with glucose
15
16
17
levels which varied over the range 0-210mg/dL, similarly centrifugally filtered and analysed
18

Analyst Accepted Manuscript

19 using ATR-FTIR (29) and the results of the current study of patient samples over the same
20
21 (patients 1-21, Table 1) and extended (patients 1-25, Table 1) range using Raman spectroscopic
22
23
analysis in the native liquid state. It is noteworthy that Raman spectroscopy yields significantly
24
25
26 lower values of RMSECV for all the Raman prediction models, suggesting higher sensitivity
27
28 and accuracy. However, the standard deviation was found to be higher and the R2 value was
29
30 found to be considerably lower for the shorter as well as larger concentration range. The
31
32
33 reduced R2 value and higher standard of deviation could be attributed to the variability in the
34
35 spectral response of the patient samples. Nevertheless, the precision of the model expressed by
36
37 the RMSECV values indicates the suitability of this technique to discriminate patients with
38
39
40
very close concentrations of blood glucose. The results demonstrate that Raman spectroscopy
41
42 is able to detect subtle variations in the glucose concentrations with similar accuracy in the
43
44 native liquid state, to the ATR-FTIR method in dried samples.
45
46
47 In order for a glucose detection method to be viable, it should be able to detect glucose in the
48
49 clinically relevant range (10-450mg/dL). Post PLSR analysis, the dataset is presented in the
50
51
52 Clarke’s error grid (Figure 5), the most common standard for evaluating the performance of a
53
54 glucose detection method used since 1987 (64). Data points that fall in zone A and B are
55
56 acceptable values. Values that fall outside A and B result in erroneous diagnosis. On the
57
58
59 Clarke’s error grid of patient samples (Figure 5), 98% of the PLSR validation dataset falls
60

21
 Analyst Page 22 of 31

1
2
3 within zone A and B, which is the zone of clinical accurate measurement with no
DOI:effect on
View Article Online
10.1039/C9AN00125E
4
5
6 clinical actions. Error can be attributed to the intrinsic variability of patient samples which
7
8 reflects their physiological state on the day. However, the results from this study are promising,
9
10 indicating that Raman spectroscopy coupled with multivariate analysis and centrifugal
11
12
filtration techniques can be used as a biochemical tool for detecting potential small biomarkers
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 from human serum/plasma.
16
17
Clarke's Error Grid Analysis
Predicted Concentration [mg/dl]

18

Analyst Accepted Manuscript

19
400
20
21
E C B
22
23 300
24
25 B
26
27 200
28
29 D
30
D
31 100
32
33
34 A C E
35 0
36 0 100 200 300 400
37
38 Reference Concentration [mg/dl]
39
40 Fig.5. PLSR validation of patient samples on Clarke’s error grid. The RMSECV was found to
41 be 1.84 mg/dL and R2 value was calculated as 0.84
42
43
44
45
46
47 4. Discussion
48
49 The potential advantages of using vibrational spectroscopy for biomarker assessment in bodily
50
51 fluids have been extensively explored in the last two decades. However, little consideration has
52
53
54
been given so far to protocols involving Raman analysis in the native liquid state of proteins.
55
56 Liakat et al. reported the in vitro prediction of physiologically relevant concentration of glucose
57
58 using mid-IR transmission light with respect to a Clarke error grid (52). Near IR and Raman
59
60

22
 Page 23 of 31 Analyst

1
2
3 spectroscopy was also used for measurement of glucose from artificial plasmaDOI:with high
View Article Online
10.1039/C9AN00125E
4
5
6 precision and accuracy (65). Although these are initial steps towards developing Raman
7
8 spectroscopy into a biochemical tool for serum/plasma analysis, measurements should be
9
10 performed on patient samples in order to ensure the relevancy of these methods. Using Raman
11
12
spectroscopy as a biochemical tool, it has been possible to detect differences in peak intensities
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 of altered serum compared to normal ones for glucose and lipid compounds (66),
16
17 multicomponent blood analysis (67) and also to determine blood glucose concentration of
18

