See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/332540026

Directional Relationship Between Vitamin D Status and Prediabetes: A New

Approach from Artificial Neural Network in a Cohort of Workers with
Overweight-Obesity

Article  in  Journal of the American College of Nutrition · April 2019

DOI: 10.1080/07315724.2019.1590249

CITATIONS READS

2 161

8 authors, including:

Luisella Maria Vigna Amedea Silvia Tirelli

Fondazione IRCCS Ca' Granda - Ospedale Maggiore Policlinico Fondazione IRCCS Ca' Granda - Ospedale Maggiore Policlinico
113 PUBLICATIONS   499 CITATIONS    111 PUBLICATIONS   1,515 CITATIONS   

SEE PROFILE SEE PROFILE

Enzo Grossi Stefano Turolo

Foundation Villa Santa Maria Fondazione IRCCS Ca' Granda - Ospedale Maggiore Policlinico
551 PUBLICATIONS   7,078 CITATIONS    24 PUBLICATIONS   177 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

European Network for Chemical Elemental Analysis by Total Reflection X-Ray Fluorescence - CA18130 ENFORCE TXRF View project

Comparison of data-driven methods for predicting groundwater redox status View project

All content following this page was uploaded by Massimo Buscema on 20 April 2019.

The user has requested enhancement of the downloaded file.

JOURNAL OF THE AMERICAN COLLEGE OF NUTRITION
https://doi.org/10.1080/07315724.2019.1590249

Directional Relationship Between Vitamin D Status and Prediabetes: A New

Approach from Artificial Neural Network in a Cohort of Workers with
Overweight-Obesity
Luisella Vigna, MDa, Amedea Silvia Tirellib, Enzo Grossic, Stefano Turolod, Laura Tomainoe, Filomena
Napolitanob, Massimo Buscemaf,g, and Luciano Riboldia
a
Department of Preventive Medicine, Occupational Health Unit, Fondazione IRCCS Ca’ Granda, Ospedale Maggiore Policlinico, Milan, Italy;
b
Department of Clinical Chemistry and Microbiology Bacteriology and Virology Units, Ospedale Maggiore Policlinico, Milan, Italy; cVilla Santa
Maria Foundation, Tavernerio, Italy; dPediatric Nephrology & Dialysis, Milano Fondazione IRCCS C
a Grande Ospedale Maggiore Policlinico
University of Milan, Milan, Italy; ePediatric Intermediate Care Unit, Department of Clinical and Community Health Sciences (DISCCO),
Fondazione IRCCS Ospedale C
aGranda-Ospedale Maggiore Policlinico, University of Milan, Milan, Italy; fSemeion Research Centre of Sciences
of Communication, Rome, Italy; gDepartment of Mathematics, University of Colorado, Denver, Colorado, USA

ABSTRACT ARTICLE HISTORY

Objective: Despite the increasing literature on the association of diabetes with inflammation, Received 11 December 2018
cardiovascular risk, and vitamin D (25(OH)D) concentrations, strong evidence on the direction of Accepted 28 February 2019
causality among these factors is still lacking. This gap could be addressed by means of artificial
KEYWORDS
neural networks (ANN) analysis.
Vitamin D; artificial neural
Methods: Retrospective observational study was carried out by means of an innovative data min- network; auto-contractive
ing analysis—known as auto-contractive map (AutoCM)—and semantic mapping followed by map; prediabetes;
Activation and Competition System on data of workers referring to an occupational-health out- inflammation; obesity;
patient clinic. Parameters analyzed included weight, height, waist circumference, body mass index sedentary workers;
(BMI), percentage of fat mass, glucose, insulin, glycated hemoglobin (HbA1c), creatinine, total occupational health
cholesterol, low- and high-density lipoprotein cholesterol, triglycerides, uric acid, fibrinogen,
homocysteine, C-reactive protein (CRP), diastolic and systolic blood pressure, and 25(OH)D.
Results: The study included 309 workers. Of these, 23.6% were overweight or obese with rates of
I, II, and III level obesity, respectively, of 40.5%, 23.3%, and 12.6%. All mean biochemical values
were in normal range, except for total cholesterol, low- and high-density lipoprotein cholesterol,
CRP, and 25(OH)D. HbA1c was between 39 and 46 mmol/mol in 51.78%. 25(OH)D levels were suffi-
cient in only 12.6%. Highest inverse correlation for hyperglycemia onset was with BMI and waist
circumference, suggesting a protective role of 25(OH)D against their increase. AutoCM processing
and the semantic map evidenced direct association of 25(OH)D with high link strength (0.99) to
low CRP levels and low high-density lipoprotein cholesterol levels. Low 25(OH)D led to changes in
glucose, which affected metabolic syndrome biomarkers, first of which was homeostatic model
assessment index and blood glucose, but not 25(OH)D.
Conclusions: The use of ANN suggests a key role of 25(OH)D respect to all considered metabolic
parameters in the development of diabetes and evidences a causation between low 25(OH)D and
high glucose concentrations.

Introduction In relation to the underlying pathophysiological factors

Epidemiological studies have well acknowledged the role of affecting the glucose metabolic pathways, there is increasing
prediabetes as the preparatory asymptomatic condition pre- evidence on the association of diabetes with obesity (2–8),
ceding overt diabetes, estimating a yearly conversion rate cardiovascular risk (9,10), and vitamin D (25(OH)D) con-
between 5% and 10% (depending on population characteris- centrations (11). Yet, the direction of causality among these
tics). Similar to diabetes, prediabetes—which is caused by factors is not totally understood; while there is some evi-
alterations in glucose metabolism—generally manifests dence in literature on the association between low levels of
through the elevation of one or more parameters, mainly 25(OH)D and increased risk for obesity, data on its involve-
impaired fasting glucose (IFG), impaired glucose tolerance ment in inflammatory pathways are few.
(IGT), and glycated hemoglobin (HbA1c). However, current With particular reference to obesity, studies in literature
evidence on the correlation among these parameters is incon- suggest an absorption of vitamin D storage on behalf of adi-
sistent, and contributions of each of these toward defining the pose tissue among subjects with obesity (12). Indeed, serum
pathological state is still the object of some debate (1). 25-vitamin D concentration results from the balance

CONTACT Luisella Vigna, MD luisella.vigna@policlinico.mi.it Department of Preventive Medicine, Occupational Health Unit, Fondazione IRCCS Ca’ Granda,
Ospedale Maggiore Policlinico, Milan, Italy
ß 2019 American College of Nutrition
2 L. VIGNA ET AL.

