Talanta 161 (2016) 31–39

Contents lists available at ScienceDirect

Talanta
journal homepage: www.elsevier.com/locate/talanta

Prediction of fermentation index of cocoa beans (Theobroma cacao L.)

based on color measurement and artificial neural networks
Noemí León-Roque a, Mohamed Abderrahim b, Luis Nuñez-Alejos c, Silvia M. Arribas d,
Luis Condezo-Hoyos b,n
a
Universidad Nacional Pedro Ruiz Gallo, Facultad de Ingeniería Química e Industrias Alimentarias, Departamento de Ingeniería en Industrias Alimentarias,
Lambayeque, Perú
b
Universidad Carlos III de Madrid, Departamento de Ingeniería de Sistemas y Automática, Leganés, Madrid, Spain
c
Programa Nacional de Alimentación Escolar Qali Warma, Cajamarca, Jaén, Perú
d
Universidad Autónoma de Madrid, Facultad de Medicina, Departamento de Fisiología, Madrid, Spain

art ic l e i nf o a b s t r a c t

Article history: Several procedures are currently used to assess fermentation index (FI) of cocoa beans (Theobroma cacao
Received 13 May 2016 L.) for quality control. However, all of them present several drawbacks. The aim of the present work was
Received in revised form to develop and validate a simple image based quantitative procedure, using color measurement and
31 July 2016
artificial neural network (ANNs). ANN models based on color measurements were tested to predict
Accepted 4 August 2016
Available online 5 August 2016
fermentation index (FI) of fermented cocoa beans. The RGB values were measured from surface and
center region of fermented beans in images obtained by camera and desktop scanner. The FI was defined
Keywords: as the ratio of total free amino acids in fermented versus non-fermented samples. The ANN model that
Cocoa included RGB color measurement of fermented cocoa surface and R/G ratio in cocoa bean of alkaline
Fermentation index
extracts was able to predict FI with no statistical difference compared with the experimental values.
Artificial neural networks
Performance of the ANN model was evaluated by the coefficient of determination, Bland-Altman plot and
Image analysis
Passing-Bablok regression analyses. Moreover, in fermented beans, total sugar content and titratable
acidity showed a similar pattern to the total free amino acid predicted through the color based ANN
model. The results of the present work demonstrate that the proposed ANN model can be adopted as a
low-cost and in situ procedure to predict FI in fermented cocoa beans through apps developed for mobile
device.
& 2016 Elsevier B.V. All rights reserved.

1. Introduction Fermentation degree of cocoa beans is traditionally measured

by ammonia (NH3) content [4]. A positive correlation between
There are several procedures currently used to assess fermen- ammonia content and fermentation level has been observed [4].
tation index (FI) of cocoa beans for quality control [1–3]. Cut-test is However, this is assessed by the Conway technique [5], which is
a simple, often used procedure, based on color changes registered time consuming and, therefore, not appropriate for routine ana-
during fermentation [1]. Cut-test consists of cutting beans lysis. Changes in the amount and composition of free amino acids
lengthwise in halves and examining their internal color using a have been reported during fermentation and, therefore, could also
score based on purple and brown beans [2,3]. However, this provide information on FI. Thus, fermented beans are character-
method is not fully quantitative and the color evaluation is very ized by high amounts of total free amino acids and increased ratio
subjective. The measurement the absorbance ratio A460 nm/ of hydrophobic free to acidic free amino acids, which provide in-
formation on the degree of fermentation [6]. A positive correlation
A530 nm of methanolic acid extracts could potentially be used
(r¼ 0.7; p o0.01) between the total amount of free amino acids
[1,3]. This absorbance ratio has shown a good non-linear re-
and the ratio of hydrophobic to acidic free amino acids has been
lationship with sensorial color category (cut-test) and Hunter L, a,
observed in raw cocoa from different origins [6]. Individual and
b color indicators during the fermentation [1], However, it has not
total free amino acids in cocoa samples can be quantified by liquid
been used as FI in fully fermented cocoabeans, where minimal
chromatography. However, this technique is not always available
color differences between samples could occur. in all laboratories and it requires a long time of analysis -ap-
proximately 1 h per sample- [6].
n
Corresponding author. Computer vision-based analytical chemistry (CV-AC), which
E-mail address: [email protected] (L. Condezo-Hoyos). focuses on chemical analysis based on color changes, is gaining

http://dx.doi.org/10.1016/j.talanta.2016.08.022
0039-9140/& 2016 Elsevier B.V. All rights reserved.
32 N. León-Roque et al. / Talanta 161 (2016) 31–39

