Spectrochimica Acta Part B 146 (2018) 106–114

Contents lists available at ScienceDirect

Spectrochimica Acta Part B

journal homepage: www.elsevier.com/locate/sab

Using LIBS to diagnose melanoma in biomedical fluids deposited on solid T

substrates: Limits of direct spectral analysis and capability of machine
learning
⁎
Rosalba Gaudiusoa,f, Ebo Ewusi-Annana, Noureddine Melikechia, , Xinzi Sunb, Benyuan Liub,
Luis Felipe Campesatoc, Taha Merghoubc,d,e
a
Department of Physics and Applied Physics, Kennedy College of Sciences, University of Massachusetts Lowell, MA 01854, United States
b
Department of Computer Sciences, Kennedy College of Sciences, University of Massachusetts Lowell, MA 01854, United States
c
Ludwig Collaborative and Swim Across America Laboratory, New York, NY 10065, United States
d
Parker Institute for Cancer Immunotherapy, New York, NY 10065, United States
e
Department of Medicine, Memorial Sloan Kettering Cancer Center, New York, NY 10065, United States
f
Institute of Nanotechnology, CNR-NANOTEC, c/o Department of Chemistry, University of Bari, 70126 Bari, Italy

A R T I C LE I N FO A B S T R A C T

Keywords: Diagnosis is crucial to increase the success rate of cancer treatments as well as the survival rate and life quality of
LIBS patients, in particular for forms of cancer that remain largely asymptomatic until metastasis. Methodologies that
Diagnosis of cancer allow the diagnosis of early-stage tumors as well as the detection of residual disease have the potential to
Melanoma improve cancer control and help monitor therapeutic outcomes. In this work, we report a Laser-Induced
Machine learning
Breakdown Spectroscopy (LIBS) approach to early diagnosis of a form of skin cancer, melanoma, based on the
Biological fluid analysis
analysis of biological fluids (blood and tissue homogenates) harvested from diseased mice and healthy controls.
Effect of substrate
We acquired femtosecond LIBS spectra and used two different approaches for the analysis: through comparison
of the emission intensity of selected analytes in healthy and diseased samples; and by using machine learning
classification algorithms (LDA, Linear Discriminant Analysis; FDA, Fisher Discriminant Analysis; SVM, Support
Vector Machines; and Gradient Boosting). We also addressed the effect of substrates on the analysis of liquid
samples, by using four different substrates (PVDF, Cu, Al, Si) and comparing their performance. We show that
with a combination of the most appropriate substrate and algorithm, we are able to discriminate between
healthy and diseased mice with accuracy up to 96% while direct analysis of LIBS spectra did not provide any
conclusive results. These series of results demonstrate that carefully designed LIBS measurements combined with
machine learning can be a powerful and practical approach for the diagnosis of cancer.

1. Introduction can significantly improve significantly the success of treatments and

ultimately the survival rate and quality of life of patients. This issue is
Cancer indicates a class of diseases related to abnormal cell growth particularly critical for kinds of cancer that develop in the absence of
in one organ or tissue, with the potential to spread to other parts of the specific symptoms and can go largely unnoticed until they metastasize,
body. Many research and medical efforts are ongoing to efficiently di- such as epithelial ovarian cancer (EOC), pancreatic cancer, and mela-
agnose and fight cancer, but the various forms of this disease are still noma [1,2].
one of the leading causes of death worldwide. Fighting cancer is very Developing large-scale screening tests is one of the most efficient
complex, in that it necessarily involves many different aspects, such as: strategies for early diagnosis of this kind of tumors. Ideally, such tests
investigating the causes of its onset; developing minimally invasive, should be rapid and minimally invasive, user-friendly, accurate (low
targeted, and, in the future, personalized therapy approaches; pro- number of false positives and false negatives), and easy to integrate in
moting prevention practices; implementing screening tests for early point-of-care structures, so as to reach and monitor large numbers of
diagnosis. The latter is a key task, as it is well-documented that de- people on a periodic basis. Laser-Induced Breakdown Spectroscopy
tecting the onset of the disease during its early stage of development (LIBS) is characterized by well-known practical advantages, which

⁎
Corresponding author.
E-mail address: [email protected] (N. Melikechi).

https://doi.org/10.1016/j.sab.2018.05.010
Received 8 March 2018; Received in revised form 23 April 2018; Accepted 11 May 2018
0584-8547/ © 2018 Published by Elsevier B.V.
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

