Contents lists available at ScienceDirect [632022]

Contents lists available at ScienceDirect
Archives of Gerontology and Geriatrics
journal homepage: www.elsevier.com/locate/archger
Identifying combinatorial biomarkers by association rule mining in the
CAMD Alzheimer's database
Balázs Szalkaia, Vince K. Grolmuszc, Vince I. Grolmusza,b,⁎, Coalition Against Major Diseases1
aPIT Bioinformatics Group, Eötvös University, H-1117 Budapest, Hungary
bUratim Ltd., H-1118 Budapest, Hungary
c2nd Department of Internal Medicine, Semmelweis University, Budapest, Hungary.
ARTICLE INFO
Keywords:
Alzheimer's diseaseAssociation rule miningCombinatorial biomarkersSCARFABSTRACT
The concept of combinatorial biomarkers was conceived when it was noticed that simple biomarkers are often
inadequate for recognizing and characterizing complex diseases. Here we present an algorithmic search method
for complex biomarkers which may predict or indicate Alzheimer's disease (AD) and other kinds of dementia. Weshow that our method is universal since it can describe any Boolean function for biomarker discovery. We applied
data mining techniques that are capable to uncover implication-like logical schemes with detailed quality
scoring. The new SCARF program was applied for the Tucson, Arizona based Critical Path Institute's CAMD
database, containing laboratory and cognitive test data for 5821 patients from the placebo arm of clinical trialsof large pharmaceutical companies, and consequently, the data is much more reliable than numerous other
databases for dementia. The results of our study on this larger than 5800-patient cohort suggest bene ficial eff ects
of high B12 vitamin level, negative eff ects of high sodium levels or high AST (aspartate aminotransferase) liver
enzyme levels to cognition. As an example for a more complex and quite surprising rule: Low or normal bloodglucose level with either low cholesterol or high serum sodium would also increase the probability of bad
cognition with a 3.7 multiplier. The source code of the new SCARF program is publicly available at http://
pitgroup.org/static/scarf.zip .
1. Introduction
Dementia is a major problem presently of high-income countries
and also an increasing concern of low-income nations worldwide.
Though sporadic before the age of 60, its occurrence is doubled by
every five years of age thereafter ( Bermejo-Pareja et al., 2008; Carlo
et al., 2002). About 40 percent of the population over 90 are a ffected,
and up to 20 percent of those between 75 and 84 su ffer from this
condition ( Prince & Jackson, 2009; Wortmann, 2012 ). The most
common cause of dementia is Alzheimer's disease (AD). The earliest
symptoms of AD include memory problems; disorientation in time andspace; and di fficulty with calculation, language, concentration and
judgment. As the disease evolves, patients may develop severe beha-vioral abnormalities and may even become psychotic. In the final stages
of the disease the su fferers are incapable of self-care and become bed-
bound, for years or even decades.
The diagnosis of AD in the great majority of the cases is done by
clinical criteria, using standardized questionnaires ( Mossello et al.,2010 ). Generally accepted evidences show that neuropathological da-
mage begins more than 20 years before those clinical signs ( Jack et al.,
2009 ), and by the time it is diagnosed, a large part of the neurons are
already irreversibly lost.
In the last years, by the combination of cerebrospinal fluid analysis,
clinical signs and neuroimaging techniques a quite reliable diagnosticmethod emerged ( Dubois et al., 2007 ). The method, however, is pro-
hibitively expensive, is not an early warning-type biomarker, and doesnot seem to be applicable for wide-scale screening of the senior popu-
lation.
Very recently, using the combination of usual clinical laboratory
data, cognitive impairment questionnaires and blood-based proteomicsassays was reported to reliably diagnose AD, without neuroimaging or
cerebrospinal fluid assays ( O’Bryant et al., 2010, 2011 ). However, early
warning biomarkers are still need to be found.
Thefinal goal of ours is finding new combinatorial biomarkers for
Alzheimer's disease. In this paper we report our results that may be usedto reach this final goal; but presently we are able to show only that
http://dx.doi.org/10.1016/j.archger.2017.08.006
Received 17 May 2017; Received in revised form 9 August 2017; Accepted 12 August 2017⁎Corresponding author at: PIT Bioinformatics Group, Eötvös University, H-1117 Budapest, Hungary.
1Data used in the preparation of this article were obtained from the Coalition Against Major Diseases database (CAMD). As such, the investigators within CAMD contributed to the
design and implementation of the CAMD database and/or provided data, but did not participate in the analysis of the data or the writing of this report.E-mail addresses: szalkai@pitgroup.org (B. Szalkai), grolmusz.vince@med.semmelweis-univ.hu (V.K. Grolmusz), grolmusz@pitgroup.org (V.I. Grolmusz).Archives of Gerontology and Geriatrics 73 (2017) 300–307
Available online 17 August 2017
0167-4943/ © 2017 Elsevier B.V. All rights reserved.
MARK

certain sets of laboratory data may make dementia (and not AD) more
probable, and certain other sets may make dementia less probable.
There are several large databases on Alzheimer's disease available
for researchers. The quality of their data obviously depends on the
methodology of the research that produced the database in question.
The most well-organized, strictly overseen and rigorously documented
experiments are perhaps conducted by the order of large pharmaceu-
tical companies in hospitals and clinics in phase 1, 2 and 3 drug trials.
Unfortunately, the detailed results of those trials are seldom published
(especially those corresponding to unsuccessful drug trials) since they
are owned by the companies that ordered the trials.
In their Alzheimer's disease database the Tucson, Arizona based
Critical Path Institute made available the results of the placebo arm of
numerous multi-million dollar clinical trials conducted by the order of
large pharmacological companies ( Rogers et al., 2012; Romero et al.,
2009, 2010 ). The data of the placebo line of the trials does not contain
proprietary information concerning the e ffects of the novel drugs under
trial, but it does contain reliable, well-organized laboratory and cog-nitive test-data, presumably in much higher quality than other, larger,
but perhaps less strictly conducted and controlled studies for AD.
Data used in the preparation of this article have been obtained from
the Coalition Against Major Diseases (CAMD) database ( Romero et al.,
2009 ). In 2008, Critical Path Institute, in collaboration with the En-
gelberg Center for Health Care Reform at the Brookings Institution,
formed the Coalition Against Major Diseases (CAMD). The Coalition
brings together patient groups, biopharmaceutical companies, and sci-
entists from academia, the US Food and Drug Administration (FDA), the
European Medicines Agency (EMA), the National Institute of Neurolo-
gical Disorders and Stroke (NINDS), and the National Institute on Aging
(NIA). Coalition Against Major Diseases (CAMD) includes over 200
scientists from member and non-member organizations. The data
available in the CAMD database have been volunteered by both CAMD
member companies and non-member organizations.
The data of the CAMD database were anonymized by an HIPAA Safe
Harbor compliant method ( Romero et al., 2009 ).
In contrast with more conservative statistical methods, we applied
data mining techniques for data analysis and combinatorial biomarkersearch. Data mining, as defi ned in Hand, Mannila, and Smyth (2001) ,i s
the analysis of large observational sets of data for finding new, still
unsuspected relations with novel, usually high-throughput methods.
Frequently, data mining uses large data sets originally collected for uses
other than the data mining analysis ( Hand et al., 2001 ).