increasing interest [7]. CV-AC has several significant advantages, box method using boxes of 1.2  0.5  0.5 m provided of two
such as simplicity of use, and the fact that it is easily combinable compartments for progressive fermentation. At 48 h, when the
with portable and widely distributed imaging devices, resulting in fermentation temperature was 40 °C, the samples were stirred
friendly analytical procedures in many areas that demand out-of- every 24 h for 6 days to avoid the excessive increase of the tem-
lab applications for in situ and real-time monitoring [7]. A limited perature (r 48 °C). The fermented cocoa beans were sun-dried to
number of studies using linear and non-linear tools to modeling a final moisture content close to 6.5%, which was measured with a
have been reported on CV-AC in food science. Thus, multiple linear cocoa moisture meter (Aqua-Boy KMP, Leeds, UK). Fermented co-
regression have been developed to predict antioxidant activity coa beans (100–200 g) were sampled from each bag using a spe-
(R2 ¼0.97, p o0.001) and total phenolic compounds (R2 ¼0.94, cific probe and the quartering method. The samples were packed
p o0.001) in colored carrot, based on CIELab data obtained by a in labeled polyethylene bags and stored at 4 °C until analysis.
computer vision system [8]. Similarly, level of ammonia in iceberg
lettuces has been predicted from color parameter (R2 ¼0.73, 2.2.3. Image acquisition, processing and analysis
p o0.001) by multiple linear regression [9]. Least-squares support 2.2.3.1. Acquisition. Forty fermented cocoa beans of each sample
vector machine, a non-linear procedure, and fractal color has been M1, M2 and M3 were randomly chosen and manually peeled off.
used to predict acrylamide content (root mean square error Firstly, whole bean image was acquired in a desktop scanner as
(RMSE) ¼ 15.70 ng/g, coefficient of determination (R2) ¼0.99) [10]. shown below and secondly, each bean was cut manually in halves
Artificial neural networks (ANNs) are very powerful tools to model using a surgical knife, as for the traditional cut-test, and image was
non-linear trends within the data where there are highly complex also acquire. Thereafter, fermented cocoa beans were individually
relationships to be modeled [11]. Cocoa bean fermentation is a ground in a coffee blender (Selecline, France) using five cycles of
complex biochemical process during which several changes occur 10 s to avoid sample heating and stored at 4 °C until analysis. Fifty
in macromolecules such as proteins and carbohydrates as well as milligrams of ground fermented cocoa beans were mixed with
in organic acids and sugars [2]. ANNs is a heuristic tool useful 1 mL of NaOH (0.28 mol L  1) and shaken at maximum speed in a
when theoretical relationship between input and output variables “Movil Rod” (J.P. Selecta S.A, Barcelona, Spain) shaker for 5 min at
is lacking [12] as fermentation index and color of cocoa beans. The room temperature. The fermented cocoa extracts were centrifuged
multiple-layer perceptron is the most used architecture ANNs to at 5000 rpm for 5 min at room temperature and supernatant was
approximate any continuous functions and for high sample/vari- stored at 4 °C until analysis. For fermented cocoa beans, the ima-
able ratio to develop a predictive model [13]. ges were acquired with an HP PSC 1510 desktop scanner (Hewlett-
The aim of this study was to predict FI in fully fermented cocoa Packard Development Company, USA) configured as follows:
beans using RGB color and absorbance, and ANNs. Color-based resolution ¼1200 dpi, color depth¼32 bits per pixel and format
ANN models for the prediction of FI were validated using total free file ¼ TIFF compressed format. A specific white cover was designed
amino acid by coefficient of determination, Bland-Altman plot and and printed in a MakerBot 3D printer (NY, USA) to cover the beans
Passing-Bablok correlation analysis. to avoid external light interference and to improve the light re-
flection [14]. The images of fermented cocoa extracts were ac-
quired using a mobile phone Motorola Moto G under constant
2. Experimental conditions of light on 96-well plate.