include limited sample preparation, fast multi-elemental response, detection (Andor Technology, DH734-18F O3). We acquired spectra
compact instrumentation, possibility of in situ analyses, and versatility, with 50 ns starting delay time after the laser pulse and 700 μs integra-
all of which can contribute to making this technique a powerful tool in tion time.
the fight against cancer. We analyzed two different kinds of biological fluids, obtained from
Despite being essentially an atomic spectroscopy technique, and as mice with melanoma and healthy controls: blood serum and homo-
such not an obvious choice for the diagnosis of diseases that proceed genates of three different tissues (lungs, spleen, lymph nodes), that
through an abnormal proliferation of cells, LIBS has proved useful to were prepared following the protocols described later on.
distinguish between biopsied cancerous tissues and adjacent healthy In the first series of experiments, we deposited 5 μl drops of each of
ones, thanks to differences in the content of trace elements. In parti- the biological fluids on PVDF membranes, and dried them for 10 min
cular, previous studies have almost consistently shown that cancerous prior to the LIBS analysis, using a tungsten IR lamp. The laser energy we
lesions have a different alkaline and alkaline earth metals content than used for this series of experiments was 1.44 mJ. In the second series of
healthy tissues [3–6]. experiments, we selected only one of the biological fluids (blood serum)
In [7], Melikechi et al. proposed for the first time to develop a LIBS- and studied the effect of different substrates on the LIBS spectra and
based “liquid biopsy” approach for the early detection of cancer, i.e. the classification accuracy. The sample preparation procedure was the
analyzed samples were not tissues (either biopsied or harvested from same, and the three employed substrates were Cu, Al, and Si. Prior to
laboratory animals), but sera. In [7], we acquired femtosecond-LIBS depositing serum, the two metallic substrates were mechanically po-
spectra of mice sera taken from animals with EOC and healthy controls lished, and all three substrates were rinsed and sonicated in 2-propanol.
of three different age groups and deposited on a solid substrate. The The laser energy we used for this series of experiments was 1.20 mJ. We
LIBS spectra were then fed to two different classification algorithms constantly monitored the laser energy during the spectra acquisition,
that were shown to be useful for the discrimination of sera from mice and spectra with intensity lower than a given threshold were auto-
with cancer and healthy ones with a maximum accuracy around 80%. matically rejected, so as to improve the signal-to-noise ratio.
Recently, Chen et al. have adopted essentially the same experimental
and computational approach to the diagnosis of lymphoma and mul- 3. Sample preparation
tiple myeloma in human serum, and have obtained classification ac-
curacies close to 100% [8]. 3.1. Mice
In the present work, we obtain LIBS spectra of blood serum and
tissue homogenates harvested from a pre-clinical model of melanoma Mouse experiments were performed in accordance with institutional
(B16-F10), an often asymptomatic and highly aggressive type of skin guidelines under a protocol approved by the Memorial Sloan Kettering
cancer (see for example [9,10]). We aim for three goals. First, to in- Cancer Center (MSKCC) Institutional Animal Care and Use Committee.
vestigate if analogous conclusions to those obtained on melanoma [6] All mice were maintained in a pathogen-free facility according to the
can be drawn from this work. In other words, is the enrichment in Mg National Institutes of Health Animal Care guidelines. C57BL/6J mice
and Ca in melanoma lesions with respect to healthy skin reported in (females, 6 to 10 weeks old) were purchased from The Jackson
skin tissues also present in biological fluids? Second, we compare the Laboratory.
performance of four different classification algorithms (LDA, Linear
Discriminant Analysis; FDA, Fisher Discriminant Analysis; SVM, Sup- 3.2. Cell line and tumor implantation
port Vector Machines; and Gradient Boosting). Third, for the first time
we address the effect of four different solid substrates (Polyvinylidene The B16-F10 mouse melanoma line was originally obtained from I.
fluoride (PVDF); Cu; Al; Si) on the signal-to-noise ratio of LIBS spectra Fidler (MD Anderson Cancer Center, Houston, TX). Cells were main-
of biological fluids and on the classification accuracy. We show that the tained in RPMI 1640 containing 7.5% fetal bovine serum (FBS) and L-
classification accuracy varies significantly with the type of algorithm glutamine. For the tumor implantation, 2 × 105 viable B16-F10 cells in
selected and of substrate used. In this particular study, the best results 100 μl of phosphate-buffered saline (PBS) were injected intradermally
are obtained with Gradient Boosting using a copper substrate, which into the right flank of C57BL/6 mice.
yields a classification accuracy close to 100%. In essence, this work
demonstrates the power of this approach even when a direct LIBS 3.3. Sample preparation protocol
spectral analysis of cancerous and non-cancerous samples does not
provide any indication of distinguishability. Mice were euthanized 2 weeks after tumor implantation, and four
different tissues (lymph nodes; spleen; lungs; blood serum) were har-
2. Experimental setup vested from tumor-bearing or non-tumor-implanted mice. Blood serum
was analyzed as such, while tissues were mechanically dissociated
The laser source and the spectroscopic system have been previously using a PowerGen 125 tissue homogenizer (Fisher Scientific) in a pro-
described [7], and will only be briefly recalled here. The plasma was tein lysis buffer (LB), with the following composition: 0.01 M Tris-HCl,
produced by focusing a 150-fs Ti-Sapphire laser (Clark-MXR, Model 0.15 M NaCl, 0.01 M MgCl2, 0.5% NP-40 in distilled water.
2210, wavelength = 775 nm) on the samples, through a fused silica
biconvex lens (focal length = 50 mm, focused spot size = 100 μm). The 4. Results and discussion
samples were mounted on a motorized and computer-controlled x-y
translation stage (scanning speed = 0.35 mm/s), to ensure that each Previous LIBS studies for cancer diagnosis have almost consistently
laser shot would ablate a fresh surface. Measurements were performed shown that cancerous tissues have a different elemental composition
in an experimental chamber filled with slight over-pressured He than healthy tissues, these differences being mostly due to the alkaline
(762 Torr), in order to reduce the spectral interference from air ele- and alkaline earth metals content. Ca and Mg levels have usually been
ments and obtain a more persistent and bright plasma. The optical found to be higher in the cancer-affected areas, with at least four dif-
emission from the plasma was collected by a fiber collimation lens 45° ferent kinds of tumor (colorectal cancer [3], breast cancer [3,4], canine
with respect to the laser beam and focused onto a 50 μm core-diameter hemangiosarcoma [5], and melanoma [6]). An exception to this trend
optical fiber, and coupled with the spectroscopic acquisition system. was observed in [11], where no Ca and Mg enrichment was found in
The latter comprised an Echelle spectrograph (Andor Technology, ME Malignant Pleural Mesothelioma (MPM), but instead MPM tissues re-
5000) for wavelength dispersion and a thermoelectrically cooled iStar sulted to be enriched in P and O and depleted in Zn and Cu. Following
Intensified Charge Coupled Device (ICCD) camera for radiation these observations, we set out to determine whether statistically