Simple biomarkers (e.g., the high level of glucose in diabetes) show
a physiological condition, related to the appearance or the status of a
disease. The concept of combinatorial biomarkers appeared around
2010, and numerous authors simply use the term in the following sense:
If—say—the high concentration of all the molecules A,Band Cchar-
acterizes a certain condition Xwell (and the high concentration of any
subset of the set { A,B,C} would not), then they say that { A,B,C}i sa
combinatorial biomarker of the condition X(Wu et al., 2012). In
O’Bryant et al. (2011) , by applying proteomics assays, a 30-protein set
was identi fied as a combinatorial biomarker of AD.
We intend to discover more involved combinatorial biomarkers that
may
contain clinical laboratory data and psychiatric test data. We de-
scribe a universal method in the following sense:
Suppose binary variables x1,x2,…,xncorrespond to Boolean clin-
ical or laboratory data; that is, xi= 1 if the value of a measurement,
denoted by Mi, is in a given interval, and xi= 0 if the value of Miis
outside of the given interval. Now, a Boolean function f(x1,x2,…,xn)
that takes values either 0 or 1, can describe anyfunctional relation
between the set of the clinical or laboratory data that corresponds to
variables x1,x2,…,xnand the presence or absence of any clinical or
biological condition denoted by A:f(x1,x2,…,xn) = 1 corresponds to
the presence of the clinical/biological condition A, and f(x1,x2,…,xn)
= 0 corresponds to the absence of the clinical/biological condition A.
Since every Boolean function fcould be given by logicalconjunctions (ANDs) and logical disjunctions (ORs) of variables x1,x2,
…,xnor their negations, our method SCARF is universally able to find
any such conjunctions or disjunctions. That is the novelty of our
method.
Note that applying more than one binary xivariable for the same
measurement Mi, one can describe the distribution of the values of M
into more than one intervals, therefore our method is applicable forquantities described by non-Boolean variables as well.
We remark that simple biomarkers relate a single —not necessarily
Boolean —variable x
1to the presence or absence of condition A.I n
combinatorial biomarkers, like in O’Bryant et al. (2011) , a 30-protein
set was identi fied as a combinatorial biomarker of AD, that is, fwas just
the conjunction of 30 variables x1through x30, corresponding to the
presence of 30 proteins in measurable concentrations. Our SCARF
program is capable to discover any combinations of conjunctions and
disjunctions and negations, therefore, it can describe an arbitrary
Boolean function on variables x1,x2,…,xn.
We make use of an adequately modi fied association rule mining
(Hand et al., 2001 ) procedure, and we also apply a new methodology
that discovers complex combinatorial biomarkers only if these complexbiomarkers have stronger implications than the simpler biomarkers.
Therefore, our program SCARF will not produce arti ficially complex
biomarkers just for the sake of complexity: the more complex is the newbiomarker, the more valid is the new implication.
1.1. Association rule mining
Our research group was among the firsts applying association rule
mining in molecular biology ( Ivan, Szabadka, & Grolmusz, 2007). Re-
cently, association rule mining has been gaining applications in drug
discovery ( Galustian & Dalgleish, 2010), in the design of clinical trials
(Epstein, 2009 ), and most recently, also in image analysis in Alzhei-
mer's research ( Chaves et al., 2011 ).
Association rule mining is a field of data mining ( Hand et al., 2001 )
developed by marketing experts for discovering implication-like rules
in uncovering customer behavior ( Agrawal, Imielinski, & Swami, 1993),
without a priori assumptions on this behaviour.
The form of an association rule is as follows: X→Y, where XandY
are disjoint subsets of a ground-set of attributes or “properties ”S, and
X→Ymeans that the set of attributes Ximplies the set of attributes Y.
Both XandYare just two disjoint lists of properties, no logical opera-
tors or clauses are allowed.
We intended to apply a generalization of this method for laboratory
and cognitive test data from the CAMD database ( Romero et al., 2009).
We analyzed how the presence/absence/severity of cognitive im-
pairment could be detected from combinations of known biomarkers,demographic information and measurements of vital signs. As an ex-
ample, consider this expression:
== ≥ → ≤ sodium high&(protein high or age 60) mmset otal 1 5 (1)
Here & stands for logical AND or conjunction. This rule states that if
blood sodium is high, AND urine protein is high OR age is at least 60,
then the total MMSE (Mini Mental State Examination) score will be at
most 15 out of 30. Let us call the left-hand side of the expression (ab-
breviated by LHS) a combinatorial marker of the right-hand side (ab-
breviated by RHS). Thus the statement above can be reformulated as
follows: high serum sodium combined with either high urine protein or
age of at least 60 is a marker of a total MMSE score less than or equal to
15.
An expression consists of elementary clauses combined by logical
operators. These elementary clauses may include equalities and in-
equalities. By substituting all elementary clauses with some wildcard,
we can obtain the pattern of an expression. For example, the expression
above is of the following pattern:
□□□ → □&( o r ) (2)B. Szalkai et al. Archives of Gerontology and Geriatrics 73 (2017) 300–307
301

During our analysis we started with a given pattern like the one
above. Then we considered all the possible logical expressions ac-
cording to this pattern, and assigned numerical values to them that
indicated the reliability and validity of the logical rules. Then we fil-
tered and sorted the vast amount of possible rules according to thesenumerical criteria, and selected the best ones. We changed a simpler
rule to a more complex rule only if the more complex rule had higher
reliability/validity than the simpler rule (see the next section for the
exact defi nitions).
1.2. Comparison of SCARF with other methods
InNanavati, Chitrapura, Joshi, and Krishnapuram (2001) a method
is presented for mining generalized disjunctive association rules. Intheir work, the authors are searching for association rules of the form
□→□ □ … □ (o r o r) (3)
by using a heuristic approach: their algorithm does not necessarily re-turn all the valid implications of the given form. In comparison,
SCARF discovers association rules of the form (2)that is, we are
interested in rules uncovering “what causes dementia ”and not, as in
Nanavati et al. (2001) ,“what is caused by dementia ”;
If SCARF contained a disjunctive normal form (an or of & s, DNF) on
the left hand side of (2), then the rule could easily be decomposed
into several valid rules. This is the reason that SCARF mines CNFs(conjunctive normal forms, i.e., an & of “or”s).
While it is possible to convert a rule of the form (3)to a rule of the
form (2)by using logic operations, converting the numerical quality
criteria (support, con fidence, lift) do not seem to be possible with an
algorithmically tractable method;
SCARF is not a heuristic algorithm: it findsevery rule of the given
form, satisfying the numerical quality criteria; the algorithm de-
scribed in Nanavati et al. (2001) is heuristic.
The scopes of logic regression analysis (e.g., Ruczinski,
Kooperberg, & LeBlanc, 2003 ) and association rule mining are quite
different. In logic regression, one intends to approximate a function
(binomial- or multinomial) on its whole domain, and the quality of theapproximation is measured by some distance-related measure on the
whole domain. Therefore, logic regression analysis targets the ambi-
tious goal of approximating, for example, the “demented ”or“not-de-
mented ”multivariate function on the whole domain of study. In SCARF,
we intend to find combinatorial biomarkers that if present, imply the
demented status. Additionally, logic regression analysis never finds a
biomarker that is true only on the 1% of the domain: the associated
distance metric will be negligible. In association rule mining, however,
it is possible to find a risk factor that holds only for 1% of the cohort,
but there, strongly implies the dementia or other disease.