2.1. Reagents and solutions 2.2.3.2. Processing and analysis. The images of fermented cocoa
beans for each sample were preprocessed before analysis using
All chemicals used were of analytical-reagent grade with no MATLAB Toolbox Image Processing. The images were corrected for
further purification. Anthrone, D-glucose, gallic acid, o-phtha- the non-uniform background illumination and converted into a
laldehyde (OPA), sodium borate, L-arginine, 2-mercaptoethanol, binary format for analysis. Images of each cocoa bean were seg-
phenol red indicator, polyvinylpyrrolidone (PVPP) and sodium mented from total images (Supplementary Fig. 1). RGB values were
hydroxide were purchased from Sigma-Aldrich (Madrid, Spain). measured using a plugins from ImageJ (http://imagej.nih.gov/ij/).
Absolute ethanol, Folin-Ciocalteu reagent, hydrochloric acid and For cocoa half beans, the mean RGB values of each two halves were
sulfuric acid were acquired from Merck (Barcelona, Spain). Milli Q calculated.
ultrapure grade water ( o18.2 mΩ) was used for the preparation
of solutions. 2.2.4. Chemical measurement
Stock solution of L-arginine (10 mg mL  1) was prepared in 2.2.4.1. Absorption spectrum of extracts from fermented cocoa
0.4 mol L  1 of hydrochloric acid and freshly working solutions 0, beans. Absorption spectra were acquired in a Synergy HT Multi-
0.125, 0.25, 0.5, 1.0 were prepared from stock by serial dilution. Mode Microplate Reader (Biotek, Rochester, VT, USA) at 400–
Glucose stock solution 1 mg mL  1 was prepared in water and it 500 nm.
was used to prepare working solutions 0, 0.0375, 0.075, 0.15,
0.3 mg L  1 by serial dilution. 2.2.4.2. Free amino acids and fermentation index. Free amino acids
from cocoa beans were recovered using an adapted and modified
2.2. Cocoa beans extraction procedure previously described [6,15], which was op-
timized to remove phenolic compounds in fermented cocoa sam-
2.2.1. Samples ples. Fifty milligrams of unshelled and ground cocoa beans were
Two raw white fine cocoa beans of criollo varieties from Piura extracted with 1 mL of aqueous hydrochloric acid solution
(M1) and Cajamarca-Perú (M2) and one native criollo variety (M3) (0.4 mol L  1), containing 5.6% w/v of PVPP to remove phenolic
from Tumbes-Peru were used. Raw samples were harvested in compounds, under constant shaking in a “Movil Rod” for 30 min at
December 2015 and fermented as below. room temperature (20 °C). The samples were centrifuged at
2100  g for 10 min at 4 °C and supernatant was stored at  20 °C
2.2.2. Fermentation and drying until analysis. The efficiency in phenolic compounds removal by
The cocoa fermentation was carried out in three different re- PVPP was checked through comparison with a sample without
gions from Perú (Piura, Cajamarca and Tumbes) according to the PVPP using a rapid microplate high-throughput methodology of
producer usual practices. Raw cocoa beans were fermented by the Folin-Ciocalteu assay [16]. Free amino acids were determined by a
 N. León-Roque et al. / Talanta 161 (2016) 31–39 33