107
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

meaningful differences would be detectable in the elemental composi- affected by the tumor) versus in biopsied cancer lesions may be ex-
tion of blood serum and tissue homogenates harvested from mice with pected. However, this specific aspect is beyond the scope of this work.
melanoma and healthy controls. In terms of normalization, as all three methods yield similar observa-
In the present paper, we used two approaches to evaluate LIBS- tions, in the following we report results obtained with only one of them:
based liquid biopsy, i.e., to discriminate between samples harvested the result obtained by dividing by the total area of the spectra.
from healthy and diseased animals: the first was a classical LIBS ap- The results obtained for the tissue homogenates, shown in
proach, based on a direct comparison between emission intensity of Fig. 2a)–c) for the total area-normalized intensities, indicate that no
spectra from diseased and healthy animals; the second was the use of statistically significant differences could be observed between healthy
supervised machine learning methods. In the following we describe the and tumor samples (analogous trends were found for C-normalized and
results that we obtained with both approaches. background-normalized intensities, not reported here). Serum, on the
other hand, enabled some discrimination between the two classes of
samples. Moreover, as shown in Fig. 3a)–c), serum provided LIBS
4.1. Direct analysis of LIBS spectra spectra with the highest signal-to-noise (S/N) ratio for most samples.
We suspect that this be due to the fact that blood serum has a darker
In a previous work focused on the LIBS analysis of melanoma tissues color than the other fluids, and is therefore more absorptive. This im-
[6], it was shown that the content of Ca, Mg and Na was higher in plies that its interaction with the laser radiation yields a more efficient
melanoma tissues than in healthy skin. To test if biological fluids con- breakdown and brighter plasma. Another possible reason could be that
tain the same elemental signatures of the presence of the disease, we serum is richer in alkaline and alkaline earth metals (and Ca in parti-
measured the intensity of these analytes in the average spectra of each cular) than other fluids. Since serum yielded the most intense spectra,
sample (blood serum and tissue homogenates). In the first series of data, we selected this fluid for investigation of the effect of substrates on the
we deposited serum and the three tissue homogenates on a polymeric signal-to-noise ratio in LIBS spectra, and, consequently, the ability to
substrate, PVDF. The LIBS spectra of the four fluids are shown in Fig. S1 discriminate between healthy and diseased animals. Our aim is to es-
a)-d) provided in the supplementary information section). We chose tablish whether changes in the solid substrate/laser beam coupling
three intense resonance transitions (Na I 589.59 nm, Ca I 422.67 nm, could further improve the signal to noise ratio. A new series of ex-
Mg I 285.21 nm) and normalized them with three methods, i.e. dividing periments were conducted using three different substrates i.e. Cu, Al,
them by: 1) the background intensity in the spectral region adjacent to and Si, all having ionization energy lower than C, H and F, the com-
the peak; 2) the intensity of an element chosen as an internal standard ponents of PVDF. Using substrates with lower ionization energy, and
(C, one of the main elements in biological samples); 3) the total in- thus lower ablation threshold, can allow operation in milder ablation
tegrated area of each average spectrum. In all the figures, we refer to conditions and can yield high-intensity spectra even when working with
samples harvested from healthy mice with the letter N (for Normal), lower laser energy.
and to samples harvested from mice with melanoma with the letter T To compare the three substrates, we estimate the S/N ratio of one C
(for Tumor). In Fig. 1a)–c) we report a comparison of the Ca I, Mg I and transition (247.86 nm) in the average spectra obtained for the three
Na I intensity normalized with the three different methods in the substrates, and reported it in Fig. 4. This shows that the target that
spectra of serum. Fig. 1 shows that the three normalization methods provides the highest S/N ratio is Cu. This may appear contrary to the
lead to similar observations, i.e., that the intensity of Ca I and Mg I is single-shot ablation thresholds reported in Table 1 (obtained in similar
slightly higher in some of the normal samples than in the tumor ones, conditions to those used for this work), therefore we provide a quali-
while no trend is observable for Na I. The Ca and Mg trend is the op- tative interpretation to rationalize this apparent contradiction. Ultra-
posite of that reported in [6] for melanoma lesions and healthy skin, but short laser ablation is often described as a non-thermal phenomenon, as
in the present case the differences were not clear enough to be mean- opposed to the mostly thermal nanosecond-laser ablation. However,
ingful, or to enable a generalization leading to an unambiguous mela- ultrashort ablation is characterized by two ablation regimes (see for
noma diagnosis. Discrepancies between the trace element content and example [12] and references therein, [13]). The first occurs at low
accumulation in fluids (serum and homogenates of tissues not directly

Fig. 1. a)–c): comparison between three different normalization methods for the transitions Ca I 422.67 nm (a), Mg I 285.27 nm (b), and Na I 589.59 nm (c) in the
spectra of blood serum deposited on PVDF. White bars represent samples harvested from healthy mice (N for Normal), black bars represent samples harvested from
mice with melanoma (T for Tumor).

