Online Prediction of Dopamine Concentration Using

EEG-Induced Type-2 Fuzzy Abduction

1
Madhuleena Dasgupta, 2Amit Konar and 1Atulya K. Nagar
1Dept. of Mathematics and Computer Science, Liverpool Hope University, UK
2
Artificial Intelligence Lab, ETCE Department, Jadavpur University, Kolkata, India
[email protected], [email protected], [email protected]

Abstract— The paper provides a novel approach to online containing Dopamine and were sent to Analytical Chemistry
prediction of Dopamine concentration level in adult humans with laboratory for measurement by chemical means. This is an
glucose as stimulus. It employs Type-2 fuzzy rules with offline and time-consuming method and is not appropriate for
Dopamine concentration level as part of the antecedent and online monitoring of Dopamine in ADHD patients. Among
average power spectral density (acquired from the non-invasive techniques, Magnetic Resonance Imaging
electroencephalography) and subject’s excitement level as parts (MRI) [19] is popular, but it is not affordable outside hospital
of the consequent. Type-2 fuzzy abduction is employed to environment. Recently, researchers developed minimally
transform the rules in a manner to compute Dopamine
invasive means of neurotransmitter measurements, such as
concentration level from the measured value of power spectral
wireless instantaneous neurotransmitter concentration System
density and subject’s excitement level. The prediction results are
comparable with the pre-processed Dopamine concentration
(WINCS) [7], reversed phase HPLC (High Performance
obtained offline. The novelty of the work lies at three distinct Liquid Chromatography) system [12], carbon nano tube multi-
levels. First it is formulated in the settings of an abductive electrode array [4], and the like. Alternative approaches of
reasoning framework. Second, a Type-2 extension of the classical measurement include measuring by Blood Test, and micro-
fuzzy abduction is proposed. Lastly, parameter tuning of the dialysis [15]. Because of excessive cost of the non-invasive
fuzzy sets to obtain optimal results of prediction is also an means, we here attempt to develop one fuzzy logic approach
additional contribution of the work. The principles adopted for capable of predicting Dopamine concentration from the online
online prediction of Dopamine concentration is useful to help measurement of brain signal parameters (such as power
children concentrate on their activities. It is equally useful to spectral density) and subjective excitement level controlled by
people suffering from Schizophrenia and other related brain instantaneous sugar-intake level. We are interested to note
diseases, where control of brain activities by using glucose or how the Dopamine concentration, induced by sugar-intake
other stimuli can be measured online from the measurement of enhances subjective attention level to reduce the effect of
Dopamine release. ADHD syndrome. The fuzzy model to be proposed would
require a few readings of offline Dopamine concentration with
Keywords— Dopamine concentration level prediction, General variation in glucose-intake to construct antecedent-consequent
Type-2 Fuzzy sets, Vertical slice approach, Fuzzy abduction implication relations. However, in real-time prediction phase,
we do not require any direct measurement of Dopamine level
I. INTRODUCTION but would predict its value using the fuzzy model to be
The human brain contains several millions of unicellular nerve proposed.
cells, called neurons, which in groups form nerve fibers to A commonsense thinking reveals that due to glucose
carry nerve impulses from one neuron to the other to perform administration, the Dopamine concentration increases, which
cognitive functionalities like sensing and processing stimuli, induce excitement in our brain. Acquisition of brain signals by
motor actuation, reasoning, learning, and many others. electroencephalography may also give an insight to subject’s
Chemical neurotransmitters play a vital role in the signal excitement level. We use a simple rule to evaluate
transduction process in a nerve fiber. Among the well-known instantaneous Dopamine concentration level from the online
neurotransmitters, Dopamine (DA) is found to have temporal measurement of EEG parameter (here average power spectral
correlations with strong excitement, creativity and attention density) and subject’s excitement level. The subject’s
[2], [17]. In absence of DA, people, especially children are excitement level will be provided by the subject as an integer
detected to have Attention Deficit Hyperactivity Disorder in [0, 99], and would be acquired by brain-computer
(ADHD) [19]. The motivation of this research is to monitor interfacing (BCI) means, so as not to disturb him otherwise to
the DA level of a subject in real-time without invasive means acquire his excitement level. We would use the well-known
through stimulation by chemical means (here using glucose P300 signal and Steady-State Visual Evoked Potential
administration of the subject). Children in the age-group 3-18 (SSVEP) to acquire the subject’s excitement level by asking
years often show the common symptoms of ADHD and need him/her to select the right integer from a 2-dimensional array
special attention by Dopamine-treatment or other means. of integers. The row and column position of the array
Dopamine can be measured through various means and resembling the right answer would be acquired by the above 2
techniques. However, measurement of Dopamine brain signals.
concentration is not easy. Early methods of Dopamine It could have been much easier, if the consequent of the
concentration measurement required extraction of tissues rule includes the Dopamine concentration level and the