2. Materials and methods
Our data source, which will be referred to as CAMD from now on
(Romero et al., 2009), was provided by the Coalition Against Major
Diseases, and consisted of the placebo arm of several drug trials. Over
5800 subjects participated in these trials including demented and not
demented people of various age and sex (see Table 1) for basic statis-
tics). Standard laboratory data that have been collected for the subjectsincluded about 300 di fferent values in blood or urine altogether. These
values were generally measured multiple times per subject (on di fferent
visit days), though each person was tested for only about 30 di fferent
values. The cognitive and psychological status of the subjects was
measured at di fferent times by standardized questionnaires ADAS-COG,
ADCS-ADL, MMSE, NPI and SIB. In addition, some genetic tests havebeen performed, e.g., ApoE and MTHFR genotypes were recorded. Vital
sign measurements (BP, pulse rate, respiratory rate and bodytemperature) have also been taken. Results concerning this dataset will
be described in greater detail below.
We transformed this large dataset into a conveniently processable
form. The CAMD database contained several rows describing one
person and these were scattered between multiple data tables. So we
collected the essential data from CAMD into one single table: this
simpli fied table contained only one row for each subject.
If a subject was tested on di fferent visit days, then we took the
average of these test results. The resulting main table for CAMD con-
sisted of around 170 columns (record fields) and 5821 rows (entries).
Our main method of processing the resulting table was association
rule mining. First, we took a given pattern like □&(□or□)→□.
Notice that the LHS (Left Hand Side) is in conjunctive normal form here(multiple OR clauses ANDed together). This pattern matches all state-ments of the following kind: “if property A is present and property B or
property C is present, then property D is present ”.
Since we are interested in implication-like association rules that
indicate
factors implying normal or demented mental state, we made
restrictions on which data columns can occur on the LHS (left hand
side) and the RHS (right hand side). Laboratory data and sex were al-
lowed on the LHS, and columns directly indicating mental status on the
RHS. Then we gave numerical constraints on the “goodness ”of a ru-
le—thus introducing an ordering on the rules. Finally we tried to fill in
all the void boxes in all possible ways to find the best rules.
If done without any optimization, this process would have yielded a
vast amount of di fferent rules that would have needed to be evaluated
“by hand ”. Even just enumerating all the possible matches to this pat-
tern would have required enormous computational resources.
Consequently, we needed to make the computation feasible: we used a
branch-and-bound approach similar to the Apriori Algorithm
(Agrawal & Srikant, 1994): if certain values for the first two boxes made
a rule fail our constraints —regardless of what would be written in the
third box —, then we threw out the rule and did not bother checking all
the possible values for the third box.
More exactly, SCARF uses a branch-and-bound algorithm for finding
association rules. The LHS is filled from left to right. When a disjunctive
clause is completed, we calculate the universe, support and LHS supportof the rule. If one of these is already too small, we will not continue
building the rule, but advance to the next value for the lastly filled
elementary clause. Thus we eliminate the rest of that branch.
The possible rules can be represented as the leaves of a tree graph.
The root is the blank pattern itself, each inner node is an incompleterule (where only some blanks at the beginning are filled with equal-
ities), and the children of a node are those —possibly incomplete— rules
where one more blank is substituted with an equality.
The algorithm performs a depth- first search (DFS) on this tree. We
start from the root node. In each step, if the current node has any un-visited children, then we descend to one of them in the tree. Otherwise,
we ascend to the parent node. This is done until we have traversed the
whole tree and arrived back to the root.
In our case, the DFS means that the blanks are filled with attribute-
value pairs, one after the other, in a recursive manner. When we reach
the last blank, the rule is evaluated, and another equality is tried for theTable 1
Basic statistics on the subjects of the CAMD data.
Age distribution Gender
distributionMMSE distribution
A: up to 65
years1093 Female 3315 A: severe cog.
impairment255
B: 66 –75 years 2070 Male 2653 B: moderate cog.
impairment611
C: 76 –85 years 2408 C: mild cog.
impairment3224
D: more than 85 397 D: normal cognition 1352B. Szalkai et al. Archives of Gerontology and Geriatrics 73 (2017) 300–307
302

last blank, until we exhausted all the possibilities. After that, we ascend
to the second last blank, and try another equality for that. Then again
we descend to the last one, and try all equalities once again. This goes
on until we visited all the nodes in the tree.
Branch-and-bound means that we need not visit all the leaves, but
rather we are able to determine at some inner nodes that the corre-
sponding subtree will not yield good enough rules. If this occurs, we
decide to ascend immediately, without visiting the descendant nodes.
This is done by calculating bounds for the parameters of descendant
nodes. If those bounds imply that the parameters will not be good en-
ough for the rules to appear in the top Nlist, then we do not traverse the
subtree.
This technique saved a considerable computational time, and made
possible this study on a 5821 cohort.
The association rule mining was done with our own new program
written in the C++ programming language, named SCARF (SimpleCombinatorial Association Rule Finder). We calculated various stan-
dard numerical values for all association rules, which would indicate
their validity. First, we de fined the universe of a rule: this is the set of
the database rows where all columns present in the rule have a knownvalue. As we mentioned before, not all subjects were tested for every-
thing, so our database contained a large amount of N/A entries. For
testing the validity of a rule, only those rows could be taken into ac-
count, where there is no N/A written to any of the columns partici-
pating in the rule.
For evaluating the validity of a rule, we continued to work with only
its universe and temporarily discarded all other rows in the database.
Next, we calculated the LHS support ,RHS support andsupport of a rule.
The LHS support is the number of the rows where the LHS is true, the
RHS support is the number of the rows where the RHS is true, and the
support is the number of the rows where both the LHS and the RHS are
true.
Then, we calculated the confidence ,lift,leverage andχ
2-statistic for a
rule. The confidence is defi ned as the conditional probability of the RHS,
assuming that the LHS is true. If one has high serum sodium combined
with high urine protein or age at least 60 in our example, then con-
fidence describes the chance of having a low MMSE score. The liftshows
how many times the presence of the LHS increases the probability ofRHS. Generally it indicates how big a risk factor the LHS is —though it is
not certain that the LHS causes the RHS, as they both may be only
consequences of some background phenomenon ( Hand et al., 2001).
The leverage is the di fference between the observed probability of
both the LHS and RHS being true, and the estimated probability we get
by assuming that the LHS and RHS are independent events. It indicates
the level of dependency between the LHS and the RHS in a way. Finally,
theχ
2-statistic is a well-known measure of the estimated dependence of
the indicator variables of the LHS and RHS. The p-value output by
SCARF comes from this χ2test.
The E-value (also calculated by SCARF) equals to the p-value mul-
tiplied by the total number of possible rules. The E-value is a more
useful measure of randomness, since if we examine many rules, there isa high probability that the p-value will be small enough, while the E-
value is insensitive for this kind of artifact.
The following table formalizes some of the above de finitions. Here P
denotes the probability measure, and P(A|B) denotes the conditional
probability of event Aon condition B:
=
=
=∧ −P
PP PConfidence (RHS|LHS)
Lift
Leverage (RHS LHS) (RHS) (LHS)P
P(RHS | LHS)
(RHS)
For the CAMD database the acceptable values were set as follows:
universe ≥500, support ≥50, con fidence≥0.5, lift ≥1.2, p-
value≤0.05. In particular, we recorded rules on data that were mea-
sured on at least 500 subjects. We defi ned the goodness of a rule to beequal to its lift.