between vitamin D intake and its release in adipose tissue. inflammatory disease, ongoing treatment for chronic condi-
Therefore, it is believed that the higher presence of body fat tion (including oral contraceptive or hormone replacement
increases the distribution volume, thus decreasing the vita- therapy for postmenopausal women), history or evidence of
min’s bioavailability (13). On the other hand, it appears that malignancy, alcohol abuse, chronic diseases impairing the
low 25(OH)D is also a contributing factor for obesity: the physiological process of production and metabolism of vita-
effects of 1,25–dihydroxyvitamin D (1,25(OH)2D) or calci- min D such as renal failure or hyperparathyroidism, and
triol, the active metabolite of vitamin D, on the absorption finally individuals receiving vitamin D or calcium
of calcium, at the cell level can directly and indirectly mod- supplementations.
ify the balance between lipolysis and lipogenesis inside the
adipocyte (14,15). Moreover, vitamin D status assessed by
serum 25(OH)D levels appears to be inversely associated Parameters and measurements
with hyperglycemia and diabetes, as evidenced by recent epi- Weight, height, waist circumference, body mass index
demiological studies (2,16–18). (BMI), and percentage of fat mass were obtained through
The lack of conclusive evidence on these associations anthropometric and impedance evaluations performed by
could be due to intrinsic methodological limitations of means of the body composition (InBody, Wunder). BMI
standard statistical techniques in describing nonlinear was calculated as the ratio between weight (kg) and height
and complex associations typically observed in (m2) and categorized in classes, according to the standard
biological systems. classification of the World Health Organization: overweight
In recent years, these limits have been overcome by the (BMI between 25.0 and 29.9 kg/m2) and obesity. Obesity was
introduction of artificial neural networks (ANN) analysis further classified into three groups according to its severity
and an innovative data mining analysis known as auto-con- (I class obesity: BMI between 30.0 and 34.9 kg/m2; II class
tractive map (AutoCM) based on ANN architecture: obesity: BMI between 35.0 and 39.9 kg/m2; III class obesity:
AutoCM allows discovery of hidden trends and associations BMI > 40.0 kg/m2). Systolic blood pressure (SBP) and dia-
among variables, using a fuzzy clustering approach (19). The stolic blood pressure (DBP) were measured and the mean of
added value of this approach in investigating biological sys- three measurements taken at 3-minute intervals by means of
tems is represented by its ability in evidencing the organiz- a conventional sphygmomanometer with the patient in
ing principles of a network of variables and therefore to supine position was recorded.
map biological processes using automatic and analytical Blood fasting tests were performed to measure glucose,
models to reconstruct the imprecise, nonlinear, and simul- insulin, HbA1c, creatinine, total cholesterol (T-Chol), low-
taneous pathways underlying a complex set of data. In the density lipoprotein (LDL) and high-density lipoprotein
last decade AutoCM has been successfully tested across the (HDL) cholesterol, triglycerides, uric acid, fibrinogen,
medical field as well (19–29). homocysteine, high-sensitivity C-reactive protein (CRP), and
The aim of the present work was, therefore, to apply this 25(OH)D concentration.
new methodological approach, AutoCM, to our database of Biochemical parameters were assessed by laboratory
a representative sample of Northern Italian working popula- routine methods on Modular P automated analyzer
tion (thus composed of healthy but sedentary individuals (Hitachi-Roche) with the relevant reference intervals or cut-
presenting several degrees of obesity). In particular, the pri- off currently used in our routine laboratory. Quantitative
mary objective of our study was to highlight relevant aspects determination of fibrinogen in citrate plasma samples was
in the development of diabetes and gain information on the obtained on automated I.L. Coagulation System
associations of main biochemical parameters for obesity, (Instrumentation Laboratory S.p.A) and HbA1c was meas-
hyperglycemia, cardiovascular risk, hypertension, inflamma- ured by ion-exchange high-performance liquid chromatog-
tion, and 25(OH)D concentration. raphy on VARIANT II Turbo Instrument (BIORAD Italia).
Vitamin D status was evaluated measuring concentrations of
its circulating form 25(OH)D, using DiaSorin’s 25-OH
Patients and methods
Vitamin D Total competitive chemiluminescence immuno-
The present is a retrospective observational study on data assay on an automated LIASON instrument (Saluggia).
from a database of 825 outpatients (591 female, 234 male; Subjects were classified as 25(OH)D severely deficient
mean age 54 ± 14 years) recruited from at the Obesity and (< 20 ng/ml) or insufficient (20–29 ng/ml) in accordance
Work Center of Fondazione Ca’ Granda, Ospedale Maggiore with the criteria outlined in the 2011 clinical practice guide-
Policlinico of Milano (Italy) between September 2013 and lines of the Endocrine Society (31) and considering 30 ng/ml
March 2015. as the minimum sufficient level. For insulin resistance,
In agreement with the purpose of our study, subjects homeostasis model assessment–insulin resistance (HOMA-
excluded from the study sample were those with serum glu- IR) was calculated as previously described by Matthews
cose above 126 mg/dl (as assessed on two separate evalua- et al. (32) using the following formula: insulin (uU/
tions) in accordance with the American Diabetes mL)
glucose (mmol/L)]/22.5.
Association (ADA) criteria for diagnosis of type 2 diabetes This retrospective observational study was approved
(30). Other exclusion criteria were ongoing diabetes treat- by the local Ethics Committees (Study registration
ment, persistent cardiovascular disease, acute or chronic number: 1370).
 JOURNAL OF THE AMERICAN COLLEGE OF NUTRITION 3