modified fluorimetric assay [17]. Briefly, a 5 mL volume of the network to retrain ¼5. Activation function (identity, logistic, Tanh,
above supernatant was mixed with 100 mL of o-phthalaldehyde exponential and Sine) were tested for hidden and output neurons.
(OPA) reagent in a 96-well plate for 10 min at room temperature in The performance of ANN models was evaluated by comparing the
the darkness. Thereafter, the fluorescence was read in a Synergy values of the coefficient of determination (R2), Bland-Altman plot
HT Multi-Mode microplate reader at excitation wavelength¼ and Passing-Bablok regression analysis.
360 7 40 nm and emission wavelength¼460 740 nm. OPA re-
agent was composed of 0.8 mg mL  1 OPA and 0.2% 2-mercap- 2.2.6. Statistical analyses
toethanol in 0.1 mol L  1 sodium borate [18]. Standard curve of L- ANN models were implemented using Statistica 13.0 (StatSoft,
arginine was used to calculate total free amino acids content and OK, USA). ANOVA One-Way was done with GraphPad Prism
was expressed as mg total free amino acid as arginine/g of sample. 5.0 statistical software (GraphPad, Inc., USA). Bland-Altman plot
Taking into account that non-fermented cocoa beans contain as and Passing-Bablok regression analyses were performed with
low as 2–4 mg/g free fat dry matter of the total free amino acid [6], MedCal Statistical Software 13.3.3 (MedCalc Software bvba, Os-
a mean value of 3 mg/g was used in this study to calculate the tend, Belgium).
novel fermentation index (FIFAa) as follows:
i = n FAa
∑i = 1 FAa fi
FIFAa =
nf 3. Results and discussion
n
where: FA is the amount of free amino acids in fermented cocoa 3.1. Fermentation of cocoa beans and changes in color properties
bean (FAaf i) and non-fermented cocoa bean (FAanf), and n is the
number of evaluated fermented cocoa beans. Fermentation is a key step to prepare fine cocoa through the
formation of potent aroma precursors [2]. Color changes ranging
2.2.4.3. Free sugar, pH and titratable acidity. To determine free su- from purple to brown have been reported during the fermentation
gar content, each fermented cocoa bean (10 mg) was extracted of cocoa beans [1]. Cut-test is a simple visual method to evaluate
with 1 mL of 80% (v/v) ethanol solution on a “movil rod” shaker at the color of the beans, which can be done on the field; therefore, it
maximum speed and room temperature for 2 h. The supernatant, is often used as FI [2,3]. On the other hand there are several che-
obtained by centrifugation at 5000 rpm for 10 min, was mixed mical methods available to assess more accurately the FI. Among
with an equal volume of PVPP solution (0.1 g mL  1) to remove others, a quantitative fermentation index for cocoa beans, based
phenolic compounds and then centrifuged as described above [19]. on absorbance ratio A460nm/A530nm of methanolic acid extracts,
Free sugars were evaluated using a microtiter anthrone-sulfuric has been proposed [1,3]. This fermentation index showed a good
acid colorimetric assay [20,21]. Briefly, 30 mL of supernatant was non-linear relationship with modified sensory color category (cut
mixed with 75 mL anthrone reagent (0.2% w/v in concentrated test) and Hunter L, a, b color indicators of fermentation evolution
sulfuric acid). The mixture was incubated at 95 °C for 3 min and at 0, 12, 36, 60, 84 and 96h [1]. However, absorbance ratio has not
then cooled to 20 °C in a thermocycler PCR (DeltaCycler I™ sys- been tested to compare fermentation index in fully fermented
tem, Ericomp Inc., CA, USA). The reaction mixtures were trans- cocoa beans, where minimal color differences can be found, as
ferred to a 96-well plate and absorbance was measured at 630 nm demonstrated in this study (Fig. 1A). Thus, CIE 1931color space
in a Synergy HT Multi-Mode Microplate Reader. The free sugar showed that color properties of 120 fermented cocoa beans have
content, estimated from glucose standard curve (0–0.3 g L  1), was chromaticity X range between 0.425 and 0.525 (Fig. 1C) and L ¼0–
expressed as grams of glucose per 100 g of sample. 30% and S¼ 60–100% (Fig. 1D); whereas, fermented and non-fer-
To pH and titratable acidity measurement, 50 milligrams of mented cocoa beans (Fig. 1B) were easily discriminated based on
cotyledon of shelled and ground cocoa beans were stirred in color properties for both color spaces (Figs. 1C and D). Because
1.0 mL of water for 30 min at room temperature [22]. The sample color change is a complex biochemical process, nonlinear and
was centrifuged at 5000 rpm for 5 min at room temperature and complex models based on color measurement (using a simple
supernatant was stored at 4 °C for pH and acidity analysis. A pH- image acquisition system) should be implemented to evaluate the
meter (Mettler-Toledo, OH, USA) was used to measure pH at 20 °C. quality of fermented cocoa beans.
Titratable acidity of eachcocoa bean was quantified using a novel
micro-titratable system (Supplementary Fig. 2). Briefly, 100 mL of 3.2. Prediction of fermentation index of cocoa beans based on color
supernatant was titrated with sodium hydroxide (0.1 mol L  1) in and artificial neural networks
presence of mixture indicator (2 mL) -neutral red containing 0.1% of
methylene blue- on 96-well plate. A syringe infusion pump (Har- 3.2.1. Fermentation index based on free amino acids content
vard apparatus, USA) at a flow of 0.1 mL min  1 was used to add From a chemical point of view, some chocolate manufacturers
sodium hydroxide. Titratable acidy was expressed as volume of quantify the NH3 content in order to evaluate the quality control of
sodium hydroxide -calculated from time and flow data- per grams cocoa beans. A positive correlation between NH3 content and
of sample. fermentation level has been observed, as traditionally measured
by ammonia content [4]. However, it is assessed by the Conway
2.2.5. Artificial neural network technique [5], which is time consuming and therefore it is not
A multiple layer perceptron (MLP) architecture of ANNs was appropriate for routine analysis. On the other hand, changes in the
applied for the modeling of total free amino acid of fermented amount and composition of free amino acids are reported during
cocoa beans as function of color parameters. RGB color of fer- fermentation. Thus, fermented seeds are characterized by high
mented cocoa beans and extracts as well as absorption spectrum amounts of total free amino acids and increased ratio of hydro-
of extracts from 400 to 450 nm were used as predictive variables. phobic free to acidic free amino acids, which provide information
AAN models were performed on one-hundred twenty fermented on the degree of fermentation [6]. A positive correlation (r¼ 0.7,
cocoa beans. A random sample size 70:15:15 was used for training, p o0.01) between the amount of total free amino acids and the
test and validation steps using a feed-forward algorithm to mini- ratio of hydrophobic to acidic free amino acids is observed in raw
mize the error function. The number of neurons in the hidden cocoa from different origins [6]. Liquid chromatography procedure
layer was set to a range from 3 to 10, network to train¼20 and is commonly used to quantify total an individual free amino acids
34 N. León-Roque et al. / Talanta 161 (2016) 31–39

Fig. 1. Digital image of fermented and non-fermented cocoa beans acquired with a desktop scanner (A) and a mobile phone (B), respectively. RGB color of surface for
fermented and non-fermented beans represented in CIE 1931 color space (C) and HSL color space (D). M1, M2 and M3 are fine cocoa beans from several Peruvian regions. (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

in cocoa samples, which is not always available in all laboratories free amino acid content (Fig. 2B). In fact, two pH-dependent pro-
and requires long time of analysis -approximately 1 h per sample- teases, an endo-peptidase and a carboxy-exo-peptidase, contained
[6]. In the present study, a very sensitive spectrofluorometric assay into cocoa cotyledons participate in the hydrolysis of globulin
adapted to microplate reader was used to assess total free amino protein storage. The optimum pH reported for carboxy-exo-pep-
acids as chemical fermentation index. High content of total free tidase, which increases hydrophobic amino acids is close to 5.8
amino acid was found for fermented cocoa beans M1 [17]. Organic acids produced during the fermentation of beans can
(11.5 70.3 mg/g sample), M2 (8.8 70.3 mg/g sample) and M3 diffuse into cotyledon, reducing the pH value and modulating the
(9.67 0.2 mg/g sample) (Fig. 2A). It is known that well fermented activity of proteases. Consequently, bean composition -depending
cocoa beans contain about 8–14 mg/g free fat but contents as high of cocoa genotype- interacts with natural microbiota-dependent
as 25 mg can occur in naturally fermented cocoa beans from Ma- fermentation in the production of free amino acids and the for-
dagascar, Papua New Guinea and Java [6] and fermentation-like mation of cocoa flavor quality [24]. In addition, although the pH of
incubation in presence of acetic acid [23]. For M1 cocoa sample, a sample M2 was higher than 5.8 -considered for non-fermented
lower pH value was found, which could explain their higher total samples- [24], M2 showed a total free amino acids content
 N. León-Roque et al. / Talanta 161 (2016) 31–39 35