Therefore we listed association rules of lift at least 1.2, i.e., only
those rules were listed where the LHS increased the probability of RHS
with at least 20%.
One of the most signi ficant novelties in our approach was filtering
out those rules which were too complicated. The SCARF program threw
out elementary clauses from the LHS as long as the overall goodness
(i.e. the lift) of the rule did not decrease by more than 2%. Then it
deleted the whole rule if its numerical values dropped below our con-
straints during the simpli fication process. In other words, we sacri ficed
some of the lift for simplicity, to avoid over fitting.
Having listed the best rules, we also determined whether the ele-
mentary clauses (like lb_as t=h,lb_f olate =l, etc.) have positive or
negative e ffect on mental state. Therefore we counted their appearances
on LHS, and classi fied these occurrences by the nature of the RHS: does
it indicate normal cognition or rather dementia? We counted how many
times an elementary clause occurred on the LHS of a rule when the RHS
indicated a positive mental state, and how many times it occurred in
rules where the RHS showed a negative state. Thus, in addition to
mining rules whose LHS could probably serve as good combinatorial
risk factor of dementia, we estimated the contribution of the individual
clauses, for example “protein= high ”to the onset of cognitive impair-
ment.
For an elementary clause, Positive score was the number of rules with
positive RHS, and Negative score was the number of rules with negative
RHS. Then we compared Positive score with Negative score : by sub-
tracting the negative score from the positive score we got a value called
simply the score of the clause. Those elementary clauses whose score
was positive were called positive clauses, and similarly, those where the
score was negative were called negative clauses.
To summarize our method: we searched for combinatorial bio-
markers using a branch-and-bound algorithm for association rule
mining; then made statistical analysis regarding elementary clauses.
The source code of the new SCARF program is publicly available at
http://pitgroup.org/static/scarf.zip .
3.
Results
The program outputs 725 rules from the CAMD database. Selected
rules, ordered by lift (i.e. “goodness ”) decreasing are listed in Table 2.
The whole set of rules is presented as Table S1 of the online supportingmaterial at http://uratim.com/CAMD .
On the LHS, clauses concerning biomarkers end in “=l”,“=h”,
“=n”, or combinations of these. Here lmeans low, hmeans high and n
means normal. If there are multiple letters (such as nh), then the cor-
responding equality states that the value is either high or normal. In
other words, single letters correspond to a value category, while mul-
tiple letters mean the union of these categories.
For example, the second rule in Table 2 was that of the second best
lift. It can be interpreted in the following way: It is likely that if serum
sodium level is elevated, and serum glucose level is either too low or
normal, then the total MMSE score will be less than 15. Note that it is
true for all rules of ours that there is not necessarily a causal relation
between the LHS and RHS, as both the LHS and RHS can be con-
sequences of an unknown process in the background.
The third rule states that “if serum sodium level is elevated, and
calcium level is either low or normal, then MMSE orientation sub-score
will be at most 2” . The seventh rule in Table 2 states that “if serum
sodium level is elevated, and body temperature is too low, then totalMMSE score will be less than 15” .
From these selected rules we can conclude that elevated sodium
combined with various other factors (not too high glucose, not too highcalcium, low temperature) might be a good indicator (or even the
cause) of mental decline.
Elementary clauses with the greatest positive e ffect on normal
cognition are listed in Table 5.B. Szalkai et al. Archives of Gerontology and Geriatrics 73 (2017) 300–307
303

Elementary clauses with the greatest negative e ffect on normal
cognition are listed in Table 6.
4. Discussion
Among the 725 rules identi fied, 513 had lift values exceeding 2.00.
Most of the rules exceeding even the 3.00 lift value had one thing in
common: the LHS contained the premise lb_sodium=h .
4.1. Liver function
The rules found suggest that having high serum levels of AST (as-
partate aminotransferase), as well as having low or high serum levels of
ALT (alanine aminotransferase) may predispose to an impaired cogni-
tion characterized by low mini mental state examination (MMSE)
scores. It should be noted that low ALT was much more rare in the
CAMD database than high ALT, so the negative e ffect should be at-
tributed mainly to high ALT. However, serum ALP (alkaline phospha-
tase) levels seem to have a controversial e ffect on mental status.
AST, ALT and ALP levels derive from the liver. Elevated ALP might
indicate bile duct obstruction. AST or ALT may elevate in a number of
cases of liver injury or damage, spreading from acute or chronic viral
infections to alcohol induced or non-alcoholic steatohepatitis. It is in-
teresting to note that elevated serum levels of AST (more than those ofALT) have been associated with impaired mental status. Although mild
elevations in serum levels of AST and ALT are nonspeci fic to the
etiology of liver injury, certain alteration patterns in these parametersmay re flect the nature of the hepatic disease. For instance, the value of
the AST/ALT ratio— also known as the De Ritis ratio —is approximately
0.8 in normal subjects, a ratio exceeding 2.00 being suggestive to al-
coholic hepatitis.
Therefore we scanned the subjects with high AST values for higher
than 2 AST/ALT ratio: we have only found 10 subjects satisfying both
conditions. In addition, only 2 rules had AST/ALT on the left-hand side.
Consequently, we may assume that high serum AST in the study sub-
jects is not typically accompanied with high De Ritis ratio (i.e. probable
alcoholic hepatitis).
The association of impaired liver function with mental decline can
be illuminated in two perspectives. On one hand, impaired liver func-
tion might be insu fficient to prevent the brain from the e ffects of certain
neurotoxins, e.g. ammonia. This happens in the case of hepaticTable 2
Several association rules of the highest lift. The lift value describes the multiplicationfactor, increasing the probability of the right hand side (RHS) if the left hand side is true.For example, our best rule (the first below) is saying that one can have the a bad result of
a cognitive test with four times higher probability if one has high serum sodium andeither low cholesterol or low or normal blood glucose level. We refer to Table 4 for the
legends applied in this table, and to subsection “Association Rule Mining ”in the In-
troduction for the simple implication-like formalism that describes the association rules
found.
(lb_sodium=h) &(lb_chol=l or lb_gluc=ln) —> mm_ori=B
Universe: 2783, LHS support: 87, RHS support: 401, Support: 50Confidence: 0.574713, Lift: 3.98859, Leverage: 0.0134618, p-value:
0,E-value: 0
3.98859
(lb_gluc=ln) &(lb_chol=l or lb_sodium=h) —> mm_ori=B
Universe: 2783, LHS support: 105, RHS support: 401, Support: 57
Confidence: 0.542857, Lift: 3.76751, Leverage: 0.0150451, p-value:
0,E-value: 0
3.76751
(lb_sodium=h) &(lb_hct=l or lb_gluc=ln) —> mm_ori=B
Universe: 2926, LHS support: 95, RHS support: 420, Support: 51
Confidence: 0.536842, Lift: 3.74, Leverage: 0.0127695, p-value: 0,
E-value: 0
3.74
(lb_sodium=h) &(bpsys=ln or lb_gluc=ln) —> mm_ori=B
Universe: 3091, LHS support: 102, RHS support: 425, Support: 52
Confidence: 0.509804, Lift: 3.70777, Leverage: 0.0122858, p-value:
0,E-value: 0
3.70777
(lb_gluc=ln) &(lb_creat=l or lb_sodium=h) —> mm_ori=B
Universe: 3091, LHS support: 99, RHS support: 425, Support: 50
Confidence: 0.505051, Lift: 3.6732, Leverage: 0.0117722, p-value:
0,E-value: 0
3.6732
(lb_sodium=h) & (age=D or lb_gluc=ln) —> mm_ori=B
Universe: 3091, LHS support: 101, RHS support: 425, Support: 51
Confidence: 0.50495, Lift: 3.67248, Leverage: 0.0120068, p-value:
0,E-value: 0
3.67248
(lb_gluc=ln) &(lb_ast=l or lb_sodium=h) —> mm_ori=B
Universe:
3091, LHS support: 101, RHS support: 425, Support: 51
Confidence: 0.50495, Lift: 3.67248, Leverage: 0.0120068, p-value:
0,E-value: 0
3.67248Table 3
Some association rules involving serum cholesterol level. We refer to Table 4 for the
legends applied in this table, and Section 1.1for the simple implication-like formalism
that describes the association rules found.