Mathematical and statistical methods been used to elegantly emphasize the causation link between
25(OH)D and glucose concentrations.
Basic statistics
Variables were expressed as mean ± SD ranges. For compari-
sons between groups, independent-samples t test was used. Brief description of ACS
A two-tailed probability value of 0.05 was considered statis- Theory. ACS is an auto-associative neural network, devel-
tically significant. Linear correlation index between variables oped by Buscema at the Semeion Research Center in Rome
was calculated by Pearson test. (34). ACS is an ANN endowed with an uncommon architec-
ture: Any couple of nodes is not linked by a single value,
but by a vector of weights, where each vector component
ANN processing
comes from a specific metric. Such “bio-diversity” of combi-
Main parameters were processed using a new data mining nations of metrics can provide interesting results when each
method based on a particular artificial adaptive system, metric describes different and consistent details of the same
AutoCM, developed by Semeion Research Center Science data set. In this situation, the ACS is an appropriate algo-
Communication, already extensively described elsewhere rithm that forces all the variables to compete among them-
(19,33). In brief, AutoCM is a specific application of ANNs selves, in different respects.
that is particularly effective for highlighting the existence of The ACS algorithm, therefore, is based on the weight
consistent patterns and/or systematic relationship and hid- matrices of other algorithms. ACS will use these matrices as
den trends and associations among variables. The AutoCM a complex set of multiple constraints to update its units in
matrix of connections preserves nonlinear associations response to any input perturbation.
among variables while capturing elusive connection patterns ACS, subsequently, works as a dynamic nonlinear asso-
among clusters that are often overlooked by traditional clus- ciative memory. Whenever any input is set on, ACS will
ter analysis. Relevant associations are plotted in an intuitive activate all its units in a dynamic, competitive, and coopera-
graphical representation, in which variables are represented tive process at the same time. This process ends once the
as nodes and the closeness of nodes accurately reflects the evolutionary negotiation among all the units will find its
association among variables. natural attractor.
The strength of the links is proportional to the weight of The ACS ANN is a complex kind of Content Addressable
the association between the two nodes: pairs of variables Memory system presenting the following new features:
whose connection weights are higher become relatively
closer and vice versa. These “strengths” of links among vari-  The ACS algorithm works using simultaneously many
ables (nodes of the graph) are quantified through a numer- weight matrices, coming from different algorithms and/
ical coefficient ranging from 0 (minimum strength) to 1 or ANNs.
(maximal strength), where the 0–1 interval reflects the range  The ACS algorithm recall is not a one-shot reaction, but
of possible values that the variable can have, thus represent- an evolutionary process where all its units negotiate their
ing the lowest end of the range as 0 and the highest value as reciprocal value.
1. As an example, the variable 25(OH)D has an absolute
natural value range from 4.3 to 83.4. According to the trans- The system is able to employ “simple” and “complex”
formation, the highest value in range in this case 83.4 is algorithms. The former entail applying straightforward for-
attributed the value 1 whereas its lowest value, 4.3, is attrib- mulas for association among variables. The latter make use
uted the value 0. All other intermediate values of the param- in turn of more ANN architectures to compute weights
eter’s natural range are scaled according to this new 1–0 through a sophisticated learning strategy. The mathematical
range. Accordingly, the value 25 becomes 0.41, the value 11 details of the algorithms’ weights definition are described in
becomes 0.18, and so on. The strength of the link can be Appendix 1.
read as the probability of transition from any state-variable ACS processing was used to identify the answers to our
to another. key research questions: (1) the prototypical variables con-
Hence, complement values (binarization) were obtained nected to the subjects with low vitamin D values and (2) the
for each of the 20 variables considered from our database. prototypical variables connected to the subjects with high
Among the 20 parameters considered were 13 biochemical blood glucose values.
indexes (glucose, 25(OH)D, insulin, HbA1c, creatinine, The whole data set was trained using the four types of
T-Chol, LDL and HDL cholesterol, triglycerides, uric acid, algorithms: Pearson linear correlation algorithm, prior prob-
fibrinogen, homocysteine, and CRP) along with age, BMI, ability algorithm, A-Temporal Diffusion Model algorithm,
SBP, DBP, waist circumference (WC), HOMA index, and and Spearman linear correlation algorithm, thus using sim-
percentage of fat mass. This pre-processing scaling allows ultaneously four different weight matrices.
then the proportional comparison among all the variables
and understanding of the existing links of each variable
when the values tend to be high or low. Ethical considerations
Finally, data have been processed using an auto-associa- Data were retrieved from an existing database recording
tive neural network developed by Buscema (34): Activation clinical anamnesis, biochemical parameters, and nutritional
and Competition System (ACS), which in this article has and lifestyle habits. Specific informed consent was collected
4 L. VIGNA ET AL.

as by routine for all patients presenting for the yearly occu- Table 1. Anthropometric, clinical, and biochemical data of population,
expressed as mean values.
pational health consultation.
N ¼ 309 Mean ± SD Min Max
Female n (%)
Results Age
BMI (kg/m2) 33.7 5.3 20 51
Patient characteristics WC (cm) 101 13 40 138
% fat 42 8 18 79
SBP (mm Hg) 128 17 90 200
Of all patients considered in the database, 309 matched the DBP (mm Hg) 80 11 50 110
inclusion/exclusion criteria. Of these (Table 1), 23.6% were Glucose (mg/dl) 92 10 70 124
overweight, 40.5% were classified into the first class of obes- Insulin (mU/ml) 13.7 8.3 2.4 68.2
HbA1c (mmol/mol) 39.1 4.8 22.4 65
ity, 23.3% were in the second class, and 12.6% were in the HOMA-IR 3.23 2.25 0.54 17.13
third class (BMI > 40 kg/m2). All mean biochemical values Total cholesterol (mg/dl) 219 39 141 409
were in normal range, except for T-Chol, HDL cholesterol, LDL cholesterol (mg/dl) 138 34 61 307
HDL cholesterol (mg/dl) 60 16 28 129
LDL cholesterol, CRP, and 25(OH)D. Mean T-Chol concen- Triglycerides (mg/dl) 119 68 31 597
trations were > 200 mg/dl, LDL cholesterol levels were > Creatinine (mg/dl) 0.80 0.19 0.45 2.76
130 mg/dl, and HDL cholesterol levels were < 65 mg/dl. Uric acid (mmol/l) 5.2 1.3 2.6 10
CRP (mg/dl) 0.50 0.81 0.01 9.91
Mean CRP level was 0.5 mg/dL, which is a borderline value Fibrinogen (mg/dl) 338 55 213 602
in healthy subjects (normal value is < 0.5 mg/dl) but high in Homocysteine 11.9 6.3 5.1 55
a setting of cardiovascular risk (where CRP < 0.1 represents 25(OH)D (ng/ml) 19.3 10.3 4.3 83.4
low cardiovascular risk and CRP > 0.3 mg/dl represents CRP ¼ C-reactive protein; DBP ¼ diastolic blood pressure; HDL ¼ high-density
lipoprotein cholesterol; LDL ¼ low-density lipoprotein cholesterol; WC ¼ waist
high cardiovascular risk) (35). As to HbA1c, 51.78% of indi- circumference; SBP ¼ systolic blood pressure.
viduals had HbA1c > 39 mmol/mol within the range
39–46 mmol/mol, which by the diagnosis of diabetes set by Table 2. Vitamin D status in the 309 considered subjects.
Expert Panel on HbA1c is defined as prediabetes (36). 25(OH)D (ng/ml) N %
In reference to 25(OH)D, 12.6% of the cohort had min- < 10 45 14.5
imum threshold levels for adequate intake for their age 10–20 151 48.9
group, while 63.4% had insufficient or deficient values, and < 30 74 23.9
30–40 31 10.0
14.5% were severely deficient (Table 2). > 40 8 2.6
The linear correlation among all parameters considered
for the onset of hyperglycemia and 25(OH)D showed the
highest inverse correlation to be with BMI, followed by WC directly connected to high WC. This last node is the starting
(Figure 1). point toward high HbA1c and high T-Chol, high blood
Patient stratification for normal and high HbA1c (prog- pressure, high glycemia and high insulin, high CRP, and
high homocysteine.
nostic factor of diabetes) revealed that approximately 50% of
Figures 3 and 4 show the results of the application of
the normoglycemic group featured low serum 25(OH)D
ACS in evaluating the causation links between 25(OH)D
concentrations. Among hyperglycemic subjects, 25(OH)D
and glucose concentrations. To approach this task, two sep-
level was lower and BMI was higher compared to normogly-
arate analyses have been carried out. In the first experiment
cemic ones, while CRP did not differ significantly (Table 3).
the question posed to the system was: What happens to the
other variables when 25(OH)D is remarkably low? In the
Semantic map of associations second experiment the question was: What happens to the
other variables when the blood glucose is remarkably high?
The semantic map based on binarization provided a graph- The activation is done putting in the first experiment the
ical representation of connections among variables, as shown variable “low 25(OH)D” at its maximum value (1.0) while
in Figure 2. The variable low 25(OH)D is located at the cen- all the other variables start from values equal to zero, and in
ter of the map like a hub (O), evidencing direct association the second experiment is carried out by putting the variable
with high link strength (0.99) with low CRP levels on the (high blood glucose to its maximum value of 1.0). It is
upper part (A) and low HDL cholesterol levels on the lower worth remembering here that all variables values are scaled
part (B). Moreover, low CRP was connected with low from 0 to 1; therefore, the 1.0 value for “low” 25(OH)D cor-
HOMA index, low triglycerides, low homocysteine, and low responds to the lowest value in the data set (70 mg/dl) and
creatinine (as seen through their distribution, graphically the 1.0 value for “high” blood glucose corresponds to the
resembling the veins of a leaf), each of which appears in highest blood glucose value in the data set (124 mg/dl).
turn to be associated to other variables until reaching high Figure 3 evidences the changes in the values of variables
25(OH)D, low WC, high HDL cholesterol, low age, low gly- after the activation with low 25(OH)D. This activation leads
cemia, low blood pressure, and low insulin. This upper sec- to a change in glucose values, which in turn leads to
tion of the diagram (A) represented the normal/healthy changes in many biomarkers of metabolic syndrome, among
condition. The lower section (B), under the hub of low which increased blood glucose insulin and HOMA index.
25(OH)D, represented the pathologic condition that passes This latter, by the way, is the first variable to respond to the
“in primis” through the low HDL cholesterol level that is activation. Figure 4 shows that the pattern of variables
 JOURNAL OF THE AMERICAN COLLEGE OF NUTRITION 5