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

antecedent would contain the Power Spectral Density (PSD) with experimental details and results where we undertake a
measure and the subject’s excitement level. However, nature comparative study of different fuzzy abduction algorithms in
sometimes does not support our expectations. For instance, the context of prediction of Dopamine levels. Conclusions are
just consider one rule: If Dopamine concentration level is listed in Section VI.
High Then average PSD is High or subject’s excitement level
is High. The rule indicates that average PSD and subjective II. SYSTEM OVERVIEW
excitement level are effects of increased Dopamine level. We
have other rules similar to the above rule with quantifiers High In the present scheme (Fig. 1), we take offline measurements
being replaced by Low, Medium etc. The rule stated above of the Dopamine (DA) concentration before and after every
cannot directly offer us any Dopamine concentration level for trials of stimulus presentation (oral intake of glucose). We
two reasons. First, during forward reasoning [3], the used the DA measurement from its concentration in blood
antecedent is known and consequent is unknown. Here the samples of the subject. We also extract average PSD in the
situation is reverse. Second, we must know the values of the Theta-bands after the subject is stimulated. We also collect the
parameters like average PSD, excitement level and their fuzzy excitation level of the subject due to oral-administration of
quantification like High etc. to assess the Dopamine glucose. This is collected from the subject’s self-assessment in
concentration level. the scale 00-99 and is obtained by using the well-known BCI-
The first problem can be solved by utilizing one interesting Speller Check-board approach using 2 signals P300 and
property of propositional contraposition [3], [13], and its SSVEP. In the proposed Check-board, we write 10 distinct
extension to fuzzy logic [1]. The early work on propositional numbers 0 to 9, 10-19, 20-29, …, 90-99 in each row in
contraposition is due to Wang, cited in [6], who has first ascending order of their occurrence in integer number system,
shown that propositional variables can be transferred after and thus covering 100 numbers in a 10-row, 10-column 2-
complementation from the antecedent to the consequent as we
dimensional array. Suppose the excitation level of the subject
do in Algebra. For example, let, p, q, r and s be four
is 37. Then the subject will fix his attention to the number 37
propositions (i.e., statements having binary truth values), and
we are given a rule: If p and q then r or s. According to Wang, in the board. The BCI system arranges all numbers in a row to
any proposition transposed from antecedent to consequent glow simultaneously one after another, until the right row
(and vice-versa) should be preceded with a negation sign and (here row=4) appears, when the subject is supposed to release
should connect with the existing antecedent by conjunction the P300 signal due to the oddball stimulus. The computer on
(AND operator) or with the consequent by disjunction (OR receiving the P300 recognizes that it identifies the right row
operator). Thus a few correct extensions of the above rule are: containing the subject’s desired 2digit number, indicating his
1) if p and not r then not q or s, 2) If p then not q or r or s, 3) excitement level. Now to recognize the column containing the
If not r then not p or not q or s. In the present context, we desired number, the BCI system will flicker each number in
modify the presumed rule in the following form: If average row 4 at different frequencies. As the subject looks at the
PSD is not High and subject’s excitement level is not High targeted number, the SSVEP will be released at a specific
Then Dopamine level is not High. frequency fixed for that column position. Once the computer
Suppose, we have certain measurements of the above 3 receives the SSVEP, it recognizes the column position
parameters: Dopamine level, excitement level and average containing the subject’s desired number. After, the row and
PSD. Using these parameters, we can construct a fuzzy column positions are recorded, the computer easily grabs the
implication relation for the transformed rule. Next, suppose number and saves it in the right system memory location,
we have certain observations of the antecedent linguistic representative of excitation level.
variables of the rule. We can construct fuzzy sets Next, we go for fuzzy encoding (fuzzification). This is
representation average PSD is not High and excitement level
done for PSD, subject’s excitation level and DA
is not high. Then by using Mamdani type fuzzy reasoning, we
concentration. This is done in two phases. In the first phase,
would be able to infer the membership function of the
subject’s Dopamine level is not High. Once we obtain the we collect 10 data samples of average PSD, DA level and
above inference, we would complement the same and thereby excitation level for three distinct classes of DA concentration
obtain the revised membership of Dopamine concentration (Low: less than 10ng/mL, Medium: around 15ng/mL, and
level being High. We would then defuzzify the inference and High: greater than 20ng/mL) [5]. So, effectively we have 30
would obtain the Dopamine concentration level of the subject. samples of DA concentration along with average PSD and
The idea presented in this paper is simple and straight- excitation levels. To design the fuzzy encoders, we first
forward but useful for online prediction of Dopamine. Such compute the mean mi and standard deviation si of 10 samples
prediction might contain small errors. However, it still has per class, and fit a Gaussian G (mi, si2) to describe the fuzzy
importance as it employs noninvasive online means of MF. However, for the Medium class, we use the complete
prediction of Dopamine concentration level from brain signal Gaussian curve, while for the Low and High classes; we use
and behavioral measurements. half of the Gaussians. To be specific, for the Low class, we
The paper is divided into 6 sections. In Section II, we
provide a system overview of the overall measurement system.
Section III and Section IV outline the novel scheme for fuzzy
abduction and the proposal to use principles of fuzzy abduction
in predicting Dopamine concentration level. Section V deals
 actual DA concentration obtained from blood samples. We use
a mean-square error function to keep it free from sign to
optimize the parameters of the Fuzzy Encoders. Details of
optimization are not given here for lack of space.
Once the parameters of the fuzzy encoders are optimized,
we get an optimized model to predict the DA concentration
from the measurements of unknown excitement level and
unknown average PSD estimates. The online system to be
used for DA concentration prediction is indicated in Fig. 3.