(lb_sodium=h) & (lb_chol=l or lb_gluc=ln) – – – > mm_ori=B
Universe: 2783, LHS support: 87, RHS support: 401, Support: 50Confidence: 0.574713, Lift: 3.98859, Leverage: 0.0134618, p-value: 0, E-value: 0
3.98859
(lb_gluc=ln) & (lb_chol=l or lb_sodium=h) – – – > mm_ori=B
Universe: 2783, LHS support: 105, RHS support: 401, Support: 57Confidence: 0.542857, Lift: 3.76751, Leverage: 0.0150451, p-value: 0, E-value: 0
3.76751
(lb_sodium=h) &(lb_chol=ln or lb_gluc=ln) —> mm_ori=B
Universe: 2783, LHS support: 106, RHS support: 401, Support: 55
Confidence: 0.518868, Lift: 3.60102, Leverage: 0.0142747, p-value:
0,E-value: 0
3.60102
(lb_chol=h) &(lb_cl=h or lb_sodium=h) —> mm_ori=B
Universe: 1420, LHS support: 71, RHS support: 304, Support: 51
Confidence: 0.71831, Lift: 3.35526, Leverage: 0.0252113, p-value:
2.22045e-016, E-value: 1.88773e −007
3.35526
(lb_chol=h) &(lb_monole=l or lb_sodium=h) —> mm_total=AB
Universe: 1364, LHS support: 73, RHS support: 325, Support: 58
Confidence: 0.794521, Lift: 3.33454, Leverage: 0.02977, p-value:
1.51101e −013, E-value: 0.00012846
3.33454
(lb_sodium=h) &(lb_monole=h or lb_chol=h) —> mm_total=AB
Universe: 1364, LHS support: 66, RHS support: 325, Support: 51
Confidence: 0.772727, Lift: 3.24308, Leverage: 0.0258608, p-value:
5.9952e −015, E-value: 5.09687e −006
3.24308
(lb_chol=h) &(lb_hbsag=h or lb_sodium=h) —> mm_attcal=B
Universe: 1164, LHS support: 67, RHS support: 312, Support: 50
Confidence: 0.746269, Lift: 2.78416, Leverage: 0.0275268, p-value:
6.2725e −011, E-value:
0.0533262
2.78416
(lb_sodium=h) &(lb_bun=h or lb_chol=h) —> mm_attcal=B
Universe: 1387, LHS support: 61, RHS support: 429, Support: 52
Confidence: 0.852459, Lift: 2.75609, Leverage: 0.023888, p-value:
8.87168e −012, E-value: 0.00754232
2.75609
(lb_sodium=h) &(lb_ca=l or lb_chol=h) —> mm_attcal=B
Universe: 1420, LHS support: 61, RHS support: 460, Support: 51
Confidence: 0.836066, Lift: 2.5809, Leverage: 0.0219996, p-value:
2.52266e −011, E-value: 0.0214466
2.5809
(lb_sodium=h) &(lb_cl=h or lb_chol=h) —> mm_attcal=B
Universe: 1420, LHS support: 66, RHS support: 460, Support: 55
Confidence: 0.833333, Lift: 2.57246, Leverage: 0.0236759, p-value:
1.65421e −010, E-value: 0.140634
2.57246B. Szalkai et al. Archives of Gerontology and Geriatrics 73 (2017) 300–307
304

encephalopathy (HE), when severe liver damage resulting in acute liver
insufficiency cannot detoxi ficate ammonia and other neurotoxins. On
the other hand, the association of elevated AST/ALT ratio with im-paired mental status proposes that another obscure element (e.g.
chronic alcohol consumption) might be the factor responsible for both
cognitive and metabolic damages.
Our results raise the possibility of a pathogenetic linkage between
liver function and mental status in patients with AD. Such linkage has
also been proposed by other studies ( Astarita et al., 2010; Sutcli ffe,
Hedlund, Thomas, Bloom, & Hilbush, 2011). One study concludes thatperipheral reduction of β-amyloid is su fficient to reduce brain β-amy-
loid and proposes that β-amyloids, which are of major pathogenic im-
portance in AD may originate from the liver ( Sutcliffe et al., 2011).
Another research found that defi cient liver production of docosahex-
aenoic acid (a neuroprotective fatty acid) correlates with impaired
cognitive status in AD patients ( Astarita et al., 2010).
To rule out the possibility when the elevated AST level is due to
some medications taken, we compiled a detailed in Table S3 (in the
supporting on-line material at http://uratim.com/CAMD ) containing
the number of subjects taking certain drugs, and the number of drug-takers with high AST. The data shows that, for example, 1929 subjects
took Donepezil, while among the Donepezil-takers, only 415 have had
high AST levels.
4.2. Serum sodium
A great number of rules (224) have high sodium on the left hand
side, all of which have impaired cognition on the right hand side. NetTable 4
The legends for Tables 2 and 3 . The MMSE-scores and age-classi fications by the letters A,
B, C and D are also de fined in this table.
age Subject age (A: ≤65 years, B: 66 –75 years, C: 76 –85 years,
D: > 85 years)
ast_alt De Ritis ratio
bpdia Diastolic blood pressurebpsys Systolic blood pressurelb_alb Serum albuminelb_alp Serum alkaline phosphataselb_alt Serum alanine aminotransferase
lb_ast Serum aspartate aminotransferase
lb_baso Basophils, particle concentrationlb_bili Serum indirect bilirubinlb_bun Blood Urea Nitrogenlb_ca Serum calciumlb_chol Serum cholesterollb_ck Serum creatine kinase
lb_cl Serum chlorine
lb_creat Serum creatininelb_eos Eosinophils, particle concentrationlb_gluc Serum glucoselb_hba1c Hemoglobin A1Clb_hbsag Hepatitis B virus surface antigenlb_hct Hematocritlb_hgb_blood Blood hemoglobin
lb_k Serum potassium
lb_ketones Ketoneslb_ldh Lactate dehydrogenaselb_lym Lymphocytes, particle concentrationlb_lymle Lymphocytes/leukocytes ratiolb_mch Mean corpuscular hemoglobinlb_mchc Mean corpuscular hemoglobin concentrationlb_mcv Mean corpuscular volume
lb_mono Monocytes, particle concentration
lb_monole Monocytes/leukocytes ratiolb_neut Neutrophils, particle concentrationlb_neutle Neutrophils/leukocytes ratiolb_ph pHlb_phos Phosphatelb_plat Platelets
lb_prot Total protein
lb_rbc_blood Red blood countlb_sodium Serum sodiumlb_tsh Thyrotropinlb_vitb12 Serum B12 vitaminlb_wbc_blood White blood countmm_attcal MMSE attention and calculation subscore (B: 0 –1, C: 2, D: 3, E:
4–5)
mm_lang MMSE language subscore (B: 0 –2, C: 3 –4, D: 5 –6, E: 7 –9)
mm_ori MMSE orientation subscore (B: 0 –2, C: 3 –4, D: 5 –7, E: 8 –10)
mm_recall MMSE recall subscore (B: 0, C: 1, D: 2, E: 3)mm_total MMSE total score (A: < 10, B: 10 –14, C: 15 –23, D:≥24)
pulse Heart rateresp Respiratory ratesex Subject sex (F: female, M: male)temper Temperature
Table 5Elementary clauses with the greatest positive e ffect on
normal cognition. The “score ”values refer to the MMSE
scores. For the de finition of the “elementary clause ”we
refer to Section 1.1.