Analyst Accepted Manuscript

19
blood samples with above-physiological levels of glucose within 5 min (68).
20
21
22
Using glucose as a model analyte, this study successfully demonstrated the feasibility of
23
24
25 employing Raman spectroscopy for detecting small biomarkers in serum after depletion of
26
27 HMWF proteins. It has been shown that optimal experimental set up for Raman analysis for
28
29 this experiment is Lab-Tek plates as substrate and measurement in the inverted geometry using
30
31
32 water immersion objective and the sample volume can be as small as 1μL. This experimental
33
34 set up is advantageous for clinical purposes where the volumes of patient samples are minimal.
35
36 After the depletion of the abundant proteins, the dominant water peak from the filtrate collected
37
38
39
after centrifugal filtration using 10kDa can be removed by using the EMSC algorithm, and
40
41 PLSR analysis applied to obtain a prediction model relating the glucose concentrations and the
42
43 intensity of glucose features. Even though the EMSC algorithm removed the underlying water
44
45
spectra effectively, there could be interference from other LMWF analytes, namely, urea (7-
46
47
48 20mg/dL). Thus, as presented in the present study, the spectral range from 1030cm-1 to 1400cm-
49
50 1 was chosen for data analysis, as this region does not contain signature peaks of urea.
51
52
53 The depletion of HMWF proteins using centrifugal filtration to detect glucose in serum using
54
55 ATR-FTIR was previously reported by Bonnier et al. (29). While the work carried out by
56
57
58 Bonnier et al. showed excellent results, the requirement of a drying step is potentially a major
59
60 drawback. Indeed, the drying process required for ATR-FTIR has been identified by Cameron

23
 Analyst Page 24 of 31

1
2
3 et al. as a potentially significant impediment to translation of the techniqueDOI:
to10.1039/C9AN00125E
clinical
View Article Online

4
5
6 applications (41). Since Raman is compatible with aqueous samples, sample drying can be
7
8 avoided and data can be recorded from the native environment, which makes the proposed
9
10 method an ideal alternative to IR. The results summarised in Table 2 and Figure 3 suggest that
11
12
Raman spectroscopy maintained high level of accuracy and predictive power and the
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 relationship between spectral variation and protein concentrations is linear, with minimal
16
17 standard deviation. The PLSR model built on varying concentration of glucose spiked in serum
18

Analyst Accepted Manuscript

19
provided an accurate prediction model (R2=0.9914, RMSECV= 1.66mg/dL) after applying pre-
20
21
22 processing steps using the EMSC based algorithm. Having established the optimal sample
23
24 preparation and analysis protocol using the spiked serum model, the same protocol was applied
25
26 to patient samples. In the case of patient samples, RMSECV and R2 values were calculated to
27
28
29 be 1.84mg/dL, and 0.84 respectively. Although ATR-FTIR provides better standard deviation
30
31 and higher R2 value, the Raman prediction model gives lower a RMSECV value, indicating
32
33 higher accuracy. The PLSR coefficient plot shows clear features of glucose, indicating the
34
35
36
prediction model is built on variations in glucose concentrations. The added advantage of
37
38 Raman spectroscopy is that the analysis can be performed on liquid samples and no additional
39
40 time delays associated with sample drying are introduced. Although the quantitative capability
41
42
of Raman can be easily demonstrated using glucose spiked with a pooled serum model, the
43
44
45 analysis of patient samples can be more complicated, the reason being the intrinsic variability
46
47 of individual samples depending upon the physiological state of the individual on that day.
48
49
50 For the purposes of a direct comparison of the two techniques, glucose concentration in human
51
52 serum was chosen, and the results detailed in Table 3 indicate that Raman in the liquid state
53
54
55 provides higher accuracy than FTIR-ATR. Although vibrational spectroscopic techniques are
56
57 unlikely to replace current techniques for routine glucose monitoring, in a more general sense,
58
59 an argument for serological applications of vibrational spectroscopic techniques has been made
60