SL: Subject’s Liking

Fig. 1. Offline Training Scheme

take the Gaussian MF for positive values of the linguistic

variables greater than the mean mi. Similarly, for the High
class, we take part of the Gaussian MF with linguistic variable
values less than their mean. This would effectively result in 3 DA concentration level
MFs looking like Fig. 2. Fig. 3. The online Scheme for DA concentration prediction

III. ABDUCTIVE RASONING USING CLASSICAL FUZZY SETS

MF Abduction refers to inferring the premises (causes) from the
observations (effects) of any rule of the form: if <premises>
then <observations>. Although abduction is not supported by
DA Concentration the logic of propositions/predicates, it has importance in many
Fig. 2. The MFs: L, M, H real-world situations. The logic of fuzzy sets offers solution to
the abductive reasoning problems. Several techniques of
abductive reasoning are available in the literature. For
We prepared the MFs from the offline data. However, we example, for a given rule: if x is A then y is B, where x and y
used them online to obtain fuzzy features. This is done by are linguistic variables and A and B are fuzzy sets, given y is
submitting the measured x-values to the right (one of 3) MFs B/, where B/ is an observed MF, we can infer A/ by computing
and its y-values is determined from the specific curve. This is the implication relation R(x, y), and then by computing the
called fuzzy encoding (fuzzification). After we obtain the MFs MF for x is A/ by using max-min composition (o) of B/oR-1(y,
of two (excitement level and average PSD) of the three x), where R-1(y, x) is the inverse fuzzy implication relation,
parameters (excitement level, average PSD and DA such that R-1o R=I, the identity matrix. One early solution to
concentration), we submit them to the abductive reasoning the abductive reasoning problem introduced above is due to
module, which infers the MF of DA concentration in any one Saha and Konar [14]. In this work, the authors used a heuristic
of 3 scales (H/L/M). We next decode (defuzzify) the MF and function to compute approximate R-1. Later in [18], the
thus obtain the defuzzified value of DA concentration. The authors offered a fast and accurate approach to compute R-1.
details of this will be given in Section III. In [1], authors attempted to extend the contraposition property
In the same way we predict DA concentration 30 times and of propositional logic in the context of fuzzy logic to handle
save the results of predicted DA concentration along with the the limitations of the previous researchers. First, the existing
inputs: excitement level, average PSD. We now prepare a set inverse method fails when there exist multiple propositions in
of training instances from these 3 parameters. The number of the antecedent of the rule. Second, computation of R-1 is
instances here is 30 and each instance has 3 parameters. We highly time-consuming. Later in a recent paper, Janarthanana
next use an Evolutionary algorithm [3] to optimize the et al. [16] proposed a novel means of using extended
parameters of the fuzzy encoders. The objective function to be contraposition property, introduced in [1], for ad hoc
optimized is to minimize the prediction error with respect to reasoning. Here, by ad hoc reasoning, the authors mean that
for a given rule with m propositions in the antecedent and n
