You are on page 1of 10

Available online at www.sciencedirect.

com
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com

ScienceDirect
00 (2021) 000–000
Procedia Computer00 (2021)192
Science 000–000
(2021) 1170–1179 www.elsevier.com/locate/procedia
www.elsevier.com/locate/procedia

25th International Conference on Knowledge-Based and Intelligent Information & Engineering


25th International Conference on Knowledge-Based
Systems and Intelligent Information & Engineering
Systems
Probabilistic
Probabilistic causal
causal model
model for
for the
the detection
detection of
of obsolete
obsolete personal
personal
information to prevent falls in the elderly
information to prevent falls in the elderly
Salma Chaieba,d,∗ b,d c,d
a,d,∗, Ali Ben Mradb,d , Brahim Hnichc,d
Salma Chaieb
a
, Ali Ben Mrad , Brahim Hnich
University of Sousse, ISITCom, 4011, Sousse, Tunisia
a University of Sousse,
b University of Sfax,ISITCom, 4011,Sfax,
ISAAS, 1013, Sousse, Tunisia
Tunisia
b University of Sfax, ISAAS, 1013, Sfax, Tunisia
c University of Monastir, FSM, 5000, Monastir, Tunisia
c University of Monastir,
d University FSM, 5000, Monastir, Tunisia
of Sfax, CES Lab, 3038, Sfax, Tunisia
d University of Sfax, CES Lab, 3038, Sfax, Tunisia

Abstract
Abstract
The information describing a person is constantly evolving and may become obsolete and contradict other information. A per-
Thedatabase,
sonal information describing
therefore, musta be
person is constantly
consistently evolving
updated upon and
the may becomeofobsolete
acquisition andobservations
new valid contradict other
that information. A per-
contradict obsolete
sonal database, therefore, must be consistently updated upon the acquisition of new valid observations that
ones contained in the database. This study focuses on proposing a novel approach for dealing with the information obsolescence contradict obsolete
ones contained
problem in the database.
in the elderly-fall This study
prevention focuses
context. Our on proposing
approach aimsa to
novel approach for
continuously dealing
monitor with information
elderly the information obsolescence
in order to detect
problem
the changein in
thetheelderly-fall
behavior of prevention
an elderlycontext. Ourtoapproach
person and aimsfrom
prevent him to continuously monitor
falls. It consists elderly information
of detecting contradictions in between
order to newly
detect
the change in the behavior of an elderly person and to prevent him from falls. It consists of detecting contradictions
acquired information about a single elderly and what we already know about that person, then identifying among his observations between newly
acquired
those thatinformation
have become about a single
obsolete and elderly
need toand what weWe
be updated. already know
propose about
a new that person,
approximate then identifying amongwhich
concept,-Contradiction, his observations
represents
those that have level
the confidence become obsolete
of having and need to beinupdated.
a contradiction We propose awhen
a set of observations new approximate concept,-Contradiction,
a causal Bayesian which represents
network is our representation model.
the
We confidence level of having a algorithm
propose a polynomial-time contradiction in a set ofobsolete
for detecting observations when a causal
information and show Bayesian
that thenetwork is our
resulting representation
obsolete model.
observations are
We propose a polynomial-time
given as an explanation AND-OR tree. algorithm for detecting obsolete information and show that the resulting obsolete observations are
given as an explanation AND-OR tree.
© 2021 The
© 2021 The Authors.
Authors. Published
Published by
by Elsevier
Elsevier B.V.
B.V.
© 2021
This The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license
is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
(https://creativecommons.org/licenses/by-nc-nd/4.0)
This is an
Peer-reviewopen access
Peer-review under article under
underresponsibility the
responsibilityofofthe CC BY-NC-ND
thescientific license
scientificcommittee
committee ofof(http://creativecommons.org/licenses/by-nc-nd/4.0/)
the KESInternational.
KES International.
Peer-review under responsibility of the scientific committee of the KES International.
Keywords: Obsolete information; Contradictory observations; Causal Bayesian network; Reasoning with uncertainty; Elderly-fall prevention
Keywords: Obsolete information; Contradictory observations; Causal Bayesian network; Reasoning with uncertainty; Elderly-fall prevention

1. Introduction
1. Introduction
Falls are common among older adults around the world and are the leading cause of traumatic death in this popula-
Falls
tion. are one-third
About common among older adults
of this senior aroundfalls
population theeach
worldyear,
and and
are the
the leading cause
risk of falls of traumatic
increases death in thiswith
proportionately popula-
age.
tion. About one-third of this senior population falls each year, and the risk of falls increases proportionately with
At 80 years old, over half of seniors fall annually. Falls, with or without injury, also have a heavy quality of life impact.age.
At 80 years old, over half of seniors fall annually. Falls, with or without injury, also have a heavy quality of life
A growing number of elderly fear falling and, as a result, limit their activities and social engagements. These impacts impact.
A growing number of elderly fear falling and, as a result, limit their activities and social engagements. These impacts

∗ Salma Chaieb. Tel.: +216 53 950 523.