24
 Page 25 of 31 Analyst

1
2
3 (28, 37). The advantages of using Raman spectroscopy over common biochemical
DOI:assays to
View Article Online
10.1039/C9AN00125E
4
5
6 quantify urea and creatinine has previously been reported(63). Moreover, Raman spectroscopy
7
8 as a biochemical tool for serum analysis is cost effective, rapid and a non-destructive method
9
10 as compared to currently employed gold standard clinical methods such as spectrophotometric
11
12
analysis ( e.g. COBAS analyser) (69). The COBAS analyser has a standard deviation of 0.04
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 mmol.L-1 (which is equivalent to 0.721 mg.dL-1), as mentioned in the Material and Methods
16
17 section, and the lower detection limit is 2mg/dL (70) with a correlation coefficient of 0.975
18

Analyst Accepted Manuscript

19
reported for immunoassays (71) and R2 of 0.990 for urea, creatinine, sugar, total protein and
20
21
22 calcium (72). Similar R2 values were calculated for pure glucose solutions and glucose spiked
23
24 serum solutions using Raman spectroscopy, suggesting that Raman spectroscopy is well-suited
25
26 for routine use as a biochemical tool for glucose analysis. However, the analysis using COBAS
27
28
29 analyser is complex and uses various enzymatic reagents, increasing the cost and chance of
30
31 inaccuracy in the results obtained. Employing well trained personnel to operate the equipment
32
33 enhances the reliability but increases the cost. Hence, Raman spectroscopy offers several
34
35
36
advantages as it is a onetime investment, easy to operate and provides rapid results with wider
37
38 information without destroying the medium. This could be translated as an alternative method
39
40 for glucose monitoring, especially in the case of hyperglycaemia. Further studies need to be
41
42
conducted to investigate other LMWF metabolites from human serum using Raman
43
44
45 spectroscopy. However, to further ensure relevancy of the results, the study should ultimately
46
47 be conducted on large number of patient samples.
48
49
50 5 Conclusion:
51
52
53 Summarising, the work presented showcases the development of an optimal methodology for
54
55 the detection of LMWF analytes from human serum using Raman spectroscopy with minimal
56
57
58 sample preparation steps and without the use of any extrinsic labels. The proposed approach
59
60 can be expeditiously employed for early detection of pathological disorders associated with

25
 Analyst Page 26 of 31

1
2
3 high or low serum/plasma proteins/biomarkers. Disease diagnosis from bodily fluids can be
View Article Online
DOI: 10.1039/C9AN00125E
4
5
6 developed into a dynamic diagnostic environment that will enable early disease diagnosis even
7
8 before the disease becomes symptomatic. Thus, analysis of bodily fluids has emerged as one
9
10 of the promising approaches to deliver crucial information about patient health and monitor
11
12
disease progression and/or therapy. Given the remarkable advances in the field over the last
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15 two decades, including sample preparation, protein fractionation, quantitation and
16
17 chemometrics, it is conceivable that vibrational spectroscopic techniques can be developed as
18