propositions in the consequent, they can infer the MF of any where c over an expression denotes its one’s complement. The
unknown proposition in the rule presuming the MFs of the statement (1) represents the implication relation: if y is not B
remaining m + n -1 propositions are known. They adopt an and z is not C then x is not A, which is same as R2 (y, z; x).
extension of the contraposition property, which entails an Now, suppose that the MFs for y is B/ and z is C/ are given,
equivalent transformation of a rule by transposing any we need to compute the MF for x is A/.We can use the
proposition from if to then side and vice-versa after changing implication relation R2(y, z; x) along with the given MFs
their signs. Once a proposition is transposed to the other side  B ( y ) and C (z ) to compute  A (x) by the following steps:
of the if-then operator, it behaves like other propositions of the 1. Compute  B  ( y) t C  ( z), where bar above a symbol
destination side. In other words, the then-side propositions are
connected by OR operator and the if-side propositions are represents its complement. For
connected by AND operators. So, after transformation of the example,  B ( y) = 1 −  B ( y).
rule, the transposed propositions would connect with existing 2. Compute the inference
consequent by OR and with existing antecedent by AND  A ( x) = ( B ( y) t C ( z)) o R2 ( y, z ; x).
operators. 3. Re-complement:  A ( x) = 1 −  A  ( x).
In the proposed DA concentration prediction problem, we
frame prediction rules by placing release of DA concentration We rewrite rule Rj in the following form:
level in the antecedent and the response of the brain, here the
average PSD in the Theta band and subject’s excitement level Rj’: If average PSD (avg-PSD) is not High and subject’s
at the consequent. For illustration, let us consider Rule Rj, given excitement level (EL) is not High then Dopamine
by concentration level (DCL) is not High.
Rj: If Dopamine concentration level is High Then average
PSD is High or Subject’s excitement level is High. With the relation introduced above we obtain
 H ( DCL) = R( EL, avg − PSD; DCL) o ( H ( EL) t  H (avg − PSD))
Here, Dopamine concentration level is the desired output.
So, we need to transform the rule so as to transpose Dopamine R( EL, avg − PSD, DCL) = Max{(1 −  H ( EL) t
concentration level at the consequent side and then perform  H (avg − PSD),  H ( DCL))}
forward reasoning. Thus, the problem falls in the domain of by Dienes-Rescher implication function.
abductive reasoning, we would demonstrate that for a given So, we obtain  H (DCL) by complementing  H (DCL).
rule of the form If x is A then y is B or z is C, where A, B and
C are 3 fuzzy sets in respective universes and x, y and z are
IV. GENERAL TYPE-2 FUZZY ABDUCTION FOR DOPAMINE
linguistic variables, we can transform the rule in the following
CONCENTRATION LEVEL PREDICTION
form as the implication relations of the 2 rules using Dienes-
Rescher implication function are identical. Here, An interval Type-2 fuzzy set (IT2FS) is a 2-tuple given by
Let R1 (x; y, z) be the implication relation for the original rule ~
and R2 (y, z; x) be the implication relation of the transformed A = [ ~ ( x),  A~ ( x)]
A
rule. Then,
R1 ( x; y, z) =  A ( x) s ( B ( y) s C ( z)) where  ~ ( x) and  A~ ( x) are called the Lower MF (LMF) and
A