Figure 1. Linear correlation among 25(OH)D levels and considered parameters.

p < 0.0001; p < 0.001; p < 0.05.

changes after activation of high blood glucose. While almost Table 3. Relationship among vitamin D, BMI, inflammation, and prediabetes.
all variables related to metabolic syndrome change their HbA1c 25(OH)D BMI CRP
values accordingly, 25(OH)D is not affected. Patients (n ¼ 309) (mmol/mol) (ng/ml) (kg/m2) (mg/dl)
This in our view is the elegant mathematical demonstra- Normal (149) 36.1 ± 2.5 20.2 ± 10 32.2 ± 4.6 0.45 ± 0.89
Prediabetes (160) 43.4 ± 3.7 18 ± 9.2 35.1 ± 5.6 0.56 ± 0.73 ns
tion that the arrows go from low 25(OH)D to high blood
BMI ¼ body mass index; HbA1c ¼ glycated hemoglobin; CRP ¼ C-react-
glucose and not vice versa. ive protein.
HbA1c normal range 20–39 mmol/mol, prediabetes range 39–47 mmol/mol.
Values are expressed as mean ± SD. Statistical differences were evaluated by
Discussion Student’s test.
p < 0.05, p < 0.01.
The present work aimed to gain insight on the complex
associations among serum 25(OH)D concentration, hyper- HbA1c, which was also confirmed in the present study, with
glycemia, and obesity in a population of sedentary Italian strong statistical significance for both BMI and HbA1c. The
workers by means of the new methodological approach of association and direction of the association among HbA1c,
data mining called AutoCM. BMI, and 25(OH)D was further illustrated by the ANN,
Despite the increasing evidence on the association among which evidences the pivotal role 25(OH)D has in the onset
obesity, 25(OH)D serum concentrations, and insulin sensi- of hyperglycemia. This is an interesting result considering
tivity, the direction of the causality relationship still remains the debate around the use of HbA1c in the diagnosis of pre-
unclear. Already in 2006, Bischof et al. (11) had demon- diabetes, either for low sensitivity and high specificity (37)
strated that adiposity was inversely associated to 25(OH)D and low positive predictive value (38). On the other hand,
concentrations, determining a decrease of vitamin D serum HbA1c has been approved as a marker for prediabetes by
levels in subjects with obesity. The study had also suggested the ADA, given its role in representing chronic exposure to
that this inverse association could have been due to an both basal and postprandial hyperglycemia in individuals
increase in metabolic clearance of vitamin D through its over a longer period, and therefore a combination of the
enhanced uptake by fatty tissue, resulting in a decrease of its pathophysiological defects underlying IFG and IGT over
bioavailability. time. As evidenced in the recent review by Sequeira and
In a previous study of ours on Italian workers (8), we Poppit, increased HbA1c values may identify different pools
found a high rate of subjects with excessive weight or obes- of prediabetic individuals including individuals with IFG
ity and with clinical signs of prediabetes featuring low aver- and IGT having different etiologies of prediabetes (39). The
age level of serum 25(OH)D. Analyses had also evidenced cutoffs for HbA1c to date are still under debate, because of
an inverse correlation between 25(OH)D with both BMI and both specificity issues as well as variability due to genetic
6 L. VIGNA ET AL.

Figure 2. Semantic connectivity map of variables created by the AutoCM algorithm.The numbers along the lines (arches) in the graph refer to the strength of the
association between two adjacent nodes. The range of this value is from 0 to 1. This value, deriving from the original weight developed by AutoCM during the train-
ing phase scaled from 0 to 1, is proportional to the strength of the connections between two variables. Here, in addition to the direction of the association (as pro-
vided by standard statistical analyses), we can also appreciate the strength of this association (link strength, [LS]).The upper right quadrant, A, indicates the
physiological condition of healthy individuals, whereas the lower right quadrant, B, indicates pathological condition.
Abbreviations: 25(OH)D ¼ 25 vitamin D; BMI ¼ body mass Index; CRP ¼ C-reactive protein; DBP ¼ diastolic blood pressure; HbA1c ¼ glycated hemoglobin;
HDL ¼ High-density lipoprotein cholesterol; LDL ¼ low-density lipoprotein cholesterol; SBP ¼ systolic blood pressure; T-Chol ¼ total cholesterol; WC ¼ waist
circumference.