Analyst Accepted Manuscript

19
a point-of-care disease monitoring system. Ultimately, the proof of concept presented in this
20
21
22 study can be easily transferable to any other low molecular weight biomarkers or therapeutic
23
24 drugs.
25
26
27 Conflicts of Interest
28
29
30 There are no conflicts to declare.
31
32
33 Acknowledgements
34
35
36 Drishya Rajan Parachalil was funded by DIT Fiosraigh scholarship. J. McIntyre was funded by
37
38 Science Foundation Ireland, PI/11/08. The Irish-French PHC Ulysses 2018 Collaborative
39
40
41
project is jointly funded by the Irish Research Council and Campus France with participation
42
43 of the French embassy in Ireland.
44
45
46
47
48 References
49
50 1. Veenstra TD, Conrads TP, Hood BL, Avellino AM, Ellenbogen RG, Morrison RS.
51 Biomarkers: Mining the Biofluid Proteome. Mol Cell Proteomics. 2005;4(4):409–18.
52
53 2. Kong K, Kendall C, Stone N, Notingher I. Raman spectroscopy for medical
54 diagnostics - From in-vitro biofluid assays to in-vivo cancer detection. Adv Drug Deliv
55 Rev. 2015;89:121–34.
56
57 3. Su S-B, Chuen T, Poon W, Thongboonkerd V. Human Body Fluid. Biomed Res Int.
58 2013;2013:2–4.
59
60 4. Busher JT. Serum Albumin and Globulin. Clin Methods Hist Phys Lab Exam.

26
 Page 27 of 31 Analyst

1
2
3 1990;497–9. View Article Online
DOI: 10.1039/C9AN00125E
4
5 5. Greening DW, Simpson RJ. Low-Molecular Weight Plasma Proteome Analysis Using
6 Centrifugal Ultrafiltration. Methods Mol Biol. 2011;728:109-24.
7
8 6. Parker CE, Borchers CH. Mass spectrometry based biomarker discovery , verification ,
9 and validation--Quality assurance and control of protein biomarker assays. Mol Oncol.
10
2014;8(4):840–58.
11
12 7. Baker MJ, Hughes CS, Hollywood KA. Biophotonics: Vibrational Spectroscopic
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
Diagnostics. Morgan & Claypool Publishers; 2016.
14
15 8. Li J, Zhang Z, Rosenzweig J, Wang YY, Chan DW. Proteomics and Bioinformatics
16 Approaches for Identification of Serum Biomarkers to Detect Breast Cancer. Clin
17
18
Chem. 2002 Aug;48(8):1296-304

Analyst Accepted Manuscript

19 9. Bast RC, Ravdin P, Hayes DF, Bates S, Fritsche H, Jessup JM, et al. 2000 Update of
20
21
Recommendations for the Use of Tumor Markers in Breast and Colorectal Cancer:
22 Clinical Practice Guidelines of the American Society of Clinical Oncology. J Clin
23 Oncol. 2001 Mar 15;19(6):1865–78.
24
25 10. Hormigo A, Gu B, Karimi S, Riedel E, Panageas KS, Edgar MA, et al. YKL-40 and
26 matrix metalloproteinase-9 as potential serum biomarkers for patients with high-grade
27 gliomas. Clin Cancer Res. 2006 Oct;12(19):5698–704.
28
29 11. Labrie F, Dupont A, Suburu R, Cusan L, Tremblay M, Gomez JL, et al. Serum prostate
30 specific antigen as pre-screening test for prostate cancer. J Urol. 1992 Mar;147(3 Pt
31 2):842–6.
32
33 12. Catalona WJ, Smith DS, Ratliff TL, Dodds KM, Coplen DE, Yuan JJ, et al.
34 Measurement of prostate-specific antigen in serum as a screening test for prostate
35 cancer. N Engl J Med. 1991 Apr;324(17):1156–61.
36
37 13. Locker GY, Hamilton S, Harris J, Jessup JM, Kemeny N, Macdonald JS, et al. ASCO
38 2006 update of recommendations for the use of tumor markers in gastrointestinal
39 cancer. J Clin Oncol. 2006 Nov;24(33):5313–27.
40
41 14. Liesenfeld DB, Habermann N, Owen RW, Scalbert A, Ulrich CM. Review of mass
42 spectrometry-based metabolomics in cancer research. Cancer Epidemiol Biomarkers
43 Prev. 2013;22(12):2182–201.
44
45 15. Kimhofer T, Fye H, Taylor-Robinson S, Thursz M, Holmes E. Proteomic and
46 metabonomic biomarkers for hepatocellular carcinoma: A comprehensive review. Br J
47
48
Cancer. 2015;112(7):1141–56.
49 16. Nowak M, Janas Ł, Stachowiak G, Stetkiewicz T, Wilczyński JR. Current clinical
50
application of serum biomarkers to detect ovarian cancer. Prz Menopauzalny.
51
52 2015;14(4):254–9.
53
17. Luque-Garcia JL, Neubert TA. Sample preparation for serum/plasma profiling and
54
55 biomarker identification by mass spectrometry. J Chromatogr A. 2007;1153(1–2):259–
56 76.
57
58
18. Fernandez-Olavarria A, Mosquera-Perez R, Diaz-Sanchez R, Serrera-Figallo M,
59 Gutierrez-Perez J, Torres-Lagares D. The role of serum biomarkers in the diagnosis
60 and prognosis of oral cancer: A systematic review. J Clin Exp Dent. 2016;8(2):184-