R2 ( y, z; x) = ( B ( y) t C ( z ))c s  A ( x)
Upper MF (UMF) respectively.
where bar over a fuzzy set denotes its complement, and t and s A general Type-2 fuzzy set is defined by
denote t- and s-norms respectively.  x,  A~ ( x),  ( x,  A~ ( x)) 
In the same manner, we can evaluate the implication
relation for the rule: If y is not B and z is not C then x is not A. where, x is the linguistic variable.  A~ ( x) is the primary MF
Let the implication relation for the last rule by Dienes-Rescher
implication function be R2 (y, z; x). and  ( x,  A~ ( x) is the secondary MF. So, a GT2FS is a 3-axis
representation of a fuzzy set. Three different ways of
We would now prove that R1 (x; y, z) = R2(y, z; x). representation of GT2FS are available in the literature. They
Proof. Let  A (x) ,  B ( y ) and  C (z ) be the MFs of x is A, y is B are popularly known as vertical slice, z-slice and wavy slice
representations [10]. We would here use the vertical slice
and z is C, where the parameters have been defined above.
representation in the present paper.
Now, by Dienes-Rescher implication function, we define the
Let us consider a GT2FS rule, given by
implication relation for the rule: If x is A then y is B or z is C
as ~ ~ ~
R1: If x is A then y is B or z is C
R1 ( x; y, z) =  A ( x) s ( B (y) s C (z)) ~ ~ ~
where A, B and C are three General Type 2 fuzzy sets in
=  A ( x) s  B ( y ) s  C ( z )
respective universes X, Y and Z, where the linguistic variables
=  ( y ) s  ( z ) s  A ( x) x  X , y Y and z  Z . For abduction we transform this rule for
B C

= ( B ( y) t C ( z))c s  A ( x) (1) R2, where,

~ ~ ~
= R2 ( y, z; x) R2: If y is B and z is C then x is A .
 .
.
 ( z , v)  ( x, u)
 ( y, u )
y y z x
z x
u1 v1 w1
 ( x, 1 - w)
x
u2 v2 w2

u3 v3 w3
w4
u4 v4
x
u v w
w1 New UMF
w2

w3
New LMF
 ( y, 1 - u )  ( z , 1 - v)  ( x, 1 - w) w4
w5
y z
y x ( 1 - w)
z x
u1
v1 w1
u2
v2 w2
u3
v3 w3
u4
v4 w4
u5
v5 w5

( 1 - u)
( 1 - v) ( 1 - w) Centroid, C

 ( y, 1 − u )  ( z , 1 - v)  ( x, 1 - w)
v1 w1
v2 w2
j j
u2
u1
v3
w4
w3
If C [ C L , C R ] Then Class  j
v4
u3

where j {low, medium, high}

u4

Pi (ui ) = {ui ( y )  Pi (vi ) = {vi ( z )  Pi ( wi ) = {wi ( x) 

.  ( y , 1 − ui ) :  i}  ( z , 1 − vi ) :  i}  ( x, wi ( x)) :  ( x)}

Pi (ui )  Pi (vi ) for  ui ,  vi  {Max Pi ( wi )}  {Min Pi ( wi )}

x wi x wi

Max Min
ui , vi ui , vi
( s norms ) (t norms ) LOW MED HIGH
 ( x, 1 - w)
x LMF
x
w1
w2

w3 UMF
w4

w5 .
( 1 - w)

Fig. 4. Graphical Representation of the Algorithm for GT2FS Reasoning, Centroid Computation and Classification