lb_vitb12=h score: 67lb_mch=h score: 25
lb_mchc=l score: 22
lb_k=h score: 17sex=M score: 10pulse=l score: 9lb_bun=l score: 8age=AB score: 4lb_mono=nh score: 3resp=ln score: 3
lb_plat=ln score: 2
lb_eos=nh score: 2lb_prot=nh score: 2Table 6Elementary clauses with the greatest negative e ffect on normal
cognition. The “score ”values refer to the MMSE scores. For the
definition of the “elementary clause ”we refer to Section 1.
temper=nh score: −10
lb_wbc_blood=h score: −10
age=BCD score: −10
lb_prot=h score: −12
lb_gluc=h score: −12
pulse=h score: −12
lb_ck=h score: −12
lb_hct=nh score: −12
lb_k=ln score: −12
lb_alp=h score: −12
lb_chol=ln score: −13
lb_ph=h score: −13
lb_hct=l score: −13
lb_alt=h score: −13
age=A score: −14
bpsys=ln score: −14
lb_creat=ln score: −14
lb_creat=h score: −16
temper=l score: −17
lb_alp=ln score: −18
lb_bun=ln score: −18
lb_alt=l score: −19
lb_wbc_blood=l score: −20
lb_chol=l score: −21
pulse=nh score: −21
lb_prot=ln score: −22
lb_bun=h score: −22
lb_plat=h score: −26
lb_gluc=ln score:−27
bpdia=ln
score:−28
age=CD score:−32
lb_chol=h score:−42
lb_ast=h score:−43
lb_ca=l score:−50
sex=F score:−57
age=D score:−99
lb_cl=h score:−173
lb_sodium=h score:−224B. Szalkai et al. Archives of Gerontology and Geriatrics 73 (2017) 300–307
305

water loss is responsible for the majority of cases of hypernatremia
(Adrogue & Madias, 2000 ). A recent publication, examining the causes
and comorbidities in patients older than 65 years, has found that themost common cause of community-acquired hypernatremia is dehy-
dration due to reduced oral intake ( Turgutalp et al., 2012 ). More in-
terestingly, they found that the most common comorbidity in this pa-tient group was AD, present in 31.4% of patients with hypernatremia
(Turgutalp et al., 2012 ). Hydration status has a signi ficant impact on
the volume of grey and white matter in the brain and on the quantity ofcerebrospinal fluid as a hallmark of ventricular enlargement
(Streitbuerger et al., 2012 ). The pattern of shrinkage in white matter
volume and increase of the ventricular system due to dehydration is
consistent with the structural brain changes observed during the pro-
gression of AD ( Streitbuerger et al., 2012 ). In another study, patients
with AD underwent bioelectrical impedance vector analysis to assessthe body cell mass and hydration status related to AD ( Buffa, Mereu,
Putzu, Floris, & Marini, 2010). Results demonstrated a tendency to-wards dehydration in patients with AD ( Buffa et al., 2010 ). Although
the association of dehydration and AD is supported by these publica-tions, the speci fic pathogenic nature of this association remains obscure
(Buffa et al., 2010; Streitbuerger et al., 2012; Turgutalp et al., 2012 ).
4.3. Vitamin B12
Our results were able to present the bene ficial impact of high levels
of vitamin B12, also known as cobalamin, on cognition. Along with
folate, vitamin B12 has an important role in the maintenance of genome
integrity ( Fenech, 2012 ). Although previous publications found asso-
ciation of low serum levels of vitamin B12 and AD ( Malaguarnera et al.,
2004; McCaddon et al., 2004), a recent systemic review on vitamin B12
status and cognitive impairment fails to declare a clear associationbetween vitamin B12 status and dementia ( O’Leary, Allman-
Farinelli, & Samman, 2012 ). However, this review also found that stu-
dies using newer and more speci fic biomarkers of vitamin B12 status
such as methylmalonic acid and holotranscobalamin were able to drawan association between mental decline and poor vitamin B12 status
(O’Leary et al., 2012 ).
Although clinically vitamin B12 defi ciency may result in macrocytic
anaemia, in the case of AD patients the occurrence of macrocytic
anaemia is rare and the neurological and hematological features are
unrelated ( McCaddon et al., 2004 ).
4.4. Hematological parameters
Additional interesting rules were detected regarding hematological
parameters. In particular, independently from each other, high valuesof mean corpuscular hemoglobin (MCH), low values of mean corpus-
cular hemoglobin concentration (MCHC), and low values of mean
corpuscular volume (MCV) were also associated with high MMSE
scores. Although high values of MCH and low values of MCHC are
present in the case of macrocyctic anaemia (with the addition of high
levels of mean corpuscular volume, low levels of hemoglobin and he-
matocrit), such solely associations should not be discussed, as they may
be coincidental.
Among the rules with lift values exceeding 2.00, other parameters of
hematological status (such as level of hemoglobin, red blood cell
number, white blood cell number) were also present. Monocyte and
eosinophil levels also appear on the left hand side of many rules with
high lift. These premises appear in combinations with various other(mostly
non-hematological) premises.
4.5. Blood cholesterol and cognition
The positive or negative e ffects of high cholesterol values to
Alzheimer's disease and cognition is a controversial issue. Some studies
(e.g., Helzner et al., 2009; Whitmer, Sidney, Selby, Johnston, & Ya ffe,2005; Zambon et al., 2010 ) show negative e ffects of high cholesterol
value for cognition, while other studies ( Mielke et al., 2005 ;Reitz,
Luchsinger, Tang, Manly, & Mayeux, 2005; Reitz, Tang,
Luchsinger, & Mayeux, 2004 ) prove the positive e ffects for cognition.
Our data supports both conclusions in a sense. That is, low, low-
normal and high cholesterol levels are all associated with impaired
mental status, but with a di fferent extent (scores −21,−13 and−42,
respectively). See Table 3 for a selection of cholesterol-related rules
from the larger Table S1 in the on-line supporting material at http://
uratim.com/CAMD .
It is worth to note that, by Table 3, elevated, low or low-normal
cholesterol levels do not necessarily mean a higher likelihood of im-paired cognition by themselves, but only combined with high sodium.
A most recent study ( Pierrot et al., 2013) shows that the neuronal
expression of amyloid precursor protein APP controls the cholesterol
24-hydroxylase mRNA levels and decreases cholesterol turnover;
therefore in certain setups, the presence of amyloid precursor proteins
imply lowered cholesterol levels.