27
 Analyst Page 28 of 31

1
2
3 193. View Article Online
DOI: 10.1039/C9AN00125E
4
5 19. Sahab ZJ, Semaan SM, Sang Q-XA. Methodology and applications of disease
6 biomarker identification in human serum. Biomark Insights. 2007;2:21–43.
7
8 20. Adkins JN, Varnum SM, Auberry KJ, Moore RJ, Angell NH, Smith RD, et al. Toward
9 a Human Blood Serum Proteome. Mol Cell Proteomics. 2002;1(12):947–55.
10
11 21. Lundblad RL. Considerations for the use of blood plasma and serum for proteomic
12 analysis. Internet J Genomics Proteomics. 2005;1(2):1–8.
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14 22. Pieper R, Gatlin CL, Makusky AJ, Russo PS, Schatz CR, Miller SS, et al. The human
15 serum proteome: Display of nearly 3700 chromatographically separated protein spots
16 on two-dimensional electrophoresis gels and identification of 325 distinct proteins.
17
18
Proteomics. 2003;3(7):1345–64.

Analyst Accepted Manuscript

19 23. Lacombe C, Untereiner V, Gobinet C, Zater M, Sockalingum GD, Garnotel R. Rapid
20
21
screening of classic galactosemia patients: A proof-of-concept study using high-
22 throughput FTIR analysis of plasma. Analyst. 2015;140(7):2280–6.
23
24. Di Girolamo F, Alessandroni J, Somma P, Guadagni F. Pre-analytical operating
24
25 procedures for serum Low Molecular Weight protein profiling. J Proteomics.
26 2010;73(3):667–77.
27
28 25. Patel KN, Patel JK, Patel MP, Rajput GC, Patel HA. Introduction to hyphenated
29 techniques and their applications in pharmacy. Pharm Methods. India: Medknow
30 Publications & Media Pvt Ltd; 2010;1(1):2–13.
31
32 26. Kuligowski J, Cascant M, Garrigues S, De La Guardia M. An infrared spectroscopic
33 tool for process monitoring: Sugar contents during the production of a depilatory
34 formulation. Talanta. 2012;99:660–7.
35
36 27. Edelmann A, Diewok J, Baena JR, Lendl B. High-performance liquid chromatography
37 with diamond ATR–FTIR detection for the determination of carbohydrates, alcohols
38 and organic acids in red wine. Anal Bioanal Chem. 2003;376(1):92–7.
39
40 28. Ioannou A. Real Time Monitoring the Maillard Reaction Intermediates by HPLC-
41 FTIR. J Phys Chem Biophys. 2016;6(2):6–10.
42
43 29. Bonnier F, Blasco H, Wasselet C, Brachet G, Respaud R, Carvalho LFCS, et al. Ultra-
44 filtration of human serum for improved quantitative analysis of low molecular weight
45 biomarkers using ATR-IR spectroscopy. Analyst. 2017;142(8):1285–98.
46
47 30. Byrne H, Sockalingum G, Stone N. Raman Microscopy: Complement or Competitor.
48 Biomed Appl Synchrotron Infrared Microspectrosc. 2011;(11):105–42.
49
50 31. Bonnier F, Petitjean F, Baker MJ, Byrne HJ. Improved protocols for vibrational
51 spectroscopic analysis of body fluids. J Biophotonics. 2014;7(3–4):167–79.
52
53 32. Bonnier F, Baker MJ, Byrne HJ. Vibrational spectroscopic analysis of body fluids:
54 avoiding molecular contamination using centrifugal filtration. Anal Methods.
55 2014;6(14):5155.
56
57 33. Bunaciu AA, Fleschin Ş, Hoang VD, Aboul-Enein HY. Vibrational Spectroscopy in
58 Body Fluids Analysis. Crit Rev Anal Chem. 2017;47(1):67–75.
59
60 34. Mitchell AL, Gajjar KB, Theophilou G, Martin FL, Martin-Hirsch PL. Vibrational