factors and ethnicity (differences in red cell turnover or and functional changes in cells leading to the development
hemoglobin glycation that account for variability in HbA1c of microvascular and macrovascular complications. CRP is
measures between populations). one of the most sensitive markers of subclinical inflamma-
Our results through the American College of Cardiology tion, and its serum concentration is an indicator of low-
approach now seem to override the debate and actually grade chronic inflammation of the arterial wall. Here, our
point out the casual relationships among HbA1c, BMI, results confirm an inverse correlation in our population also
and 25(OH)D. for CRP, in agreement with evidence from current literature
The semantic map shows low 25(OH)D to be the discrimin- supporting the association among cardiometabolic risk,
ant factor between normal and pathologic status. The first link hypovitaminosis D, and inflammation (41).
in the pathologic status is to low 25(OH)D and low HDL chol- Taken together, the ACS approach shows low 25(OH)D
esterol which, passing through high WC, links to increased concentration to be a cause of diabetes in subjects with
HbA1c. The distribution of links we have observed for our obesity or overweight rather than the other way around.
population highlights the association between high values for Supporting evidence for the implication of vitamin D in
WC and HbA1c—confirming what is already described in lit- type 2 diabetes comes from a large number of cross-sec-
erature (40)—and the link of visceral fat mass to increased tional studies, which have generally reported an inverse
inflammation (CRP and fibrinogen) in the path leading to pro- association between vitamin D status and prevalent hyper-
gression to a pathological state. This is a clear association glycemia. Serum 25(OH)D concentration is a widely used
among 25(OH)D, inflammation, and metabolic status. marker of vitamin D levels, as circulating 25(OH)D reflects
Our analyses also suggest a protective role of high endogenous synthesis from exposure to sunlight as well as
25(OH)D against increase in WC, BMI, and abdominal fat. dietary intake. Even if there is no consensus for what is con-
By ACS we were able to demonstrate a causation between sidered an appropriate cutoff for optimal 25(OH)D serum
low 25(OH)D and high glucose concentrations. The concentrations, the most frequently used cutoff for vitamin
“normal” status begins with 25(OH)D linked to a low CRP. D deficiency in adults and suggestion of clinical osteomal-
Indeed, chronic hyperglycemia causes irreversible structural acia is serum 25(OH)D < 10 ng/mL.
 JOURNAL OF THE AMERICAN COLLEGE OF NUTRITION 7

Figure 3. The Activation and Competition system (ACS) and low 25(OH)D levels. ACS defined the profile below for the low 25(OH)D concentrations (read from left
to right) was described by high HOMA index, high insulin and glycemia, and high inflammation index.

Clinical intervention studies support that vitamin D, or diabetes, on the other this risk can be reduced by 40% to
its active metabolite 1,25–dihydroxyvitamin D (1,25 70% with adequate lifestyle and dietary changes (43). This
(OH)2D), improves insulin sensitivity (41,42). However, change, however, will difficultly occur if left to individual ini-
there is still need to investigate whether the incidence of tiative and could be more likely achievable through a struc-
diabetes could be reduced by achieving and maintaining suf- tured effort at level of national health care systems. In the
ficient 25(OH)D levels. Some help might be provided by the context of population screening (44), the mandatory yearly
recent article published by Alquist et al. (42) in which occupational checkups already provide an ideal setting and
authors divided diabetic patients into five subgroups with opportunity for large-scale capillary intervention, at no add-
differing disease progression and risk of diabetic complica- itional costs. These yearly appointments could be the occasion
tions. They concluded that this new substratification might for empowering individuals in taking their health into their
eventually help to tailor and target early treatment to own hands and providing them with strategies and thera-
patients who would benefit most, thereby representing a first peutic goals to work toward from one year to the next.
step toward precision medicine in diabetes.
From a more practical point of view, our findings are
Study limitations
extremely important from an epidemiological and prevention
perspective; if on one hand the presence of prediabetes sig- Considering that the ANN approach requires large data-
nificantly increases the chances of individuals experiencing bases, we realize there might be some limits as to sample
8 L. VIGNA ET AL.

Figure 4. The Activation and Competition system (ACS) and high blood glucose concentrations. ACS defined the profile below for the high blood glucose concen-
trations (read from left to right) was describe by high HOMA index, insulin and high inflammation index with no interactions with 25(OH)D concentrations.

size. Moreover, the study population was drawn specifically The use of a specific AutoCM processing and the seman-
within the occupational setting, thus limiting to individuals tic map evidence a clear association among 25(OH)D,
in working age (in Italy, mean age is older than 40 years). inflammation, and metabolic status. In particular, it suggests
Therefore, it does not provide insight on other age groups a protective role of high 25(OH)D against increase in WC,
such as younger adults from 15 years on, which would cer- BMI, and abdominal fat. Moreover, it confirmed causation
tainly be interesting to evaluate. Nonetheless, we hope to between low 25(OH)D and high glucose concentrations.
encourage a common sharing of clinical data of prediabetes Occupational health checkups could be useful in large-scale
status in order to obtain a precise prediction and a better screening and preventive interventions for individuals at risk
algorithm (44). for chronic conditions such as diabetes and inflammatory-
related disease (45).
Conclusion
The use of neural network mapping suggests a key role of
vitamin D respect to all metabolic parameters considered in Acknowledgments
our study in the development of prediabetes, drawing a The authors are grateful to Manuella Walker for editorial assistance in
physiopathologic road map for obesity and type 2 diabetes. preparing the manuscript.
 JOURNAL OF THE AMERICAN COLLEGE OF NUTRITION 9