Fig. 2. Total free amino acid content and pH of fermented fine cocoa beans M1, M2 and M3. Values are means 7 SD (n ¼3). ***p o 0.001.

Fig. 3. Experimental (A) and predicted fermentation index (B) using six artificial neural network models based on color measurement of fermented fine cocoa beans used for
the validation step. *, &p o 0.05, **, &&p o0.01 and ***p o0.001. Asterisk was used to compare M2 (n ¼6) and M3 (n¼ 5) with M1 (n¼ 7), and & symbol reflects the statistical
significance between M2 and M3 samples. M1, M2 and M3 are fine cocoa beans from several Peruvian regions. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

comparable to fermented cocoa samples: 8–14 mg/g free fat spectrophotometric measurement shows a complex relationship
sample [6]. Finally, in the present study a novel fermentation index with TCD [3] in order to be employed as predictor of fermentation
using total free amino acid content measured in a microplate index. Another study on cocoa bean fermentation index and the
reader (see Section 2.2.3.2) was employed to validate artificial Hunter L, a, and b indicators has shown a high regression coeffi-
neural networks models based on color analysis. cient (R2 ¼0.98) and color measurement is suggested as an ap-
propriate predictor to estimate the fermentation index. However,
3.2.2. Artificial neural networks models based on color measurement this color indicator was not able to distinguish between fully fer-
The natural fermentation and fermentation-like incubation of mented cocoa samples [1]. Therefore, simple color measurement is
cocoa beans produce changes in color of whole beans, including not useful to predict fermentation index and does not allow
surface and center regions. Thus, brown pigments may also be comparing fermented cocoa samples at the end of the process.
produced from complexation of condensed tannin, a high mole- Computer vision-based analytical chemistry (CV-AC) focused
cular weight product of flavonoid polymerization with protein, via on chemical analysis and based on color changes is gaining in-
hydrogen bonding, whereas anthocyanins are degraded [1,2]. creasing interest [7]. CV-AC has several significant advantages,
Color change measurements are the basis of common cut test, such as simplicity of use, and the fact that it is easily combinable
which is employed to evaluate samples quality in terms of fer- with portable and widely available imaging devices, resulting in
mentation degree [1–3]. Thus, fermentation index measured as the friendly analytical procedures in many areas that demand out-of-
ratio of A460 nm/A530 nm, has been proposed to evaluate the lab applications for in situ and real-time monitoring [7]. CV-AC
fermentation degree [1]. Changes in total color difference (TCD) might be non-invasive (green analytical procedure) because sam-
values were obtained for raw and fermented cocoa samples at ples do not require any previous chemical treatment [25]. CV-AC is
different times, but fermentation index based on extensively employed as a tool for food classification but a limited
36 N. León-Roque et al. / Talanta 161 (2016) 31–39

Fig. 4. Digital image (A) and representative absorption spectrum of extracts (B) from fermented fine cocoa beans. Extracts were prepared in 0.28 mol L  1 sodium hydroxide,
transferred to a 96-well plate and image was acquired with a mobile phone under natural light. Absorption spectrum was measured in a microplate reader from 400 nm to
500 nm. M1, M2 and M3 are fine cocoa beans from several Peruvian regions.

number of studies are reported in food science. Multivariate classification, according to ammonia content and the prediction of
models obtained by multiple linear regression have been devel- ammonia content in lettuces. In another study, principal compo-
oped to predict antioxidant activity (AA) and total phenolic (TP) in nent analysis (PCA) and least-squares support vector machine
colored carrot from CIELab data obtained by a Computer Vision combined with fractal color were used to classify biscuits ac-
System (CVS). The AA and TP predicted by the proposed models cording to acrylamide level as well as to predict acrylamide con-
showed a good correlation with the experimental AA (R2 ¼0.97, tent (root mean square error (RMSE) ¼15.70 ng/g, coefficient of
p o0.001) and TP (R2 ¼ 0.94, p o0.001) [9]. Similarly, a multi-re- determination (R2) ¼0.99) [10].
gression model that predicts the level of ammonia in iceberg let- In the present study different models based on color mea-
tuces has been developed from color parameter provided by a surement and ANNs were developed in order to predict the level
computer vision system (R2 ¼0.73, p o0.001). However, the au- of fermentation of fine cocoa beans. ANN models were tested
thors suggest the exploration of alternative approaches such as (Section 2.2.5) using different measurements of color, including
artificial neural networks (ANNs) to implement models for the those obtained with a scanner, mobile phone and even absorption
 N. León-Roque et al. / Talanta 161 (2016) 31–39 37

Fig. 5. Predicted and experimental total free amino acids content in fermented fine cocoa beans in training, test and validation steps (A) and for test and validation steps
(B) for ANNs model VI. Color-based ANN model VI was validated by Bland-Altman plot (C) and Passing-Bablok correlation analysis (D). (For interpretation of the references to
color in this figure legend, the reader is referred to the web version of this article.)