28
 Page 29 of 31 Analyst

1
2
3 spectroscopy of biofluids for disease screening or diagnosis: Translation from
DOI: the
View Article Online
10.1039/C9AN00125E
4
5
laboratory to a clinical setting. J Biophotonics. 2014;7(3–4):153–65.
6 35. Nyuwi KT, Gyan Singh CH, Khumukcham S, Rangaswamy R, Ezung YS, Chittvolu
7
SR, et al. The role of serum fibrinogen level in the diagnosis of acute appendicitis. J
8
9 Clin Diagnostic Res. 2017;11(1):PC13-PC15.
10
36. Tekin IO, Pocan B, Borazan A, Ucar E, Kuvandik G, Ilikhan S, et al. Positive
11
12 correlation of CRP and fibrinogen levels as cardiovascular risk factors in early stage of
continuous ambulatory peritoneal dialysis patients. Ren Fail. 2008;30(2):219–25.
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15
37. Stec JJ, Silbershatz H, Tofler GH, Matheney TH, Sutherland P, Lipinska I, et al.
16 Association of fibrinogen with cardiovascular risk factors and cardiovascular disease
17 in the Framingham Offspring Population. Circulation. 2000;102(14):1634–8.
18

Analyst Accepted Manuscript

19 38. Ariëns RAS. Elevated fibrinogen causes thrombosis. Blood. 2011;117(18): 4687–8.
20
21
39. Hong LF, Li XL, Luo SH, Guo YL, Zhu CG, Qing P, et al. Association of fibrinogen
22 with severity of stable coronary artery disease in patients with type 2 diabetic mellitus.
23 Dis Markers. 2014;2014:485687.
24
25 40. Paraskevaidi M, Morais CLM, Lima KMG, Snowden JS, Saxon JA. Differential
26 diagnosis of Alzheimer ’ s disease using spectrochemical analysis of blood. PNAS.
27 2017; 114 (38):929-38
28
29 41. Cameron JM, Butler HJ, Palmer DS, Baker MJ. Biofluid spectroscopic disease
30 diagnostics: A review on the processes and spectral impact of drying. J Biophotonics.
31 2018;11(4):1–12.
32
33 42. Shaw RA, Kotowich S, Leroux M, Mantsch HH. Multianalyte Serum Analysis Using
34 Mid-Infrared Spectroscopy. Ann Clin Biochem. 1998 Sep 1;35(5):624–32.
35
36 43. Gajjar K, Trevisan J, Owens G, Keating PJ, Wood NJ, Stringfellow HF, et al. Fourier-
37 transform infrared spectroscopy coupled with a classification machine for the analysis
38 of blood plasma or serum: a novel diagnostic approach for ovarian cancer. Analyst.
39 2013;138(14):3917–26.
40
41 44. Shaw RA, Low-Ying S, Man A, Liu K-Z, Mansfield C, Rileg CB, et al. Infrared
42 Spectroscopy of Biofluids in Clinical Chemistry and Medical Diagnostics. In:
43 Biomedical Vibrational Spectroscopy. John Wiley & Sons, Inc.; 2007. p. 79–103.
44
45 45. Bonnier F, Brachet G, Duong R, Sojinrin T, Respaud R, Aubrey N, et al. Screening the
46 low molecular weight fraction of human serum using ATR-IR spectroscopy. J
47
48
Biophotonics. 2016;9(10):1085–97.
49 46. Spalding K, Bonnier F, Bruno C, Blasco H, Board R, Benz-de Bretagne I, et al.
50
Enabling quantification of protein concentration in human serum biopsies using
51
52 attenuated total reflectance – Fourier transform infrared (ATR-FTIR) spectroscopy.
53 Vib Spectrosc. 2018;99:50–8.
54
55 47. Nabers A, Perna L, Lange J, Mons U, Schartner J, Güldenhaupt J, et al. Amyloid blood
56 biomarker detects Alzheimer ’ s disease. EMBO Mol Med. 2018;(May):1–11.
57
58
48. Roy S, Perez-Guaita D, Andrew DW, Richards JS, McNaughton D, Heraud P, et al.
59 Simultaneous ATR-FTIR Based Determination of Malaria Parasitemia, Glucose and
60 Urea in Whole Blood Dried onto a Glass Slide. Anal Chem. 2017 May