108
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

Fig. 2. a)–c): emission intensities of Ca I 422.67 nm, Mg I 285.27 nm, and Na I 589.59 nm normalized over the total integrated area of spectra of tissue homogenates
deposited on PVDF (a) lungs, b) lymph nodes, c) spleen). White bars represent samples harvested from healthy mice (N for Normal), black bars represent samples
harvested from mice with melanoma (T for Tumor).

Fig. 3. a)–c): signal-to-noise (S/N) ratio of Ca I 422.67 nm a), Na I 589.59 nm b) and Mg I 285.21 nm c) in spectra of serum and homogenates of lungs, lymph nodes
and spleen, deposited on PVDF. The letters N and T respectively indicate samples harvested from healthy mice and from mice with melanoma.

fluences, during which no evidence of thermal effects and melting is

observed, the second at high fluences, during which a plasma forms, the
ablation depth increases and thermal effects appear, even with single
pulse ablation [14]. In LIBS experiments, including ours, the working
ablation regime is the second, since an emitting plasma is formed.
Table 1 shows that Al has lower ablation threshold, but also lower
melting temperature and thermal conductivity than Cu and Si. This
implies that, for this metal, thermal effects, such as the formation and
accumulation of molten material in the laser-induced crater, and the
consequent decrease of ablation efficiency, can play a significant role,
even with ultrashort pulses [14–17]. On the other hand, a comparison
between Cu and Si shows that the formation of a Cu plasma can be
facilitated by its lower ionization potential and higher thermal con-
ductivity, which can reduce the impact of thermal effects despite a
lower melting point. These observations can account for a higher ab-
Fig. 4. Signal-to-noise (S/N) ratio of C I 247.86 nm in spectra of serum de- lation efficiency for Cu than for the other two substrates. The S/N ratio
posited on Cu, Si, and Al. The letters N and T respectively indicate samples of species originating from solutions or fluids deposited on the target
harvested from healthy mice and from mice with melanoma. surface is related to the efficiency of laser/substrate coupling and to the
amount of material ablated from the underlying substrate, which

Table 1
physical properties of Cu, Si and Al. In the ablation threshold column, the acronym SP indicates that the reported values refer to Single Pulse ablation.
Element Ionization energy (eV) [22] Melting point (K) [23] Thermal conductivity at 300 K (W/cm K) [23] Ablation threshold (J/cm2)

Cu 7.726 1358 4.01 0.86 [24] 250 fs, 800 nm, SP

Si 8.152 1687 2.37 0.405 [24] 250 fs, 800 nm, SP
Al 5.985 934 1.48 0.4 [16] 180 fs, 775 nm, SP

109
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

Table 2
Excitation temperature and electron density determined for the spectra of serum deposited on Cu, Al and Si substrates.
Sample Cu series Al series Si series

−3 −3
T (K) Ne × 10 17
(cm ) T (K) Ne × 10 17
(cm ) T (K) Ne × 1017 (cm−3)

Substrate 7830 ± 490 3.9 ± 0.1 8140 ± 290 5.8 ± 0.1

Normal 7750 ± 620 4.2 ± 0.1 9000 ± 690 6.4 ± 0.1 6600 ± 130 4.7 ± 0.1
Normal 8340 ± 380 4.79 ± 0.09 7620 ± 540 5.8 ± 0.1 6700 ± 340 4.6 ± 0.1
Normal 79230 ± 570 4.8 ± 0.1 8080 ± 460 5.0 ± 0.1 6570 ± 360 5.1 ± 0.1
Tumor 8070 ± 580 5.14 ± 0.08 8300 ± 470 5.8 ± 0.1 6350 ± 300 5.3 ± 0.1
Tumor 8930 ± 370 4.85 ± 0.08 8380 ± 550 5.7 ± 0.1 6160 ± 60 4.7 ± 0.1
Tumor 8200 ± 510 5.54931 ± 0.08 8470 ± 440 5.7 ± 0.1 6430 ± 230 4.7 ± 0.2