∗ Salma
E-mailChaieb.
address:Tel.: +216 53 950 523.
salma.chaieb2@yahoo.com
E-mail address: salma.chaieb2@yahoo.com
1877-0509 © 2021 The Authors. Published by Elsevier B.V.
1877-0509
This © 2021
is an open Thearticle
access Authors. Published
under by Elsevier B.V.
the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
1877-0509
This © 2021
is an open Thearticle
access Authors. Published
under by Elsevier B.V.
the scientific
CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer-review under
This is an open responsibility
access of the
article under committee
the CC BY-NC-ND oflicense
the KES(https://creativecommons.org/licenses/by-nc-nd/4.0)
International.
Peer-review under responsibility of the scientific committee of the KES International.
Peer-review under responsibility of the scientific committee of KES International.
10.1016/j.procs.2021.08.120
Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179 1171
S. Chaieb et al. / 00 (2021) 000–000

also lead to enormous economic losses. Consequently, in order to achieve a good degree of fall prevention among
older adults, the development of automatic monitoring and fall detection systems that call for help from caregivers
such as lifestyle guides, clinical-community partnerships, etc. becomes a serious necessity. In this respect, several fall
prevention strategies and tools are being developed and tested involving treating physicians, other health care team
members, patients and some of their family members [1, 10, 19, 11]. To ensure the best service, these systems require
consistent and up-to-date personal information about the elderly collected from hospitals, mutual health, and other
associations and organizations in caring for the elderly. Information describing health status, profile, activities and
conditions of a person is usually provided by widely distributed sources and is often uncertain and unreliable. Further-
more, it is constantly evolving over time and may become obsolete with the arrival of new information (supposed to
be certain) that contradicts it with respect to a representation model. For example, consider an older adult about whom
you know the following information: He is in good health and goes shopping regularly. Now, you learn that he had a
violent car accident in which he lost his sight and his autonomy (new event supposed to be certain). As a blind person
with loss of autonomy cannot shop regularly (general knowledge), the new piece of information contradicts the older
one (he goes shopping regularly) and leads to revise it since it’s clearly obsolete.
The objective of our work is to continuously control and monitor personal information of the older adults in real
time in order to detect the change in the behavior of an elderly person. To do this, we propose an Obsolete Information
Detection System (OIDS) that proceeds as follows: when a new observation about an older adult is recorded, our OIDS
checks if this observation contradicts what we already know about that person stored in the database row reserved for
him. If so, it identifies among these observations those that are responsible for this contradiction, i.e., that have become
obsolete and need to be updated, while keeping the new acquired one. The general aim of the system is to provide
a consistent and updated elderly database. For treating physicians, having such an information base for each of their
patients, in particular for elderly patients, can contribute to the improvement of falls prevention as well as various
aging-related diseases. It is, in fact, an innovative and effective way to help doctors follow their patients by providing
them with some information (or predictive values in case of lack of information) at the right time for the target person.
In this work, we choose to use a Causal Bayesian network (CBN) [12, 14] to represent our knowledge. Several facts
justify this choice: first, our approach is primarily based on studying dependencies and causal relationships between
the given observations to provide us with an understanding, comprehension, and insight into the occurrence of some
events, namely the contradiction. Indeed, it explains the causes of the occurrence of a contradiction between given
observations (why do we have a contradiction?), as well as the different alternatives to remove this contradiction (what
are the possible observations responsible for contradiction? What if we update the values of some observations?). So,
CBN represents a flexible useful tool in this respect as it can be used to formalize, measure, and deal with different
unfairness (contradictory) scenarios underlying a dataset. Second, we work within an uncertain environment char-
acterized by the lack of information as we may not have observations on all the characteristics of an elderly person.
CBNs are thus a powerful probabilistic model that allows reasoning with incomplete data. Third, in real life, the newly
acquired information may be uncertain, contrary to what we have supposed. So CBN proves to be very relevant for
managing such a type of information. Our OIDS is practical since its time complexity is quadratic in the number of
CBN variables. The results of our approach are very accurate and refined– given as an explanation AND-OR tree that
encodes all possible obsolete observations ordered in a particular way.
A closely related area is belief revision theory (BRT) which is a classical topic for knowledge bases in both the
logical case and the probabilistic one [2, 7]. This research area deals with the changing of information stored in
knowledge bases when they are contradicted by newly arriving information. Although similar in some ways, our
approach does not address the same problem as defined in BRT. In BRT, revising the knowledge, both in a logical
and probabilistic contexts, means changing the representation model in response to changes in the domain. However,
in our approach, the general knowledge is supposed to be stable and is in the form of a CBN that describes relations
between some characteristics of a given subject.
We augmented our discussion with an interesting case study: the elderly-fall prevention, which we would work on
it during different sections of this paper to elaborate the concepts better.
The rest of the paper is organized as follows. In Section 2, we describe the construction details of the CBN for
preventing falls in older adults. In Section 3, we present our theoretical contribution, which consists of proposing a
new concept regarding -Contradiction among observations, studying its theoretical properties, giving the obsolete
1172 Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179
S. Chaieb et al. / 00 (2021) 000–000