Conflict of interest statement 18. Reis JP, von M€ uhlen D, Miller ER, 3rd, Michos ED, Appel LJ.
Vitamin D status and cardiometabolic risk factor in the United
The authors have no conflict of interest to disclose. Sates adolescent population. Pediatrics. 2009;3:e371–9. doi:
10.1542/peds.2009-0213.
19. Ulivieri FM, Piodi LP, Grossi E, Rinaudo L, Messina C, Tassi
References AP, Filopanti M, Tirelli A, Sardanelli F. The role of carboxy-ter-
minal crosslinking telopeptide of type I collagen, dual x-ray
1. Bansal N. Prediabetes diagnosis and treatment: a review. World J absorptiometry bone strain and Romberg test in a new osteopor-
Diabetes. 2015;6(2):296–303. ISSN 1948-9358 (online). doi: otic fracture risk evaluation: a proposal from an observational
10.4239/wjd.v6.i2.296. study. PLoS One. 2018;13(1):e0190477. doi:10.1371/journal.
2. Pannu PK, Piers LS, Soares MJ, Zhao Y, Ansari Z. Vitamin D pone.0190477.
status is inversely associated with markers of risk for type 2 dia- 20. Gironi M, Saresella M, Rovaris M, Vaghi M, Nemni R, Clerici
betes: a population based study in Victoria, Australia. PLoS One. M, Grossi E. A novel data mining system points out hidden rela-
2017;12(6):e0178825. doi:10.1371/journal.pone.0178825. tionships between immunological markers in multiple sclerosis.
3. Bobbioni-Harsch E, Pataky Z, Makoundou V, Laville M, Disse E, Immun Ageing. 2013;10(1):1. doi:10.1186/1742-4933-10-1.
Anderwald C, Konrad T, Golay A, Risc I. From metabolic nor- 21. Copped
e F, Grossi E, Buscema M, Migliore L. Application of
mality to cardiometabolic risk factors in subjects with obesity. artificial neural networks to investigate one-carbon metabolism
Obesity. 2012;20(10):2063–9. doi:10.1038/oby.2012.69. in Alzheimer’s disease and healthy matched individuals. PLoS
4. Patel S, Farragher T, Berry J, Bunn D, SIlman A, Symmons D. One. 2013;8(8):e74012. doi:10.1371/journal.pone.0074012.
Association between serum vitamin D metabolite levels and dis- 22. Gallucci M, Zanardo A, Bendini M, Di Paola F, Boldrini P,
ease activity in patients with early inflammatory polyarthritis. Grossi E. Serum folate, homocysteine, brain atrophy, and auto-
Arthritis Rheum. 2007;56(7):2143–9. doi:10.1002/art.22722. CM system: the treviso dementia (TREDEM) study. J Alzheimers
5. Guillot X, Semerano L, Saidenberg-Kermanac’h N, Falgarone G, Dis. 2014;38(3):581–7. doi:10.3233/JAD-130956.
Boissier M-C. Vitamin D and inflammation. Joint Bone Spine. 23. Compare A, Calugi S, Marchesini G, Shonin E, Grossi E,
2010;77(6):552–7. doi:10.1016/j.jbspin.2010.09.018. Molinari E, Grave RD. Emotionally focused group therapy and
6. Mangin M, Shinha R, Fincher K. Inflammation and vitamin D: dietary counseling in binge eating disorder. Effect on eating dis-
the infection connection. Inflamm Res. 2014;63(10):803–19. doi: order psychopathology and quality of life. Appetite. 2013;71:
10.1007/s00011-014-0755-z. 361–8. doi:10.1016/j.appet.2013.09.007.
7. Mousa A, Misso M, Teede H, Scragg R, de Courten B. Effect of 24. Grossi E, Podda GM, Pugliano M, Gabba S, Verri A, Carpani G,
vitamin D supplementation on inflammation: protocol for a sys- Buscema M, Casazza G, Cattaneo M. Prediction of optimal
tematic review. BMJ Open. 2016;6(4):e010804. doi:10.1136/ warfarin maintenance dose using advanced artificial neural
bmjopen-2015-010804. networks. Pharmacogenomics. 2014;15(1):29–37. doi:10.2217/pgs.
8. Vigna L, Cassinelli L, Tirelli AS, Felicetta I, Napolitano F, 13.212.
Tomaino L, Mutti M, Barberi CE, Riboldi L. 25(OH)D levels in 25. Copped
e F, Grossi E, Lopomo A, Spisni R, Buscema M, Migliore
relation to gender, overweight, insulin resistance, and inflamma- L. Application of artificial neural networks to link genetic and
tion in a cross-sectional cohort of Northern Italian workers: evi- environmental factors to DNA methylation in colorectal cancer.
dence in support of preventive health care programs. J Am Coll Epigenomics. 2015;7(2):175–86. doi:10.2217/epi.14.77.
Nutr. 2017;36(4):253–60. doi:10.1080/07315724.2016.1264280. 26. Buscema M, Grossi E, Montanini L, Street ME. Data mining of
9. Skaaby T, Thuesen BH, Linneberg A. Vitamin D, cardiovascular determinants of intrauterine growth retardation revisited using
disease and risk factors. Adv Exp Med Biol. 2017;996:221–30. novel algorithms generating semantic maps and prototypical dis-
doi:10.1007/978-3-319-56017-5_18. criminating variable profiles. PLoS One. 2015;10(7):e0126020.
10. Norman PE, Powell JT. Vitamin D and cardiovascular disease. doi:10.1371/journal.pone.0126020.
Circ Res. 2014;114(2):379–93. doi:10.1161/CIRCRESAHA. 27. Narzisi A, Muratori F, Buscema M, Calderoni S, Grossi E.
113.301241. Outcome predictors in autism spectrum disorders preschoolers
11. Bischof MG1, Heinze G, Vierhapper H. Vitamin D status and its undergoing treatment as usual: insights from an observational
relation to age and body mass Index. Horm Res. 2006;66(5): study using artificial neural networks. Neuropsychiatr Dis Treat.
211–5. doi:10.1159/000094932. 2015;11:1587–99. doi:10.2147/NDT.S81233.
12. Carrelli A, Bucovsky M, Horst R, Cremers S, Zhang C, Bessler 28. Drago L, Toscano M, De Grandi R, Grossi E, Padovani EM,
M, Schrope B, Evanko J, Blanco J, Silverberg SJ, et al. Vitamin D Peroni DG. Microbiota network and mathematic microbe
storage in adipose tissue of obese and normal weight women. J mutualism in colostrum and mature milk collected in two differ-
Bone Miner Res. 2017;32(2):237–424. doi:10.1002/jbmr.2979. ent geographic areas: Italy versus Burundi. ISME J. 2017;11(4):
13. Worstman J, Matsuoka LY, Chen TC, Lu Z, Holick MF. 875–84. doi:10.1038/ismej.2016.183.
Decreased bioavalability of vitamin D in obesity. Am J Clin 29. Morelli V, Palmieri S, Lania A, Tresoldi A, Corbetta S, Cairoli E,
Nutr. 2000;72(3):690–3. doi:10.1093/ajcn/72.3.690. Eller-Vainicher C, Arosio M, Copetti M, Grossi E, et al.
14. Krout D, Schaar A, Sun Y, Sukumaran P, Roemmich JN, Singh Cardiovascular events in patients with mild autonomous cortisol
BB, Claycombe-Larson KJ. The TRPC1 Ca2þ-permeable channel secretion: analysis with artificial neural networks. Eur J
inhibits exercise-induced protection against high-fat diet-induced Endocrinol. 2017;177(1):73–83. doi:10.1530/EJE-17-0047.
obesity and type II diabetes. J Biol Chem. 2017;292(50): 30. American Diabetes Association. Diagnosis and classification of
20799–807. doi:10.1074/jbc.M117.809954. diabetes mellitus. Diabetes Care. 2014;37(Suppl 1):S81–S90. doi:
15. Zemel MB. Proposed role of calcium and dairy food components 10.2337/dc14-S081.
in weight management and metabolic health. Phys Sportsmed. 31. Holick MF, Binkley NC, Bischoff-Ferrari HA, Gordon CM,
2009;37(2):29–39. doi:10.3810/psm.2009.06.1707. Hanley DA, Heaney RP, Murad MH, Weaver CM, Endocrine
16. Rolim MC, Santos BM, Conceiç~ao G, Rocha PN. Relationship Society. Evaluation, treatment, and prevention of vitamin D defi-
between vitamin D status, glycemic control and cardiovascular ciency: an Endocrine Society clinical practice guideline. J Clin
risk factors in Brazilians with type 2 diabetes mellitus. Diabetol Endocrinol Metab. 2011;96(7):1911–30. doi:10.1210/jc.2011-0385.
Metab Syndr. 2016;8(1):77. doi:10.1186/s13098-016-0188-7. 32. Matthews DR, Hosker JP, Rudenski AS, Burnett MA, Darling P,
17. Manickam B, Neagu V, Kukreja SC, Barengolts E. Relationship Bown EG, Turner RC. Homeostasis model assessment: insulin
between glycated hemoglobin and circulating 25-hydroxyvitamin resistance and beta-cell function from fasting plasma glucose and
D concentration in African American and Caucasian American insulin concentrations in man. Diabetologia. 1985;28(7):412–9.
men. Endocr Pract. 2013;19(1):73–80. doi:10.4158/EP12168.OR. doi:10.1007/BF00280883.
10 L. VIGNA ET AL.