Table 1 spectrum of extracts. The simplest ANN model I based only on

Artificial neural network models for the prediction of fermentation index in cocoa color of the surface (Fig. 1) and center (Supplementary Fig. 3) was
beans based on color measurement.
able to detect differences of fermentation index in three fine cocoa
ANN Predictors variable based on Coefficient of determination (R2) bean varieties. These results were comparable to experimental
model color measurement data of sample randomly selected in the validation test of ANN
Training Test Validation (Figs. 3A and B), although a higher statistical difference was de-
tected between M1 and M2 (p o0.01). It was consistent with the
I RGB surface and RGB centera 0.69 0.69 0.60
II RGB surface, RGB center and 0.66 0.60 0.54
fact that experimental FI and predicted values of M1 and M3
RGB NaOH extractb samples did not show statistical differences (p 40.05), whereas
III RGB surface and absorption 0.65 0.69 0.67 the prediction of model I was different to experimental data for M2
spectrum of NaOH extractc sample (p o0.05). To improve the prediction capacity of the ANN
IV RGB surface, RGB center and 0.82 0.81 0.63
model I, an aqueous extract of 120 fermented beans was prepared
absorption spectrum of
NaOH extractd in sodium hydroxide (0.28 mol L  1) for 5 min (Fig. 4). It allowed
V RGB surface and absorbance 0.55 0.66 0.62 the efficient extraction of brown color and soluble proteins, which
at 420 nme are modified during the fermentation process. Moreover, flavo-
VI RGB surface and RG ratio of 0.55 0.76 0.59 noids which are also changed during fermentation are able to
NaOH extractf
react with sodium hydroxide [26]. However, the inclusion of RGB
Absorption spectrum from 400 to 450 nm was used in the ANN models. values of extracts in model II did not improve the prediction of
Multilayer perceptron: Hidden unit ¼ HU; networks to train¼NT; networks to re- fermentation index measured as the total free amino acids. In fact,
tain¼ NR; Activation function for hidden unit (HUF) and output unit (OUF). statistical differences between the predicted and experimental
a
HU ¼ 6, NT¼5, NR¼ 1, HUF ¼ Tanh and OUF¼ Identity. data were higher than those predicted using model I (Figs. 3A and
b
HU¼ 9, NT¼ 9 and NR ¼1, HUF ¼ Tanh and OUF¼ Exponential.
c B). Difference between predicted and experimental fermentation
HU¼ 14, NT¼5 and NR¼ 1, HUF ¼Tanh and OUF ¼Logistic.
d
HU¼ 17, NT¼ 12 and NR¼ 1, HUF ¼ Exponential and OUF¼ Tanh. index for M1 was also found (p o0.05). For models III and IV,
e
HU ¼4, NT¼5 and NR¼ 1, HUF ¼ Tanh and OUF¼ Tanh. spectrum data from 400 to 450 nm together with RGB values of
f
HU¼ 7, NT ¼4 and NR¼ 1, HUF ¼ Logistic and OUF¼ Logistic. cocoa beans were included as predictor variables. Experimental
and predicted fermentation index for all samples were not dif-
ferent but differences were found for predicted values of M2 and
M3 samples (p o0.05) and the highest significance between M1
38 N. León-Roque et al. / Talanta 161 (2016) 31–39