contributes to form and sustain the plasma [18,19]. Therefore, this can healthy and diseased samples. As an example, in Fig. S2a)–f)
account for the fact that we observe the highest S/N ratio with the (Supplementary information) we report the normalized intensity bar
target that can be ablated with the highest efficiency. Other effects such charts obtained for Ca with the three different substrates.
as those related to the surface conditions, in particular roughness, may We used the Boltzmann distribution and Saha equation to determine
affect the ablation efficiency by locally increasing the beam irradiance the number densities of these analytes, in order to include in our count
and reducing the sample reflectivity ([20] and references therein), as also the ionized fraction of each of the investigated species (all three of
well as the distribution of the deposited fluids [21]. them being highly ionizable: EI Ca = 6.11 eV; EI Mg = 7.65 eV; EI
We did not investigate these effects in details, but we did determine Na = 5.14 eV [22]). It is important to underline that the selected
the plasma parameters of the LIPs produced with the three substrates, transitions are strong resonance ones, so they can be affected by self-
so to characterize the breakdown process in the three cases, as well as absorption, even though the analytes are at trace levels. We report the
to investigate if the plasma parameters could provide additional insight relative number density results in Fig. S3a)–c) (Supplementary in-
to discriminate healthy and diseased subjects. Table 2 shows that no formation), as an example only for the Cu substrate series, which shows
statistically significant differences could be observed between the that even with this correction, the differences between tumor and
plasma parameters of tumor and normal samples, and that the only healthy samples do not appear meaningful. In particular, the Ca density
actual differences could be ascribed to the different substrates. In par- appears higher in tumor samples than in healthy ones, and though this
ticular, the plasma temperature obtained with the Si substrate is lower observation is consistent with the literature about LIBS analysis of
than those with the metal substrates, which can further account for the cancer tissues, and in particular of melanoma, we believe it is not ne-
lower S/N observed for this target. Electron density, on the other hand, cessarily meaningful, as it is not clearly detectable with the other
is highest for Al (the element with the lowest first ionization energy) substrates (in particular, as previously mentioned, we observed an op-
than for the other substrates. In addition, we used the plasma para- posite trend with the PVDF substrate). On the basis of this investigation,
meters to calculate the relative number densities of the three matrix therefore, we conclude that, unlike what has been reported in the lit-
elements, so to qualitatively check our hypothesis of a higher ablation erature about the direct LIBS analysis of cancer tissues, we are not able
efficiency of the Cu substrate. For these additional calculations we used to reliably and unequivocally identify the presence of the disease in
the Boltzmann distribution and the Saha equation, valid for plasmas in samples that are not cells or tissues affected by cancer.
Local Thermodynamic Equilibrium (LTE) [25], in order to determine,
respectively, the relative number density of atoms and of ions of each
species. We report the number densities in Fig. 5. It shows that, with the 4.2. Machine learning approach
sole exception of one sample, the number density of the matrix element
is higher with the Cu substrate than with Al and Si, which is consistent The second part of our investigation was dedicated to establishing if
with our interpretation of the S/N ratio trend. machine learning tools applied to LIBS spectra could discriminate be-
Like in the series of experiments performed on PVDF, the normal- tween healthy and diseased samples despite the fact that the direct
ized emission intensities of Ca, Na, Mg transitions did not provide un- analysis does not yield conclusive indications. We tested four algo-
ambiguous information that can be used to discriminate between rithms, with the intent to compare their performance and identify the
most suitable for the present task: Linear Discriminant Analysis (LDA),
Fisher Discriminant Analysis (FDA), Support Vector Machines (SVM)
and Gradient Boosting.
Linear Discriminant Analysis (LDA) was already discussed in [7]
and will only be briefly recalled here. LDA is a supervised learning
approach that identifies the separating hyperplanes between different
classes by assuming normal class-conditional distribution models.
Features are projected to linear vector subspaces and then classified.
The class of an unknown sample was determined by computing score
values for the various classes using the score functions and data fea-
tures, and the sample was assigned using maximum likelihood decision
rules. Feature extraction was done using the statistical dependency (SD)
between features and associated class labels with a quantized feature
space [26], in order to limit the contribution of non-discriminatory data
points, reduce the dimensionality of the original dataset and avoid over
fitting. FDA is a very similar learning approach to LDA, and is used for
Fig. 5. Relative number density of the matrix element. The transitions em- discrimination between two classes. Support vector machines (SVM) is
ployed were: Cu I 330.80 nm, Al I 266.04 nm, Si 288.15 nm, and the spectro- a discriminative classifier that distinguishes one class from another by
scopic parameters were obtained from [22]. The letters N and T respectively finding an optimal hyperplane that maximizes the separation between
indicate samples harvested from healthy mice and from mice with melanoma. the two classes. The members of both classes that are closest to the