~ ~ ~ (FOU) [11] at each possible value of the linguistic variables

Let A, B and C be MFs representing CLOSE-TO-CENTRE-
. ~ ~ ~ with peak at the centre of the FOU.
VALUE of the support of A, B and C . We use Gaussian ~
~ ~ ~ Let ui ( y) be a primary MF for the fuzzy set B .
MFs to describe A, B and C , where the Gaussians are
 ( y  , 1 − i ( y )) be the secondary MF of 1 −  i ( y ). Let vi (z ) be
represented by mean and variance of data distributions in X, Y ~
and Z. We use triangular vertical slices to
~ ~ ~ a primary MF for the fuzzy set C .  ( z  , 1 − vi ( z )) be the
represent A, B and C . secondary MF of 1 − vi ( z ). Let wi (x) be a primary MF for the
~
GT-2 MF Construction: We modeled diurnal variations with fuzzy set A .  ( x , 1 − wi ( x)) be the secondary MF of 1 − wi ( x).
10 samples per class per subject in DA concentration level,
average PSD and subject’s excitement level in a day using We follow the steps below and compute the inference and its
Gaussian curves as introduced before. Next, we collect the centroid.
data for 10 days. So, for 10 days, we have 10 Gaussian MFs
for each linguistic variable (DCL, avg-PSD, EL). We take the 1. We take the t-norms and s-norms of
maximum and minimum of the 10 MFs to obtain the UMF and { ( y, 1-u )   ( z, 1-v) : u {u1 , u 2 , u3 .....u n }
LMF of the Type-2 fuzzy set. To construct secondary MFs, we and v {v1 , v2 , v3 ....vn }}.
construct isosceles triangles in the footprint of uncertainty
 If  ( y, 1 − u) has n terms and  ( z, 1 − v) has n terms sugar-intake. So, we keep a small time-slot to ask the subject to
then we have n2 terms. plan his/her excitement-level, which is captured from him/her
by BCI means using P300 and SSVEP during next 8 seconds.
2. We next take maximum of the t-norms and minimum
of the s-norms to produce GLB (Greatest Lower
Sugar
Bound) and LUB (Least Upper Bound) respectively. X
Intake
PSD ST* P300 SSVEP
3. Next we process the consequent GT2FS MF by the
following steps. 2s 6s 8s 2s 4s 4s
a) For a given x = x , compute *
ST: The subject decides about excitement level in [0, 99].
Pi ( x) =  ( x, wi ( x))  wi ( x) for wi ( x)  [wi , wi ]
Fig. 5. The structure of the stimulus
where wi and wi are the minimum and maximum
values of wi in the FOU (Footprint of 00 01 02 03 04 05 06 07 08 09
10 11 12 13 14 15 16 17 18 19
Uncertainty) along the x = x line. Repeat it 20 21 22 23 24 25 26 27 28 29
for x = x ,  x. 30 31 32 33 34 35 36 37 38 39
b) Find the minimum and maximum values of 40 41 42 43 44 45 46 47 48 49
Pi (x) along the x line in the FOU. Let 50 51 52 53 54 55 56 57 58 59
60 61 62 63 64 65 66 67 68 69
[ Pi ( x)]min and [ Pi ( x)]max for all feasible x = x. 70 71 72 73 74 75 76 77 78 79
c) These [ Pi ( x)]min and [ Pi ( x)]max for all 80 81 82 83 84 85 86 87 88 89
90 91 92 93 94 95 96 97 98 99
feasible x = x are joined to obtain new LMF
referred as LM F  and UMF referred as Fig. 6. BCI Speller Check Board
~
UMF  within the FOU of A.
We here briefly outline the feature extraction and function
4. Obtain Min (GLB, LMF ) and Min (LUB, UMF ) and the approximator selection to detect brain signals like SSVEP and
area under the two mins represent here the FOU. P300. It may be remembered that the P300 signal is required
5. In case more than one rules fire we take the union of to determine the row address of the desired number in the
the FOUs produced by each rule and re-declare the check-board. Further, the SSVEP signal is decoded to obtain
the right column of the board containing the desired number.
resulting inferential space as the new FOU. Here, we select Adaptive Autoregressive (AAR) [8]
6. Use Karnik-Mandel (KM) algorithm to obtain the left parameters and Hzorth parameters [8] to recognize P300 and
end point centroid (Cl) and right end point centroid PSD in Theta band to recognize the SSVEP. We use the
(Cr). Finally obtain C= Centroid= (Cl + Cr)/2. General Type-2 Fuzzy function approximator introduced
above to decode the P300 and SSVEP signals.
7. This centroid represents the predicted DA
We here undertake 2 experiments. Fig. 7 proposes a
concentration. scheme for computing the measurement error (ME) in the
8. If the centroid C falls in one of three possible prediction of DA-concentration. In experiment 1, we attempted
centroid ranges for Low, Medium and High classes: to measure prediction accuracy by measuring ME for 10
[LowCl, LowCr], [MedCl, MedCr], [HighCl, HighCr], then sessions of each of 30 subjects. It is apparent that error is found
to lie within a margin of 10%.
the corresponding class is announced as the class of
the measured Dopamine instance.
It is to be noted that for optimal parameter setting we need to
adapt the variance of the primary Gaussian MFs using
Evolutionary Optimization algorithm following the structure
given in Fig. 1. Here we used Differential Evolution (DE)
algorithm to undertake the optimization of parameters of the
fuzzy GT2FS function approximation system.