5. Conclusions
A 5821-patient, high-quality database was analyzed with original
methods for combinatorial biomarkers of dementia. We have found
some novel and also some already well established relations connected
to cognition characteristics in a 5821 patient cohort. The already es-
tablished findings prove the validity of our datamining approach, and
the new findings, related to MCH, ALP and AST levels prove its power.
Some more controversial biomarkers, including cholesterol level, were
also re-discovered, and were placed into context of other attributes for
negative and positive e ffects to cognition.
Conflicts of interest
The authors declare no con flicts of interest.
Acknowledgements
Data used in the preparation of this article have been obtained from
the Coalition Against Major Diseases (CAMD) database ( Romero et al.,
2009 ). In 2008, Critical Path Institute, in collaboration with the En-
gelberg Center for Health Care Reform at the Brookings Institution,
formed the Coalition Against Major Diseases (CAMD). The Coalition
brings together patient groups, biopharmaceutical companies, and sci-
entists from academia, the US Food and Drug Administration (FDA), the
European Medicines Agency (EMA), the National Institute of Neurolo-
gical Disorders and Stroke (NINDS), and the National Institute on Aging
(NIA). Coalition Against Major Diseases (CAMD) includes over 200
scientists from member and non-member organizations. The data
available in the CAMD database have been volunteered by both CAMD
member companies and non-member organizations.
Data used in the preparation of this article were obtained from the
Coalition Against Major Diseases database (CAMD; http://codr.cpath.
org). As such, the investigators within CAMD contributed to the designand implementation of the CAMD database and/or provided data, but
did not participate in the analysis of the data or the writing of thisreport.
Special
thanks to Neville J., Kopko S., Broadbent S., Avilés E.,
Stafford R., Solinsky C., Bain L.J., Cisneroz M., Romero K. and
Stephenson D.
BS was supported through the new national excellence program of
the Ministry of Human Capacities of Hungary. VG is supported by theVEKOP-2.3.2-16 program of National Research, Development and
Innovation O ffice of Hungary.B. Szalkai et al. Archives of Gerontology and Geriatrics 73 (2017) 300–307
306

Appendix A. Supplementary data
Supplementary data associated with this article can be found, in the
online version, at http://dx.doi.org/10.1016/j.archger.2017.08.006 .
References
Adrogue, H. J., & Madias, N. E. (2000). Hypernatremia. New England Journal of Medicine,
342(May (20)), 1493 –1499. http://dx.doi.org/10.1056/NEJM200005183422006 .
Agrawal, R., & Srikant, R. (1994). Fast algorithms for mining association rules in large
databases. In B. Jorge, M. J. Bocca, & C. Zaniolo (Eds.). Proc. of the 20th international
conference on very large data bases (VLDB ‘94), vol. 1215 (pp. 487 –499).
Agrawal, R., Imielinski, T., & Swami, A. N. (1993). Mining association rules between sets
of items in large databases. In P. Buneman, & S. Jajodia (Eds.). Proceedings of the 1993
ACM SIGMOD international conference on management of data (pp. 207 –216).
Astarita, G., Jung, K. M., Berchtold, N. C., Nguyen, V. Q., Gillen, D. L., Head, E., et al.
(2010). De ficient liver biosynthesis of docosahexaenoic acid correlates with cognitive
impairment in Alzheimer's disease. PLoS ONE, 5 (9), e12538. http://dx.doi.org/10.
1371/journal.pone.0012538 .
Bermejo-Pareja, F., Benito-Leon, J., Vega, S., Medrano, M. J., Roman, G. C., & in Central
Spain (NEDICES) Study Group ND (2008). Incidence and subtypes of dementia in
three elderly populations of central Spain. Journal of the Neurological Sciences,
264(January (1 –2)), 63 –72.
Buffa, R., Mereu, R. M., Putzu, P. F., Floris, G., & Marini, E. (2010). Bioelectrical im-
pedance vector analysis detects low body cell mass and dehydration in patients with
Alzheimer's disease. Journal of Nutrition Health and Aging, 14 (December (10)),
823–827.
Carlo, A. D., Baldereschi, M., Amaducci, L., Lepore, V., Bracco, L., Maggi, S., et al. (2002).
Incidence of dementia, Alzheimer's disease, and vascular dementia in Italy. The ILSAStudy. Journal of the American Geriatrics Society, 50 (January (1)), 41 –48.
Chaves, R., Gorriz, J. M., Ramirez, J., Illan, I. A., Salas-Gonzalez, D., & Gomez-Rio, M.
(2011). E fficient mining of association rules for the early diagnosis of Alzheimer's
disease. Physics in Medicine and Biology, 56 (September (18)), 6047 –6063. http://dx.
doi.org/10.1088/0031-9155/56/18/017 .
Dubois, B., Feldman, H. H., Jacova, C., Dekosky, S. T., Barberger-Gateau, P., Cummings,
J., et al. (2007). Research criteria for the diagnosis of Alzheimer's disease: Revising
the NINCDS-ADRDA criteria. Lancet Neurology, 6 (August (8)), 734 –746. http://dx.
doi.org/10.1016/S1474-4422(07)70178-3 .
Epstein, R. J. (2009). Unblocking blockbusters: Using Boolean text-mining to optimise
clinical trial design and timeline for novel anticancer drugs. Cancer Informatics, 7 ,
231–238.
Fenech, M. (2012). Folate (vitamin B9) and vitamin B12 and their function in the
maintenance of nuclear and mitochondrial genome integrity. Mutation Research,
733(May (1 –2)), 21 –33.http://dx.doi.org/10.1016/j.mrfmmm.2011.11.003 .
Galustian,
C., & Dalgleish, A. G. (2010). The power of the web in cancer drug discovery
and clinical trial design: Research without a laboratory? Cancer Informatics, 9 ,3 1–35.
Hand, D. J., Mannila, H., & Smyth, P. (2001). Principles of data mining. MIT Press .
Helzner, E. P., Luchsinger, J. A., Scarmeas, N., Cosentino, S., Brickman, A. M., Glymour,
M. M., et al. (2009). Contribution of vascular risk factors to the progression in
Alzheimer's disease. Archives of Neurology, 66 (March (3)), 343 –348. http://dx.doi.
org/10.1001/archneur.66.3.343 .
Ivan, G., Szabadka, Z., & Grolmusz, V. (2007). Being a binding site: Characterizing residue
composition of binding sites on proteins. Bioinformation, 2 (5), 216 –221.
Jack, C. R., Lowe, V. J., Weigand, S. D., Wiste, H. J., Senjem, M. L., Knopman, D. S., et al.
(2009). Serial PIB and MRI in normal, mild cognitive impairment and Alzheimer's
disease: Implications for sequence of pathological events in Alzheimer's disease.
Brain, 132 (May (Pt 5)), 1355 –1365. http://dx.doi.org/10.1093/brain/awp062 .
Malaguarnera, M., Ferri, R., Bella, R., Alagona, G., Carnemolla, A., & Pennisi, G. (2004).
Homocysteine, vitamin B12 and folate in vascular dementia and in Alzheimer's dis-ease. Clinical Chemistry and Laboratory Medicine, 42 (9), 1032 –1035. http://dx.doi.
org/10.1515/CCLM.2004.208 .
McCaddon, A., Tandy, S., Hudson, P., Gray, R., Davies, G., Hill, D., et al. (2004). Absence
of macrocytic anaemia in Alzheimer's disease. Clinical and Laboratory Haematology,
26(August (4)), 259 –263. http://dx.doi.org/10.1111/j.1365-2257.2004.00618.x .