29
 Analyst Page 30 of 31

1
2
3 16;89(10):5238–45. View Article Online
DOI: 10.1039/C9AN00125E
4
5 49. Hands JR, Abel P, Ashton K, Dawson T, Davis C, Lea RW, et al. Investigating the
6 rapid diagnosis of gliomas from serum samples using infrared spectroscopy and
7
cytokine and angiogenesis factors. Anal Bioanal Chem. 2013;405(23):7347–55.
8
9 50. Parachalil DR, Brankin B, McIntyre J, Byrne HJ. Raman spectroscopic analysis of
10
high molecular weight proteins in solution – considerations for sample analysis and
11
12 data pre-processing. Analyst . 2018;143(24):5987–98.
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
51. Daines TL, Morse KW. Determination of Glucose in Blood Serum. J Chem Educ.
14
15
1976;53(2):126–7.
16 52. Liakat S, Bors KA, Huang T, Michel APM, Zanghi E, Gmachl CF, et al. In vitro
17
18
measurements of physiological glucose concentrations in biological fluids using mid-

Analyst Accepted Manuscript

19 infrared light. Biomed Opt Express. 2013;4(7):233–9.
20
21
53. Illingworth J. Methods of enzymatic analysis: Third edition. Biochemical
22 Education.1985;13(1):38
23
54. Bell C. Clinical Guide to Laboratory Tests. 3rd edition. Norbert W. Tietz, ed.
24
25 Transfusion. Wiley/Blackwell; 2018 Sep 17;35(11):972.
26
55. Kerr LT, Hennelly BM. A multivariate statistical investigation of background
27
28 subtraction algorithms for Raman spectra of cytology samples recorded on glass slides.
29 Chemom Intell Lab Syst. 2016;158:61–8.
30
31
56. Kohler A, Kirschner C, Oust A, Martens H. Extended multiplicative signal correction
32 as a tool for separation and characterization of physical and chemical information in
33 Fourier transform infrared microscopy images of cryo-sections of beef loin. Appl
34 Spectrosc. 2005 Jun;59(6):707–16.
35
36 57. Liland KH, Kohler A, Afseth NK. Model-based pre-processing in Raman spectroscopy
37 of biological samples. J Raman Spectrosc. 2016;47(6):643–50.
38
39 58. Joss L, Müller EA. Machine Learning for Fluid Property Correlations: Classroom
40 Examples with MATLAB. J Chem Educ. 2019; 0(0).
41
42 59. Hen XIS, Iang LXU, Hubin SYE, Ong RHU, In LINGJ, Anyang HXU, et al.
43 Automatic baseline correction method for the open-path Fourier transform infrared
44 spectra by using simple iterative averaging. Opt Expres. 2018;26(10):609–14.
45
46 60. Wold S, Sjöström M, Eriksson L. PLS-regression: A basic tool of chemometrics.
47 Chemom Intell Lab Syst. 2001;58(2):109–30.
48
49 61. Russell S, Norvig P. Artificial Intelligence: A Modern Approach. 3rd ed. Upper Saddle
50 River, NJ, USA: Prentice Hall Press; 2009.
51
52 62. Söderholm S, Roos YH, Meinander N, Hotokka M. Raman spectra of fructose and
53 glucose in the amorphous and crystalline states. J Raman Spectrosc.
54 1999;30(11):1009–18.
55
56 63. Saatkamp CJ, de Almeida ML, Bispo JAM, Pinheiro ALB, Fernandes AB, Silveira L.
57 Quantifying creatinine and urea in human urine through Raman spectroscopy aiming at
58 diagnosis of kidney disease. J Biomed Opt. 2016;21(3):037001.
59
60 64. Clarke WL, Cox D, Gonder-Frederick LA, Carter W, Pohl SL. Evaluating Clinical