33. Buscema M, Grossi E. The semantic connectivity map: an adapt- In this situation, the ACS is an appropriate algorithm that
ing self-organising knowledge discovery method in data bases. forces all the variables to compete among themselves, in dif-
Experience in gastro-oesophageal reflux disease. Int J Data Min
Bioinform. 2008;2(4):362–404.
ferent respects.
34. Buscema M, Newman F, Massini G, Grossi E, Tastle WJ, Liu AK.
The ACS algorithm, therefore, is based on the weight
Assessing post radiotherapy treatment involving brain volume dif- matrices of other algorithms. ACS will use these matrices as
ferences in children: an application of adaptive systems method- a complex set of multiple constraints to update its units in
ology. In: Tastle WJ editor. Data mining application using artificial response to any input perturbation.
adaptive systems. Chapter 1. New York (NY): Springer Science- ACS, subsequently, works as a dynamic non linear asso-
Business Media. 2010. pp 1–23. doi:10.1007/978-1-4614-4223-3_1. ciative memory. Whenever any input is set on, ACS will
35. Osman R, L’Allier PL, Elgharib N, Tardif J-C. Critical appraisal
of C-reactive protein throughout the spectrum of cardiovascular activate all its units in a dynamic, competitive and coopera-
disease. Vasc Health Risk Manage. 2006;2(3):221–37. doi: tive process at the same time. This process will end up
10.2147/vhrm.2006.2.3.221. when the evolutionary negotiation among all the units will
36. International Expert Committee. International expert committee find its natural attractor.
report on the role of the A1C assay in the diagnosis of diabetes. The ACS ANN is a complex kind of Content
Diabetes Care. 2009;32(7):1327–34.
Addressable Memory (C.A.M.) system. Compared to the
37. Guo F, Moellering DR, Garvey WT. Use of HbA1c for diagnoses
of diabetes and prediabetes: comparison with diagnoses based on classic associative memory by Hinton [2], McClelland and
fasting and 2-hr glucose values and effects of gender, race, and Rumelhart [3] and Grossberg [4–5–6], ACS presents the fol-
age. Metab Syndr Relat Disord. 2014;12(5):258–68. doi:10.1089/ lowing new features:
met.2013.0128.
38. Gosmanov AR, Wan J. Low positive predictive value of hemo-  The ACS algorithm works using simultaneously many
globin A1c for diagnosis of prediabetes in clinical practice. Am J
Med Sci. 2014;348(3):191–4. doi:10.1097/MAJ.0000000000000223. weight matrices, coming from different algorithms and/
39. Sequeira I, Poppitt S. HbA1c as a marker of prediabetes: a reli- or ANNs;
able screening tool or not? Insights Nutr Metabol. 2017;1(1):  The ACS algorithm recall is not a one-shot reaction, but
21–9. an evolutionary process where all its units negotiate their
40. Kahn HS, Cheng YJ. Comparison of adiposity indicators associ- reciprocal value;
ated with fasting-state insulinemia, triglyceridemia, and related
risk biomarkers in a nationally representative, adult population.
Diabetes Res Clin Pract. 2018;136:7–15. doi:10.1016/ To compute the weight matrices for the ACS algorithm,
j.diabres.2017.11.019. one can follow different approaches; we will refer to them,
41. Chen X, Wu W, Wang L, Shi Y, Shen F, Gu X, Jia Z. respectively, as ‘simple’ and ‘complex’ algorithms. The for-
Association between 25-hydroxyvitamin D and epicardial adipose mer entail applying straightforward formulas for association
tissue in Chinese non-obese patients with type 2 diabetes. Med among variables. The latter make use in turn of more ANN
Sci Monit. 2017;23:4304–431.
42. Osati S, Homayounfar R, Hajifaraji M. Metabolic effects of vita- architectures to compute weights through a sophisticated
min D supplementation in vitamin D deficient patients (a dou- learning strategy.
ble-blind clinical trial). Diabetes Metab Syndr. 2016;10(2 Suppl
1):S7–S10. doi:10.1016/j.dsx.2016.01.007. 3.2.2. ACS weights: Simple Algorithms
43. Ahlqvist E, Storm P, K€ar€aj€am€aki A, Martinell M, Dorkhan M, The matrix of associations of M variables from a dataset
Carlsson A, Vikman P, Prasad RB, Aly DM, Almgren P, et al.
Novel subgroups of adult-onset diabetes and their association with N patterns can easily be constructed by computing the
with outcomes: a data-driven cluster analysis of six variables. linear associations between any couple of the M variables:
Lancet Diabetes Endocrinol. 2018;6(5):361–9. P
N
44. Tamayo T, Rosenbauer J, Wild SH, Spijkerman AM, Baan C, ðxi;k −x i Þ  ðxj;k −x j Þ
½  k¼1
Forouhi NG, Herder C, Rathmann W. Diabetes in Europe: an Wi;jL ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ;
update. Diabetes Res Clin Pract. 2014;103(2):206–17. doi: P N P N

10.1016/j.diabres.2013.11.007. ðxi;k −x i Þ  2

ðxj;k −x j Þ 2
k¼1 k¼1
45. LeFevre ML, U.S. Preventive Services Task Force. Screening for
vitamin D deficiency in adults: U.S. Preventive Services Task
½ 
Force recommendation statement. Ann Intern Med. 2015;162(2): −1  Wi;jL  1; i; j 2 ½1; 2; :::; M
133–40. doi:10.7326/M14-2450. ½ 
The association matrix, Wi;jL ,
is a square matrix where all
½ 
the main diagonal entries are zero. The matrix Wi;jL has,
APPENDIX 1 however, some limitations. It considers only linear relation-
ships among variables, and it is not sensitive to the fre-
Activation and Competition System quency and to the distribution of the variables across the
3.2.1. Theory dataset. To compensate these limitations, we compute
½ 
ACS is an auto-associative neural network, developed by M. another association matrix, Wi;jP based on the distribution
Buscema in 2009 at the Semeion Research Center in Rome probability of co-occurrence of any couple of the M
[1]. ACS is an ANN endowed with an uncommon architec- variables:
ture: any couple of nodes is not linked by a single value, but P
N P
N
by a vector of weights, where each vector component comes
1
N2  xi;k  ð1−xj;k Þ  ð1−xi;k Þ  xj;k
½  k¼1 k¼1
from a specific metric. Such ‘bio-diversity’ of combinations Wi;jP ¼ − ln
PN P
N
of metrics can provide interesting results when each metric
1
N2  xi;k  xj;k  ð1−xi;k Þ  ð1−xj;k Þ
k¼1 k¼1
describes different and consistent details of the same dataset.
 JOURNAL OF THE AMERICAN COLLEGE OF NUTRITION 11