110
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

hyperplane serve as support vectors. The separating hyperplane is se- calculations were run, the first using the spectra as such, the second
lected by optimizing the margin between the two classes. For data removing the main spectral features of the matrix elements, so to
whose feature space is nonlinear, a kernel is used to transform the data evaluate and rule out the possibility of their having a role in the clas-
into a linear space [27,28]. In this work the support vector machine sification. The obtained results were practically identical, which in-
implementation within the Waikato Environment for Knowledge Ana- dicates that the matrix features were not involved in the discrimination
lysis (WEKA) software with a second-degree polynomial kernel was between cancerous and healthy samples, thus in the following we only
used. A ten-fold cross validation was performed [29]. report the one obtained without these features.
For LDA, FDA and SVM, the average spectrum of the clean PVDF The results obtained with the four methods are expressed in terms of
substrate was subtracted from the spectra of the fluids deposited on the the following metrics: total classification accuracy, defined as the per-
substrate itself. No substrate spectrum subtraction was done in the case centage of correctly classified samples from both classes (healthy/
of the Cu, Al, and Si substrates. In both cases, each spectrum with total cancer); Positive Predictive Value (PPV), or sensitivity, defined as the
integrated area that did not fall within one standard deviation about the percentage of samples correctly classified as cancerous; and Negative
average for the 100 shots were rejected, while the remaining were Predictive Value (NPV), or specificity, defined as the percentage of
normalized by their total integrated area. All calculations were per- samples correctly classified as healthy. Our models were trained using
formed using MATLAB [30]. The analysis was performed over the all spectra acquired from all samples, and then tested with cross-vali-
spectral range 250–680 nm to limit the computational cost while at the dation sub-sets of different dimensions. We opted for cross-validation
same time including the spectral region with the most meaningful due to the limited number of available samples, though a better ap-
spectral transitions. proach is an independent external validation, i.e. testing the models
Boosting is a machine learning meta-algorithm that combines a set with data sets not included in the training set (see for example [32,33]
of weak classifiers into a strong classifier. A weak classifier usually has a and references therein).
simple structure and performs only slightly better than random gues-
sing. These weak classifiers are typically trained iteratively and en- 4.2.1. Classification results
sembled in a special way (e.g., weighted according to their individual Fig. 6a)–d) reports the classification accuracy, NPV and PPV ob-
accuracy) to boost the overall performance of the classifier. We used the tained with the first three series of data, i.e. with the four biological
regression tree as the weak classifier in our experiments. The Gradient fluids deposited on PVDF, while the total classification accuracy values
Boosting algorithm considers additive models of the following form: are reported in Table 2 for an immediate comparison.
M
Some considerations can be made based on the data of Fig. 6a)–d)
F (x ) = ∑ αm hm (x ) and Table 3. First, the classification accuracy varied substantially both
m=1 (1) with the analyzed fluid and with the algorithm selected. With the sole
exception of the serum data, the accuracy of LDA, FDA and SVM was
where F(x) is the final model, hm(x) are the weak classifiers and αm are lower than 70% or, in some instances, even only slightly higher than
the weights for each weak classifier determined by its performance. The 50% (a clear indication that in these cases the algorithms failed to
additive model is built in a forward stage-wise fashion: provide any discrimination between the two classes.). On the other
Fm (x ) = Fm − 1 + αm hm (2) hand, Gradient Boosting had a much better performance, with accuracy
of 80% or higher for all the samples. Interestingly, for all the algorithms
At each stage, the model is trying to choose a model that satisfies the the best classification accuracy was obtained with the serum sample,
following equation: which is consistent with the results of the direct analysis of LIBS spectra
y = Fi (x ) + hi (3) and with the fact that the spectra obtained with this fluid provided
spectra with higher S/N than the others (see Fig. 3a)–d)). The Gradient
where y indicates the true classification of sample x, and y − Fi(x) are Boosting data reported in Table 3 and Fig. 6 were obtained using 100
called residuals. These are the parts that existing model is not able to spectral features and a 5-fold cross validation. The number of features
calculate appropriately. In order to compensate for these residuals was selected based on a preliminary optimization, which results we
during each stage, the gradient boost model employs an iterative pro- report in Fig. 7a). Here, classification accuracy obtained with the four
cess for the construction of the additive model. At each iteration, the fluids is plotted as a function of the number of features with 5-fold cross
model attempts to choose a weak classifier that compensates the re- validation. Once the threshold of 100 features is exceeded, the increase
siduals of the existing model. Ultimately, this process minimizes the in classification accuracy is moderate or absent, indicating that in-
overall cost function. creasing further the number of features used for the calculation only
In this work, feature selection and classification were performed increases its computational cost. Therefore, 100 features were selected
with gradient boost models. A simple model with 100 regression trees as the best compromise between computational cost and classification
was used for feature selection. Each regression tree performs feature accuracy.
selection by choosing appropriate split points. Among all the regression For the second series of data, i.e. spectra of serum deposited on Cu,
trees, the more frequently a feature is used in the split points of a tree, Si and Al substrates, we performed an analogous preliminary optimi-
the more important is the feature in the model. The importance of each zation for gradient boost, and plotted the results in Fig. 7b). In this case,
feature (wavelength in the spectrum) in the data was obtained with this the plateau of classification accuracy was reached with 50 features,
approach and we chose the features with higher importance in the therefore in Table 4 and Fig. 8a)–d) we compare the results of LDA, FDA
classification. For classification, we implemented a more complicated and SVM with Gradient Boosting obtained with this optimal number of
gradient boost model with a larger number of regression trees. We features (and 5-fold cross validation). For the Cu substrate, the max-
changed the structure for each regression tree by increasing the imum number of features extracted by the algorithm was 368, while for
minimum number of samples at leaf nodes to simplify each weak the other two substrates, up to 500 features could be extracted and used
classifier, which can help reduce overfitting. We also use subsampling for the calculations.
to enhance the performance of the model [31] as well, i.e. at each Table 4 and Fig. 8 show that, while in the PVDF substrate series,
iteration, the base classifier is trained on a fraction of all training LDA provides at least a modest classification accuracy with serum (and
samples. no classification with the other fluids), in the second series of experi-
To rule out the possibility of mathematical artifacts and to optimize ments this algorithm is unable to classify with all three substrates,
the algorithm itself in terms of its computational cost and accuracy, we which makes it an unsuitable choice for the present task. The other
performed some additional calculations with this method. Two series of algorithms, instead, show a net improvement in the classification

111
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

Fig. 6. a)–d): PPV, NPV and classification accuracy obtained with LDA a), FDA b), SVM c), and Gradient Boosting d) of the four biological fluids deposited on PVDF.