Mielke, M. M., Zandi, P. P., Sjogren, M., Gustafson, D., Ostling, S., Steen, B., et al. (2005).
High total cholesterol levels in late life associated with a reduced risk of dementia.Neurology, 64 (May (10)), 1689 –1695. http://dx.doi.org/10.1212/01.WNL.
0000161870.78572.A5 .
Mossello, E., Ballini, E., Mello, A. M., Tarantini, F., Simoni, D., Baldasseroni, S., et al.
(2010). Biomarkers of Alzheimer's disease: From central nervous system to per-iphery? International Journal of Alzheimer's Disease, 2011 , 342980. http://dx.doi.org/
10.4061/2011/342980 .Nanavati, A. A., Chitrapura, K. P., Joshi, S., & Krishnapuram, R. (2001). Mining gen-
eralised disjunctive association rules. In L. L. Henrique Paques, & D. Grossman (Eds.).Proceedings of the tenth international conference on Information and knowledge man-
agement (CIKM ‘01)(pp. 482 –489). New York, NY, USA: ACM Available from http://
dl.acm.org/citation.cfm?id=502666 http://dl.acm.org/cita-
tion.cfm?id=502666http://dl.acm.org/citation.cfm?id=502666.
O’Bryant, S. E., Xiao, G., Barber, R., Reisch, J., Doody, R., Fairchild, T., et al. (2010). A
serum protein-based algorithm for the detection of Alzheimer disease. Archives of
Neurology, 67 (September (9)), 1077 –1081. http://dx.doi.org/10.1001/archneurol.
2010.215 .
O’Bryant,
S. E., Xiao, G., Barber, R., Reisch, J., Hall, J., Cullum, C. M., et al. (2011). A
blood-based algorithm for the detection of Alzheimer's disease. Dementia and Geriatric
Cognitive Disorders, 32 (1), 55 –62.http://dx.doi.org/10.1159/000330750 .
O’Leary, F., Allman-Farinelli, M., & Samman, S. (2012). Vitamin B12 status, cognitive
decline and dementia: A systematic review of prospective cohort studies. British
Journal of Nutrition, 108 (December (11)), 1948 –1961. http://dx.doi.org/10.1017/
S0007114512004175 .
Pierrot, N., Tyteca, D., D ’auria, L., Dewachter, I., Gailly, P., Hendrickx, A., et al. (2013).
Amyloid precursor protein controls cholesterol turnover needed for neuronal activity.
EMBO Molecular Medicine, 5 (April (4)), 608 –625. http://dx.doi.org/10.1002/emmm.
201202215 .
Prince, M., & Jackson, J. (2009). World Alzheimer Report 2009 Alzheimer's Disease
International Available from http://www.alz.co.uk/research/ files/
WorldAlzheimerReport.pdf .
Reitz, C., Tang, M. X., Luchsinger, J., & Mayeux, R. (2004). Relation of plasma lipids to
Alzheimer disease and vascular dementia. Archives of Neurology, 61 (May (5)),
705–714. http://dx.doi.org/10.1001/archneur.61.5.705 .
Reitz, C., Luchsinger, J., Tang, M. X., Manly, J., & Mayeux, R. (2005). Impact of plasma
lipids and time on memory performance in healthy elderly without dementia.Neurology, 64 (April (8)), 1378 –1383. http://dx.doi.org/10.1212/01.WNL.
0000158274.31318.3C .
Rogers, J. A., Polhamus, D., Gillespie, W. R., Ito, K., Romero, K., Qiu, R., et al. (2012).
Combining patient-level and summary-level data for Alzheimer's disease modelingand simulation: A beta regression meta-analysis. Journal of Pharmacokinetics and
Pharmacodynamics, 39 (October (5)), 479 –498. http://dx.doi.org/10.1007/s10928-
012-9263-3 .
Romero, K., de Mars, M., Frank, D., Anthony, M., Neville, J., Kirby, L., et al. (2009). The
Coalition Against Major Diseases: Developing tools for an integrated drug develop-ment process for Alzheimer's and Parkinson's diseases. Clinical Pharmacology and
Therapeutics, 86 (October (4)), 365 –367. http://dx.doi.org/10.1038/clpt.2009.165 .
Romero, K., Corrigan, B., Tornoe, C. W., Gobburu, J. V., Danhof, M., Gillespie, W. R., et al.
(2010). Pharmacometrics as a discipline is entering the “industrialization ”phase:
Standards, automation, knowledge sharing, and training are critical for future suc-cess. Journal of Clinical Pharmacology, 50 (September (9 Suppl.)), 9S –19S. http://dx.
doi.org/10.1177/0091270010377788 .
Ruczinski, I., Kooperberg, C., & LeBlanc, M. (2003). Logic regression. Journal of
Computational and Graphical Statistics, 12 (3), 475 –511. http://dx.doi.org/10.1198/
1061860032238 .
Streitbuerger,
D. P., Moller, H. E., Tittgemeyer, M., Hund-Georgiadis, M., Schroeter, M. L.,
& Mueller, K. (2012). Investigating structural brain changes of dehydration usingvoxel-based morphometry. PLoS ONE, 7 (8), e44195.. http://dx.doi.org/10.1371/
journal.pone.0044195 .
Sutcliffe, J. G., Hedlund, P. B., Thomas, E. A., Bloom, F. E., & Hilbush, B. S. (2011).
Peripheral reduction of beta-amyloid is su fficient to reduce brain beta-amyloid:
Implications for Alzheimer's disease. Journal of Neuroscience Research, 89 (June (6)),
808–814. http://dx.doi.org/10.1002/jnr.22603 .
Turgutalp, K., Ozhan, O., Gok Oguz, E., Yilmaz, A., Horoz, M., Helvaci, I., et al. (2012).
Community-acquired hypernatremia in elderly and very elderly patients admitted tothe hospital: Clinical characteristics and outcomes. Medical Science Monitor,
18(December (12)), CR729 –CR734 .
Whitmer, R. A., Sidney, S., Selby, J., Johnston, S. C., & Ya ffe, K. (2005). Midlife cardio-
vascular risk factors and risk of dementia in late life. Neurology, 64 (January (2)),
277–281. http://dx.doi.org/10.1212/01.WNL.0000149519.47454.F2 .
Wortmann, M. (2012). Dementia: A global health priority –Highlights from an ADI and
World Health Organization report. Alzheimer's Research and Therapy, 4 (September
(5)), 40. http://dx.doi.org/10.1186/alzrt143 .
Wu, W., Juan, W. C., Liang, C. R. M. Y., Yeoh, K. G., So, J., & Chung, M. C. M. (2012).
S100A9, GIF and AAT as potential combinatorial biomarkers in gastric cancer diag-
nosis and prognosis. Proteomics Clinical Applications, 6 (April (3 –4)), 152 –162. http://
dx.doi.org/10.1002/prca.201100050 .
Zambon, D., Quintana, M., Mata, P., Alonso, R., Benavent, J., Cruz-Sanchez, F., et al.
(2010). Higher incidence of mild cognitive impairment in familial hypercholester-
olemia. American Journal of Medicine, 123 (January (3)), 267 –274. http://dx.doi.org/
10.1016/j.amjmed.2009.08.015 .B. Szalkai et al.
Archives of Gerontology and Geriatrics 73 (2017) 300–307
307