V. EXPERIMENTS AND RESULTS

We used the stimulus structure shown in Fig. 5 for our
experiment. The stimulus structure begins with a fixation cross
for 2 seconds to draw subject’s attention. Next 6 seconds are
used for presentation of sugar to the subject. The effect of
sugar appears 3 to 4 seconds later on PSD. So, we preserve 8
seconds for extraction of PSD in the Theta band. Now, it is our Fig. 7. Configuration required for Experiment 1
curiosity to know about subject’s excitement level due to
TABLE-I : COMPUTATION OF MEASUREMENT ERROR FOR 10 SESSIONS OF ONE ACKNOWLEDGMENT
SUBJECT
The first author is grateful to Liverpool Hope University for
%ME
Subject Session DAactual DApredicted ME
[ME *100]/DAactual
granting a VC’s scholarship for pursuing a doctoral degree in
1 19.13 20.00 0.9 5.75 Computer Science.
2 18.57 21.00 2.4 7.00
3 10.53 22.00 11.5 9.50 REFERENCES
4 12.80 18.00 5.2 6.25
5 26.67 19.00 7.7 7.50 [1] A. Chakraborty, A. Konar, N R. Pal, L C. Jain, “Extending the
1
6 12.05 20.00 8.0 8.30 Contraposition Property of Propositional Logic for Fuzzy Abduction,”
IEEE Transactions on Fuzzy Systems, vol. 21, August 2013.
7 28.57 19.00 9.6 3.50
8 25.64 21.00 4.6 7.80 [2] A. D. Mosley, D S. Romaine, Ali. Samii, The Encyclopedia of
9 23.53 18.00 5.5 8.50 Parkinson’s Disease, 2nd ed., Facts on File Inc., 2004.
10 11.49 20.00 8.5 8.70 [3] A. Konar, Artificial Intelligence and Soft Computing: Behavioral and
Cognitive Modelling of the Human Brain, CRC Press, 2000.
In experiment 2, we attempted to measure the accuracy of [4] C. A. Marsden, Measurement of neurotransmitter release in vivo,
Chichester: Wiley, 1984.
the proposed GT2FS function approximator. From literature,
[5] D. G. Gardner, D. Shoback, Greenspan’s Basic and Clinical
we obtained the ranges of DA concentration level for low, Endocrinology, 9th ed., McGraw-Hill Companies, 2011, Appendix.
medium and high classes. They are described by [ LowCl, LowCr],
[6] E. A. Dougherty and C R. Giardina, Mathematics for AI and
[MediumCl, MediumCr] and [HighCl, HighCr]. Once the defuzzified Autonomous Systems, Prentice-Hall, Englewood Cliffs, 1988.
output C falls in any one of the three classes, we declare that [7] F. Agnesi, et al., “Wireless Instantaneous Neurotransmitter
class-label←the class boundary including the centroid C. Concentration System based amperometric detection of Dopamine,
Table II provides the prediction accuracy of the proposed Adenosine, and Glutamate for intraoperative neurochemical
function approximator and the existing ones. It is apparent monitoring,” Journal of Neurosurgery, vol. 111(4), October 2009.
from the table that the proposed GT2FS function approximator [8] G. Dornhege, J R. Millan, T. Hinterberger, D J.McFarland, K R. Müller,
outperforms the existing ones by prediction accuracy. “Toward Brain-Computer Interfacing,” MIT Press, 2007.
[9] J. An, Z. Wang, Q. Yang, Z. Ma, “A SVM Function Approximation