Table 3 Table 4
Comparison of the classification accuracy obtained with four different algo- comparison of the classification accuracy obtained with serum deposited on the
rithms and the spectra of the four biological fluids deposited on PVDF (Gradient four different substrates. The following laser energy values were used: 1.2 mJ
Boosting: 100 features). for Cu, Si, and Al; 1.44 mJ for PVDF (the PVDF data reported here for com-
parison were obtained with the same number of features as the ones calculated
Sample LDA FDA SVM Gradient Boosting
for the other substrates, i.e. 50, and a 5-fold cross-validation).
Serum 73.6 ± 0.6% 76.3 ± 0.6% 78.3 ± 0.2% 92 ± 1% Substrate LDA FDA SVM Gradient Boosting
Lymph node 51.0 ± 0.2% 50.0 ± 0.9% 53.7 ± 0.5% 80 ± 4%
Lungs 57.6 ± 0.9% 68.6 ± 0.6% 69.4 ± 0.3% 85.5 ± 0.8% Cu 53.0 ± 0.5% 85 ± 0.6% 93.8 ± 0.3% 96.3 ± 0.8%
Spleen 66.1 ± 0.7% 68.3 ± 0.7% 65.7 ± 0.3% 86 ± 2% Si 50.4 ± 0.1% 63.0 ± 1.1% 63.3 ± 0.4% 87 ± 4%
Al 55.2 ± 0.7% 66.4 ± 0.9% 83.9 ± 0.5% 87 ± 4%
PVDF 73.6 ± 0.6% 76.3 ± 0.6% 78.2 ± 0.4% 90 ± 2%
accuracy, in particular with the Cu substrate, while Si provides the
worst results, and Al and PVDF results are comparable. Analogously to
what was previously observed (the four different fluids providing dif- matrix and experimental conditions (C is a matrix element in PVDF,
ferent S/N ratios), the fact that the Cu substrates provides the best thus its normalized intensity may not be used to quantitatively compare
classification accuracy can be reasonably related to its high S/N ratio, the S/N ratio with the experiments on metals and Si). As already ob-
reported in Fig. 4. On the other hand, a direct comparison of the S/N served for the previous series, also for the series of serum deposited on
ratio obtained with PVDF is not straightforward due to the different Cu, Si and Al, the best classification accuracy is obtained with Gradient

Fig. 7. a)–b): trend of classification accuracy obtained with gradient boost as a function of the number of features with 5-fold cross-validation for the two series of
experiments: a) serum and tissue homogenates deposited on PVDF; b) serum deposited on Cu, Si, and Al.

112
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

Fig. 8. a)–d): PPV, NPV and classification accuracy obtained with LDA a), FDA b), SVM c), and Gradient Boosting d) of serum deposited on Cu, Si, and Al.

Boosting. two classes, by efficiently processing large sets of data, and to suc-
To optimize this algorithm in terms of finding the best compromise cessfully classify cancer and non-cancer samples. Moreover, our results
between computational cost (and thus speed), accuracy and precision, demonstrate that the choice of the optimal substrate and algorithm can
we made a final test, by changing the dimension of the subset for cross play a crucial role to obtain high classification accuracy.
validation. Fig. 9 displays the classification accuracy obtained with the
optimal number of features selected for the second series of experi-
ments, 50, and three different cross validations, i.e. 10-fold, 5-fold and 5. Conclusions
Leave-One-Out Cross Validation (LOOCV). Fig. 9 shows that the accu-
racy is virtually independent of the dimension of the subset used for In this work we applied LIBS to the diagnosis of a skin cancer,
cross validation, and therefore that using a 5-fold cross validation ra- melanoma, by analyzing four different biological fluids (blood serum
ther than a 10-fold may be a suitable choice to reduce the computa- and three tissue homogenates, lungs, lymph nodes, and spleen) har-
tional cost while still keeping high accuracy (but with a slight loss of vested from diseased mice and healthy controls, and deposited on solid
precision). All the three cross validation methods lead to satisfactory substrates prior to the LIBS analysis. With one of the fluids (serum), we
results (> 80%) with all three substrates, and Cu provides the best, also studied the effect on LIBS spectra and on classification accuracy of
with classification accuracy of 96%. These data prove that, even when changing the laser-sample coupling by using four different substrates
the direct univariate LIBS analysis of biomedical samples with very (PVDF, Cu, Al, and Si). For this analysis, we used two approaches. First,
similar spectral signatures did not provide conclusive results, statistical we used a direct comparison of the emission intensity and relative
methods were able to identify and exploit the variability between the number density of alkaline and earth-alkaline metals in cancerous and
healthy samples. Second, we used a machine learning approach. Unlike
reports in the literature for LIBS of melanoma lesions and adjacent
Cu healthy skin, the direct analysis of spectra of blood and fluids coming
Si from tissues not directly affected by the disease did not yield clear and
100
Al
unambiguous discrimination between biomedical samples from dis-
eased and healthy mice. However, carefully selected machine learning
Classification accuracy %

80
methods could provide high classification accuracy, sensitivity (true
positives), and specificity (true negatives). We compared the results of
60 four different algorithms (Linear Discriminant Analysis, LDA; Fisher
Discriminant Analysis, FDA; Support Vector Machines, SVM; and
40 Gradient Boosting) and found that Gradient Boosting provided the
highest classification accuracy. Moreover, we show that the use of
20 different substrates could significantly affect the classification accuracy,
with Cu providing the best results in the employed experimental con-
ditions. This study shows the powerful capability of machine learning
0
10-fold 5-fold LOOCV tools as compared to traditional or direct methods of analysis of LIBS
spectra for classification purposes. Future work will involve testing the
Fig. 9. Comparison between the classification accuracy obtained with Gradient method with different kinds of asymptomatic tumors at early stages or
Boosting and 10-fold, 5-fold, and LOO cross validation. residual disease after treatment, the use of samples harvested from