30
 Page 31 of 31 Analyst

1
2
3 Accuracy of Systems for Self-Monitoring of Blood Glucose. Diabetes Care.DOI:
1987 Sep
View Article Online
10.1039/C9AN00125E
4
5
1;10(5):622 LP-628.
6 65. Xue J, Chen H, Xiong D, Huang G, Ai H, Liang Y, et al. Noninvasive Measurement of
7
Glucose in Artificial Plasma with Near-Infrared and Raman Spectroscopy. Applied
8
9 Spectroscopy, 2014;68(4):428–33.
10
66. Cássia R De, Borges F, Navarro RS, Giana HE, Tavares FG, Fernandes AB, et al.
11
12 Detecting alterations of glucose and lipid components in human serum by near-
infrared Raman spectroscopy. Res Biomed Eng. 2015;31(2):160–8.
Published on 04 April 2019. Downloaded by University of California - Santa Barbara on 4/7/2019 1:18:23 PM.

13
14
15
67. Berger AJ, Koo T, Itzkan I, Horowitz G, Feld MS. Multicomponent blood analysis by
16 near-infrared Raman spectroscopy. Appl Opt. 1999;38(13):2916-26
17
18
68. Berger AJ, Itzkan I, Feld MS. Feasibility of measuring blood glucose concentration by

Analyst Accepted Manuscript

19 near-infrared Raman spectroscopy. Spectrochim Acta A Mol Biomol Spectrosc.
20 1997;53:287–92.
21
22 69. Imai K. Clinical Chemistry and Immunoassay Testing Supporting the Individual
23 Healthy Life. Hitachi Review. 2008;57:1–7.
24
25 70. Pas S, Molenkamp R, Schinkel J, Rebers S, Copra C, Seven-Deniz S, et al.
26 Performance evaluation of the new Roche cobas AmpliPrep/cobas TaqMan HCV test,
27 version 2.0, for detection and quantification of hepatitis C virus RNA. J Clin
28 Microbiol. 2013 Jan;51(1):238–42.
29
30 71. Gammeren AJ Van, Gool N Van, Groot MJM De, Christa M. Analytical performance
31 evaluation of the Cobas 6000 analyzer – special emphasis on trueness verification.
32 Clin Chem Lab Med. 2008;46(6):863–71.
33
34 72. Ahmed A, Alam JM, Ali H, Sultana I, Fazal S. Comparative precision analysis of
35 Precision-controls on automated chemistry analyzers. International Journal of
36 Innovative Science Engineering and Technology. 2015;2(8):1-13.
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60

31