½  X ½n
−1  Wi;jP  þ1; x 2 ½0; 1; i; j 2 ½1; 2; :::; M ui  Wi;j
k

½  X
If we scale linearly this new matrix, Wi;jP , in the same Ecci ¼ a 
iM
½ 
k
Wi;j >0;
½  Nk;iE
interval as for the linear matrix, Wi;jL we get two comparable kQ
X ½n
hyper-surfaces into the same metric space. ui  Wi;j
k
X iM
Inii ¼ a  ½
k
Wi;j <0;
3.2.3. ACS weights: Complex Algorithms kQ Nk;iI
Ei ¼ Ecci þ b  Inputi Inputi >0;
An alternative route can be that of using more complex Ii ¼ Inii þ b  Inputi Inputi <0;
algorithms, such as ANNs themselves, to compute the   
½n ½n ½n
Neti ¼ Max−ui  Ei þ ui −Min  Ii −Deci  ui −Rest ;
matrix of the weights connecting the dataset variables. This  
di ¼ Neti  1:0−u2i ;
choice yields two important results. First, we can define X
each weight taking into account global interactions among H½n ¼ d2i ;
iM
variables (i.e., the simultaneous associations among all of ½ 
uinþ1 ¼ ui þ di ;
½ n
them), and not simply coupled interactions as in the associ- ½nþ1
Deci
½ n
¼ Deci  e−ui ;
2

ation matrices above. Second, we work with nonlinear speci-

fications of the algorithm, that allow to handle even H ½n is the cost function of ACS to be minimized.
extremely complicated relationships among the data- Subsequently, when, the algorithm terminates.
set variables. More specifically:
In particular, we consider the Auto-Contractive Maps ðMax−ui Þ  Ei þ ðui −MinÞ  Ii −Deci  ðui −RestÞ ¼ 0
(AutoCM) [8, 9].
Once the AutoCM has been trained [8,9,17,18,19,20], we Max  Ei −ui  Ei þ ui  Ii −Min  Ii −Deci  ui þ Rest  Deci ¼ 0
can transform the trained weight matrix, wi;jðnÞ , into a new
metric as follows: ð−Ei þ Ii −Deci Þ  ui þ Max  Ei −Min  Ii þ Rest  Deci ¼ 0
f ðxÞ ¼ the function scales linearly the argument;
Max  Ei −Min  Ii þ Rest  Deci
−1  x  þ1; ui ¼
½  Ei −Ii þ Deci
Wi;jA ¼ new Auto CM weigths matrix:
When Max ¼ 1; Min ¼ −1; Rest ¼ 0:1, then:
½ 
Wi;jA ¼ f ðwi;j Þ:
Ecci þ Ii −0:1  Deci
ui ¼
3.2.4. Activation & Competition System Algorithm Ecci −Ii þ Deci
ACS is a non linear associator, whose cost function is based
on the minimization of the energy among units, whenever
the system is activated by an external input. We call it
We have already said that the ACS ANN is partially inspired
Activation & Competition System (ACS for short). It is
to a previous ANN presented by Grossberg [25 - 267. But
defined as follows:
their differences are so marked that we need to present ACS
M ¼ Number of Variables − Units; as a new ANN:
Q ¼ Number of weights matrices;
i; j 2 M;
k 2 Q;  ACS works using simultaneously many weight matrices
Wi;jk
¼ value of connection between the i−th and coming from different algorithms, while Grossberg’ IAC
the j−th units of the k−th matrix; uses only one weight matrix;
Ecci ¼ global excitation to the i−th unit coming from the other units;  ACS weight matrices represent different mappings of the
Inii ¼ global inhibition to the i−th unit coming from the other units;
Ei ¼ final global excitation to the i−th unit; same dataset and all the units (variables) are processed in
Ii ¼ final global inhibition to the i−th unit; the same way, while Grossberg’ IAC just works when the
½n ¼ cycle of the iteration; dataset presents only a specific kind of architecture;
½n
ui ¼ state of the i−th unit at cycle n;  The ACS algorithm can use any combination of weight
H½n ¼ amount of units updating at cycle n; matrices, coming from any kind of algorithm. The only
Neti ¼ Net Input of the i−th unit;
di ¼ delta update of the i−th unit; constraint is that all the values of every weight matrix
Inputi ¼ value of the i−th external input : −1  Inputi  þ1; have to be linearly scaled into the same range (typically
½ 
Nk;iE ¼ number of positive weights of the k−th matrix to the i−th unit; between -1 and þ1), while Grossberg’ IAC can work
½
Nk;iI ¼ number of negative weights of the k−th matrix to the i−th unit; only with static excitations and inhibitions.
Max ¼ Maximum of activation : Max ¼ 1:0;  Each ACS unit tries to learn its specific value of decay,
Min ¼ Minimum of activaction : Min ¼ −1:0; during its interaction with the other units, while
Rest ¼ rest value : Rest ¼ −0:1;
½n ½ 
Decayi ¼ Decay of activaction the i−th unit at cycle n : Decayin¼0 ¼ 0:1; Grossberg’ IAC works with a static decay parameter for
a ¼ scalar for the Ei and Ii net input to each unit; all the variables;
b ¼ scalar for the external input;  The ACS architecture is a circuit with symmetric weights
e ¼ a small positive quantity close to zero: (vectors of symmetric weights), able to manage a dataset
with any kind of variables (Boolean, categorical,
12 L. VIGNA ET AL.

continuous, etc.), while Grossberg’ IAC can work only 2. Hinton GE, Anderson JA editors (1989) Parallel Models of
with specific types of variables Associative Memory. Psychology press.
3. McClelland JL, Rumelhart DE (1988) Explorations in parallel
distributed processing: A handbook of models, programs, and
Appendix references exercises. MIT Press, Boston, MA.
4. Grossberg S (1976) Adaptive pattern classification and universal
1. Buscema M, Newman F, Massini G, Grossi E, Tastle WJ, and recoding: Part I: Parallel development and coding of neural feature
Liu AK (2013) Assessing Post Radiotherapy Treatment Involving detectors. Biological Cybernetics 23: 121–134. PMID: 974165
Brain Volume Differences in Children: An Application of 5. Grossberg S (1978) A theory of visual coding, memory, and
Adaptive Systems Methodology, Chapter 1, pp 1–23, in W.J. development. In Leeuwenberg ELJ, Buffart HFJM editors. Formal
Tastle (ed.), Data Mining Applications Using Artificial Adaptive Theories of Visual Perception. John Wiley & Sons, New York.
Systems, doi: 10.1007/978-1-4614-4223-3_1, Springer 6. Grossberg S (1980) How does the brain build a cognitive code?
Science þ Business Media New York. Psychol Rev 87: 1–51. PMID: 7375607

View publication stats