and M2 (p o0.001) (Figs. 3A and B). The ratio A460 nm/A530 nm, a Acknowledgments
quantitative fermentation index for cocoa beans [1,3], was not
included because under extraction conditions residual anthocya- The authors are indebted to University Carlos III of Madrid
nins detected at 530 nm are degraded to colorless compounds (Spain) for the funding received within the strategic action in
[27]. The absorbance ratio (A460 nm/A530 nm) has very limited Robotics, Computer Vision and Automation (Project: 2012/00605/
use for in situ quantification of fermentation index in cocoa beans 002). Dra. León-Roque thanks the Fondo Nacional de Desarrollo
because it requires an organic solvent in the preparation of ex- Científico, Tecnológico y de Innovación Científica (FONDECYT,
tracts and long incubation times of 16–18 h [1,3]. Furthermore, Perú) - Movilización Nacional e Internacional en Ciencia, Tecno-
when absorbance at 420 nm and RGB of the bean surface were logía e Innovación (Grant number 111-2015-FONDECYT-DE) for her
included as predictor variables, the prediction of fermentation Posdoctoral Fellowship. The authors also are indebted to Eng. José
index was similar to those obtained with ANN model I. Apparently, Fernando Reyes Córdova (Cooperativa Agraria Nor Andino Ltda.,
the ANN models that included absorbance values do not improve Piura, Perú), Mr. Lauriano Narcizo Mendoza Herrera (Centro Po-
their prediction capacity of the fermentation index based on total blado Menor Santa Cruz, distrito de Bellavista, Cajamarca, Perú)
free amino acids (Fig. 3B). Interestingly, the inclusion of R/G ratio and Mr. Oscar Velásquez Ramírez (farmer from Canana, Jaén, Ca-
of extracts calculated from images increased the quality of the jamarca, Perú) for providing fermented cocoa samples. Finally, the
prediction for the fermentation index (Model VI). It is probably authors would like to extend their gratitude to Dr. Pilar Rupérez
related to the significant correlation found between R/G ratio of from Instituto de Ciencia y Tecnología de Alimentos y Nutrición
extracts and the total free amino acid in the fermented cocoa (ICTAN), Consejo Superior de Investigaciones Científicas (CSIC),
samples (r ¼0.2414, p o0.01). Similar statistical significance was Madrid, Spain for her critical review of the manuscript.
found for predicted and experimental values of fermentation index
(Figs. 3A and B). Performance model, measured as coefficient of
determination (R2) of 0.76 and 0.59 was reached for test and va- Appendix A. Supporting information
lidation, respectively (Figs. 5A and B and Table 1). Table 1 sum-
marizes all ANN models based on color measurement tested in the Supplementary data associated with this article can be found in
prediction of fermentation index from total free amino acid con- the online version at http://dx.doi.org/10.1016/j.talanta.2016.08.022.
tent. Finally, Bland-Altman plot confirmed that model VI can be
used to predict total free amino acids in fermented cocoa beans
(validation step) and therefore, the fermentation index (Fig. 5C). References
When Passing-Bablok regression analysis was applied, no bias for
the proportionality was found for fluorimetric assay and ANN [1] S.G. Ilangantileke, T. Wahyudi, M.G. Bailon, Assessment methodology to pre-
dict quality of cocoa beans for export, J. Food Qual. 14 (1991) 481–496.
model VI (95% CI of slope ¼ 0.9766–4.6967), which means that [2] W. Amoa-Awua, Methods of cocoa fermentation and drying, in R.F. Schwand
model VI had no significant deviation from linearity. However, the and G.H. Fleet, Cocoa and Coffee Fermentations, First ed., CRC Press - Taylor &
intercept showed bias (95% CI of slope¼  12.7368 to  0.3333) Francis Group, NY, USA, 2015, pp. 71-116.
[3] S. Misnawi, B. Jinap, S. Nazamid Jamilah, Effects of incubation and polyphenol
(Fig. 4B); i.e., a value of correction should be included in Model VI oxidase enrichment on colour, fermentation index, procyanidins and as-
to predict total free amino acids in fermented cocoa beans. Ex- tringency of unfermented and partly fermented cocoa beans, Int. J. Food Sci.
perimental data of total sugar content in mg/g sample Technol. 38 (2003) 285–295.
[4] C. Hue, Z. Gunata, A. Bergounhou, S. Assemat, R. Boulanger, F.X. Sauvage,
(M1 ¼0.75 70.14; n¼7, M2 ¼0.37 70.04; n¼6 and F. Davrieux, Near infrared spectroscopy as a new tool to determine cocoa
M3 ¼ 0.407 0.03; n¼5) and titratable acidity in mL NaOH fermentation levels through ammonia nitrogen quantification, Food Chem.
0.1 mol L  1/g sample (M1 ¼29.3 71.4; n ¼7, M2 ¼ 12.4 7 0.9; n¼ 6 148 (2014) 240–245.
[5] E.J. Conway, A. Byrne, An absorption apparatus for the micro-determination of
and M3 ¼19.0 70.7; n¼ 5) in the three fermented cocoa samples certain volatile substances: the micro-determination of ammonia, Biochem. J.
showed a similar pattern to total free amino acids as fermentation 27 (1933) 419–429.
index and the quality of prediction of ANN model VI. Despite the [6] C. Rohsius, R. Matissek, R. Lieberei, Free amino acid amounts in raw cocoas
from different origins, Eur. Food Res. Technol. 222 (2005) 432–438.
performance of model VI to predict fermentation level of cocoa [7] L.F. Capitán-Vallvey, N. López-Ruiz, A. Martínez-Olmos, M.M. Erenas, A.
beans is lower than a recently developed near infrared (NIR) J. Palma, Recent developments in computer vision-based analytical chemistry:
spectroscopy procedure [4], it is a high-tech technology procedure a tutorial review, Anal. Chim. Acta 899 (2015) 23–56.
[8] B. Pace, M. Cefola, P. Da Pelo, F. Renna, G. Attolico, Non-destructive evaluation
which is not readily available to be used by cocoa manufacturers of quality and ammonia content in whole and fresh-cut lettuce by computer
and producers. Therefore, the proposed procedure based on color vision system, Food Res. Int. 64 (2014) 647–655.
measurement and ANNs could be considered as an alternative [9] B. Pace, M. Cefola, F. Renna, M. Renna, F. Serio, G. Attolico, Multiple regression
models and computer vision systems to predict antioxidant activity and total
low-cost procedure to easily predict the fermentation index in phenols in pigmented carrots, J. Food Eng. 117 (2013) 74–81.
cocoa beans. [10] H. Lu, H. Zheng, Fractal colour: a new approach for evaluation of acrylamide
contents in biscuits, Food Chem. 134 (2012) 2521–2525.
[11] F. Marini, Artificial neural networks in foodstuff analyses: trends and per-
spectives a review, Anal. Chim. Acta 635 (2009) 121–131.
4. Conclusions [12] L. Wang, Support Vector Machine: Theory and Application, First ed., Springer-
Verlag, Berlin, Germany, 2005.
[13] D.J. Livingstone, Artificial Neural Networks: Methods and Applications, First
A quantitative and simple model based on color measurement ed., Humana Press, NY, USA, 2008.
from digital images and artificial neural networks useful to predict [14] M. Abderrahim, S.M. Arribas, L. Condezo-Hoyos, A novel high-throughput
fermentation index as total free amino acids for fine fermented image based rapid Folin-Ciocalteu assay for assessment of reducing capacity in
foods, Talanta 152 (2016) 82–89.
cocoa beans was developed and validated. The developed vision- [15] A. Marseglia, G. Palla, A. Caligiani, Presence and variation of γ-aminobutyric
based analytical assay included RGB values of beans surface and acid and other free amino acids in cocoa beans from different geographical
extracts as predictor variables. It could be applicable by manu- origins, Food Res. Int. 63 (Part C) (2014) 360–366.
[16] L.M. Magalhães, F. Santos, M.A. Segundo, S. Reis, J.L.F.C. Lima, Rapid microplate
facturers and producers for an in situ quality control and quanti- high-throughput methodology for assessment of Folin-Ciocalteu reducing
fication of the fermentation level of processed cocoa beans. Finally, capacity, Talanta 83 (2010) 441–447.
it is worth mentioning that image acquisition, processing and [17] G.H. Fisher, I. Arias, I. Quesada, S. D'Aniello, F. Errico, M.M. Di Fiore,
A. D'Aniello, A fast and sensitive method for measuring picomole levels of
analysis can be further developed and used for the implementa- total free amino acids in very small amounts of biological tissues, Amino Acids
tion of analytical tools as apps in mobile devices. 20 (2001) 163–173.
 N. León-Roque et al. / Talanta 161 (2016) 31–39 39