information detection algorithm, and discussing its complexity. In Section 4, we propose a practical solution method
for the obsolescence problem. Finally, we conclude in Section 5 and outline our future directions.

2. Causal Bayesian network for Elderly Fall-Prevention


2.1. CBN background
In this work, we use CBN to represent our information base. Causal Bayesian network [12, 14] used to model causal
relationships is a graphical model which consists of a model structure and a set of parameters. CBN = (G, Θ), where
G = (X, E) is a directed acyclic graph with nodes X = {X1 , ..., Xn } corresponding to random variables representing the
elderly characteristics, and directed edges E that connect these nodes and correspond to causal relationships among
them. The direct causes of a variable Xi are the variables that will change the distribution of Xi as we vary them,
as we perfectly intervene in the whole system. Indeed, given two variables X1 and X2 , we say that X1 precedes X2
causally if experimental interventions that change the value of X1 can affect the distribution of X2 but not vice versa.
We use capital letters such as X1 , X2 , X3 to denote variable names and lowercase letters such as x1,1 , x2,1 , x3,2 for
specific values, also called observations, taken by those variables. Sets are denoted by boldface capital letters. The
second component of the couple, namely Θ, represents a set of parameters that quantifies the network and are stored
as a conditional probability table (CPT) for each variable. The joint distribution over X, given by the chain rule (1), is
defined by the product of the local probability distributions on each variable Xi , P(Xi | Pa(Xi )), where Pa(Xi ) denotes
the set of the parents of the node Xi in the graph.
n

P(X) = P(Xi | Pa(Xi )) (1)
i=1

For more details about CBNs, the reader can refer to the book of Causality edited by Pearl [13].

2.2. Model structure and parameters

As part of the Elderly-Fall Prevention project, we have access to a real-life database Elderly-Data that contains
information on the elderly. It is collected during the elderly appointments with their attending physician in both the
University Hospital Center of Lille and Valenciennes over a 9-year period (2005-2014). The Elderly-Data includes
about 1174 patient records, each of these records was described by 207 patient-history features (binary, denoting
presence or absence of a feature or continuous, expressing the value of a feature). The construction of the CBN
structure is conducted using the following steps: first, by reviewing the medical literature, health statistic reports,
and by analyzing the database Elderly-Data, we identified 25 features out of 207 that represent the most significant
influences of falls among elderly. In [3], authors have reported that targeted risk factors that may increase elderly
falls were postural hypotension, polypharmacy (4+ prescriptions), sensory impairments (eyes/ears/feet), and mobility
(balance/gait) issues. In [15], authors have reported several other risk factors that may increase elderly falls and
that include environmental risk, and various health conditions including muscle weakness, cognitive impairments,
diabetes, hypotension, and osteoporosis. Recent epidemiological findings of associations between hearing capacity
and motor functions have proven the strong association between hearing loss, poor balance, and falls [5]. Many
patients experience difficulty with movement, including balance issues and gait disturbances, after stroke [6]. Stroke
may also result in decreased truck muscle strength and limited trunk coordination, frequently determining loss of
autonomy due to the trunk impairment [17]. Factors associated with increased stroke risk in multivariate analysis
included age, diabetes, and cardiovascular disease [16]. Whenever a directed edge is added between two variables, we
ask for expert assistance. A total of 3 experts were interviewed to validate the structure and evaluate the relationships
between the nodes. We estimate that elicitation of the structure took about 16 hours with the experts. This includes
model refinement sessions, where the previously obtained structure was reassessed in a group setting. The structure of
our model is shown in Fig. 1. We believe that it models reasonably causal interactions among the selected variables.
In order to draw useful information from the resulting CBN, knowledge of the joint probability is required. The
information in CPTs may be obtained from empirical data, direct measurements, related model outputs, or expert
opinion under no other data available. In our work, we used available statistics, bibliographic searches, observational
S. Chaieb
Salma Chaieb et al. et al. / 00 (2021)
/ Procedia 000–000
Computer Science 192 (2021) 1170–1179 1173