Approach with good performances in Interpolation and Extrapolation,”
TABLE- II : PERFORMANCE OF FUNCTION APPROXIMATOR BASED ON Proceedings of the Fourth International Conference on Machine
PREDICTION ACCURACY Learning and Cybernetics, pp. 1648-1653, August 2005.
Function Approximator Prediction Accuracy [10] J. M. Mendel, “General Type-2 Fuzzy Logic Systems Made Simple: A
GT2FS Function Approximator 97.46% Tutorial,” IEEE Transactions on Fuzzy Systems, vol. 22, October 2014.
(Proposed) [11] J. Mendel, D. Wu, “Perceptual Computing: Aiding People in Making
Type-1 Fuzzy Function Approximator 89.20% Subjective Judgements,” Wiley-IEEE Press, April 2010.
(Proposed) [12] L. Studer, “Noninvasive Dopamine determination by reversed phase
Support Vector Machine [9] 90.34% HPLC in the medium of free-floating roller tube cultures of rat fetal
IT2FS Function Approximator 92.28% ventral mesencephalon: A tool to assess dopaminergic tissue prior to
Polynomial Curve fitting 88.86% grafting,” Brain Research Bulletin, vol. 41, pp. 143-150, 1996.
Back-propagation Neural Learning 88.92% [13] N. Nilsson, Artificial Intelligence: A New Synthesis, Morgan Kaufmann
Support Vector Machine with Radial 90.20% Publishers, 1998.
Basis function Kernel
[14] P. Saha, A. Konar, “A heuristic algorithm for computing the mix-min
inverse fuzzy relation,” International Journal of Approximate
VI. CONCLUSIONS Reasoning, vol. 30, pp. 131-147, September 2002.
The paper proposed one novel approach to predict Dopamine [15] R. C. Rayne, “Neurotransmitter Methods,” Humana Press, 1997.
concentration level in human adult subjects using glucose as [16] R. Janarthanan, A. Konar, A. Chakraborty, “Propositional syntax and
the stimulus. General Type-2 fuzzy abduction has been utilized semantics induced knowledge re-structuring in a fuzzy logic network for
to predict the level of Dopamine from the measure of average ad hoc reasoning,” International Journal of Approximate Reasoning, vol.
82, pp. 138-160, 2017.
PSD and subject’s excitement level. Experiments undertaken
[17] S. C. Bhatnagar, Neuroscience: For the Study of Communicative
on 30 subjects confirm that the prediction error with respect to Disorders, 2nd ed., Lippincott Williams and Wilkins, 2002.
measured real values of Dopamine is well within the [18] S. Chakraborty, A. Konar, L C.Jain, “An efficient algorithm to
acceptance limit. The non-invasive means of Dopamine Computing Max-Min Post-Inverse Fuzzy Relation for Abductive
concentration level prediction would have enormous merits, Reasoning,” IEEE Transactions on Systems, Man and cybernetics-Part
particularly in people with ADHD. A Systems and Humans, vol. 40, pp. 158-169, January 2010.
[19] V. Adisetiyo, et al., “Multimodal MR Imaging of Brain Iron in Attention
Deficit Hyperactivity Disorder: A Noninvasive Biomarker That Responds
to Psychostimulant Treatment?” Radiology, vol. 272, pp. 524-532,
August 2014.