113
R. Gaudiuso et al. Spectrochimica Acta Part B 146 (2018) 106–114

humans, and the experimental optimization of the sample deposition Anal. At. Spectrom. 31 (2016) 252.
technique to increase and stabilize the signal-to-noise ratio, which we [12] J. Cheng, C. Liu, S. Shang, D. Liu, W. Perrie, G. Dearden, K. Watkins, Opt. Laser
Technol. 46 (2013) 88.
expect to lead to a further improvement of the classification accuracy. [13] N.N. Nedialkov, S.E. Imamova, P.A. Atanasov, J. Phys. D. Appl. Phys. 37 (2004)
638.
Acknowledgements [14] A.Y. Vorobyev, V.M. Kuzmichev, N.G. Kokody, P. Kohns, J. Dai, C. Guo, Appl. Phys.
A 82 (2006) 357.
[15] B. Le Drogoff, F. Vidal, Y. von Kaenel, M. Chaker, T.W. Johnston, S. Laville, J. Appl.
We would like to thank the Kerith Foundation for its support of this Phys. 89 (2001) 8247.
work. We also thank, the Swim Across America, Ludwig Institute for [16] W. Perrie, M. Gill, G. Robinson, P. Fox, W. O'Neill, Appl. Surf. Sci. 230 (2004) 50.
[17] J. Krüger, P. Meja, M. Autric, W. Kautek, Appl. Surf. Sci. 186 (2002) 374.
Cancer Research, the Parker Institute for Cancer Immunotherapy and [18] A. De Giacomo, C. Koral, G. Valenza, R. Gaudiuso, M. Dell'Aglio, Anal. Chem. 88
the NIH/NCI Cancer Center Support Grant (P30 CA008748) for their (2016) 5251.
support of one of the coauthors (LFC). [19] M.A. Aguirre, S. Legnaioli, F. Almodóvar, M. Hidalgo, V. Palleschi, A. Canals,
Spectrochim. Acta B 88 (2013) 79–80.
[20] L.J. Radziemski, D.A. Cremers (Eds.), Laser-induced Plasmas and Applications,
Appendix A. Supplementary data Marcel Dekker, New York, 1989.
[21] D. Bae, S.-H. Nam, S.-H. Han, J. Yoo, Y. Lee, Spectrochim. Acta B 113 (2015) 70.
Supplementary data to this article can be found online at https:// [22] NIST database, https://www.nist.gov/pml/atomic-spectra-database , Accessed
date: February 2018.
doi.org/10.1016/j.sab.2018.05.010. [23] D.R. Lide, Handbook of Chemistry and Physics, CRC Press, LLC, 2004.
[24] C.S.R. Nathala, A. Ajami, W. Husinsky, B. Farooq, S.I. Kudryashov, A. Daskalova,
References I. Bliznakova, A. Assion, Appl. Phys. A 122 (2016) 107.
[25] J.A.M. Van der Mullen, Phys. Rep. 191 (1990) 109.
[26] J. Pohjalainen, O. Rasanen, S. Kadioglu, Comput. Speech Lang. 29 (2013) 1.
[1] B.W. Stewart, P. Kleihues (Eds.), World Cancer Report 2014, IARC Press, Lyon, [27] T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning: Data
2003. Mining, Inference, and Prediction, Springer, New York, 2009.
[2] R.L. Siegel, K.D. Miller, A. Jemal, CA Cancer J. Clin. 67 (2017) 7. [28] L. Liang, T. Zhang, K. Wang, H. Tang, X. Yang, X. Zhu, Y. Duan, H. Li, Appl. Opt. 53
[3] A. El-Hussein, A.K. Kassem, H. Ismail, M.A. Harith, Talanta 82 (2010) 495. (2014) 544.
[4] F. Ghasemi, P. Parvin, N.S. Hosseini Motlagh, A. Amjadi, S. Abachi, Appl. Opt. 55 [29] E. Frank, M.A. Hall, I.H. Witten, The WEKA Workbench. Online Appendix for “Data
(2016) 8227. Mining: Practical Machine Learning Tools and Techniques”, fourth edition, Morgan
[5] A. Kumar, F. Yueh, J.P. Singh, S. Burgess, Appl. Opt. 43 (2004) 5399. Kaufmann, 2016, p. 2016.
[6] J.H. Han, Y. Moon, J.J. Lee, S. Choi, Y. Kim, S. Jeong, Biomed. Opt. Express 7 [30] MATLAB Version 9.0.0, The MathWorks Inc., Natick, Massachusetts, 2016.
(2015) 57. [31] J. Friedman, Comput. Stat. Data Anal. 38 (2002) 367.
[7] N. Melikechi, Y. Markushin, D.C. Connolly, J. Lasue, E. Ewusi-Annan, [32] S. Moncayo, S. Manzoor, F. Navarro-Villoslada, J.O. Caceres, Chemom. Intell. Lab.
S. Makrogiannis, Spectrochim. Acta B 123 (2016) 33. Syst. 146 (2015) 354.
[8] X. Chen, X. Li, X. Yu, D. Chen, A. Liu, Spectrochim. Acta B 139 (2018) 63. [33] J. Gottfried, D.A. Cremers, L. Radziemski (Eds.), Handbook of Laser Induced
[9] H. Sadozai, T. Gruber, R.E. Hunger, M. Schenk, Front. Immunol. 8 (2017) 1617. Breakdown Spectroscopy, second edition, John Wiley & Sons, Ltd, 2013, pp.
[10] T.H. Erlich, D.E. Fisher, G. Ital. Dermatol. Venereol. 1 (2018) 68. 223–241.
[11] M. Bonta, J.J. Gonzalez, C.D. Quarles Jr., R.E. Russo, B. Hegedus, A. Limbeck, J.

114