[18] F.C. Church, D.H. Porter, G.L. Catignani, H.E. Swaisgood, An o-phthalaldehyde [23] D. Kadow, N. Niemenak, S. Rohn, R. Lieberei, Fermentation-like incubation of
spectrophotometric assay for proteinases, Anal. Biochem. 146 (1985) 343–348. cocoa seeds (Theobroma cacao L.) – reconstruction and guidance of the fer-
[19] B. Tian, B. Xie, J. Shi, J. Wu, Y. Cai, T. Xu, S. Xue, Q. Deng, Physicochemical mentation process, LWT - Food Sci. Technol. 62 (2015) 357–361.
changes of oat seeds during germination, Food Chem. 119 (2010) 1195–1200. [24] E.O. Afoakwa, A. Paterson, M. Fowler, A. Ryan, Flavor formation and character
[20] A. Laurentin, C.A. Edwards, A microtiter modification of the anthrone-sulfuric in cocoa and chocolate: a critical review, Crit. Rev. Food Sci. Nutr. 48 (2008)
acid colorimetric assay for glucose-based carbohydrates, Anal. Biochem. 315 840–857.
(2003) 143–145. [25] S. Garrigues, M. de la Guardia, Non-invasive analysis of solid samples, TrAC
[21] F. Abderrahim, E. Huanatico, R. Repo-Carrasco-Valencia, S.M. Arribas, M. Trends Anal. Chem. 43 (2013) 161–173.
C. Gonzalez, L. Condezo-Hoyos, Effect of germination on total phenolic com- [26] E.W. Trease GE, Pharmacognosy, 15 ed., Saunders Publishers, London, UK,
pounds, total antioxidant capacity, Maillard reaction products and oxidative 2002.
stress markers in canihua (Chenopodium pallidicaule), J. Cereal Sci. 56 (2012) [27] A. Castañeda-Ovando, Md.L. Pacheco-Hernández, M.E. Páez-Hernández, J.
410–417.
A. Rodríguez, C.A. Galán-Vidal, Chemical studies of anthocyanins: a review,
[22] M. Senanayake, E.R. Jansz, K.A. Buckle, Effect of different mixing intervals on
Food Chem. 113 (2009) 859–871.
the fermentation of cocoa beans, J. Sci. Food Agric. 74 (1997) 42–48.