age diabetes

hearingDeficit heartDisease

cardiovascularDrugs gender

visionDeficit strokeTIA

muscleImpairment drugsNb

hypotension dementia

balanceDisorder autonomyLoss

gaitDisorder driveCar environmentalRisk

doShopping livesAlone fallsNb

leaveHome fearFalling telealarm fracture nursingHome

Fig. 1. CBN for Elderly Fall-Prevention.

data when it is possible, and solicit the experts to estimate the parameters of the model. Although CBN formal-
ism allows both discrete and continuous variables, all exact general-purpose algorithms for CBNs deal with models
containing only discrete variables. In order to take advantage of these algorithms, we started by discretizing the con-
tinuous variables, namely age, drugsNb, and fallsNb, while accepting the minimum loss of variable information. Our
discretization is based on method of Equal Frequency Discretization [9, 18].
Prior probability distributions are simply relative counts of various outcomes for each of the variables gender,
age, smoking, and environmentalRisk. However, although the prior probabilities can be learned reasonably accurately
from the database of hundreds of records, the conditional probabilities present more of a challenge. In cases where
several variables directly precede a variable in question, individual combinations of their values may be very unlikely
to the point of being absent from the data file. In these cases, and given the lack of observational data, we solicited
experts to define and validate the probability distribution for the rest of the CBN variables, except for the two variables
driveCar and fallsNb. Thus, for each of these variables, the simple average of distributions from each expert was used
to produce the corresponding CPT for the final model. For the two variables driveCar and fallsNb and since they have
6 parents, the definition of their probability distribution by experts is a bit complicated. Thus, their CPTs are learned
from the observational data and then checked by the experts.
In this work, we consider the CBN as a stable component and its updating is not part of the objective of this paper.
The CBN is passed as a parameter to our OIDS. It is used only to represent causal influences between observations
and to compute the conditional probability of one node, given values assigned to the other attributes.

3. Theoretical results
We denote by N the number of CBN’s variables and by E the number of CBN’s edges. We use onew and Onew to
distinguish a new observation onew on a variable Onew that may already be observed or not. We define two sets, which
we will use extensively throughout this paper:
OBS = {(Xi , xi, j ), i ∈ I ⊆ {1, .., N}, xi, j ∈ D(Xi )} the set of pairs representing a variable and its observed value,
relating to a single individual of the database, OBS’ = OBS \ (Onew , onew ). When there was a previous observation on
1174 Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179
S. Chaieb et al. / 00 (2021) 000–000

Onew , the new one replaces it, meaning that (Onew , onew ) replaces the previous element related to Onew in OBS. When
no previous observation on the variable Onew was present in OBS, the element (Onew , onew ) is added to the set OBS.
Theoretically, a contradiction between observations occurs when the conditional probability of the new observation
given other observations is zero: P((Onew = onew )|OBS’) = 0. However, since the CBN is a probabilistic knowledge
representation model, there is always a degree of uncertainty related to the inference we draw from it. Hence, in
practice, we shall not expect the conditional probabilities to be exactly zero but rather very close to 0. For that reason,
we introduce a contradiction probability tolerance value  to reflect those uncertainties related to the probabilistic
dependencies among the variables in the CBN. The approximate contradiction is thus defined as follows:
Definition 1 (-Contradiction). Given a CBN, a set of observed variables OBS’, a new observation onew on a variable
Onew , and a real number 0 ≤  ≤ 1. OBS’ is -Contradictory to onew when P(Onew = onew |OBS’) ≤ .
We refer to a set of all contradictory observations to onew as Sonew and show how to restrict this set to a smaller
subset of obsolete observations. At this stage, we can define a set of obsolete observations as follows:
Definition 2 (Set of obsolete observations). Given a CBN, a new observation onew on Onew and a set of observed
variables OBS’ that is -Contradictory to onew , a set Sonew ⊆ OBS’ of obsolete observations is an -Contradictory set
to onew that is minimal and such that the withdrawal of Sonew from OBS’ restores its consistency with onew .
The obsolete information detection approach is based on this fundamental definition. Indeed, an -Contradiction
occurs when there is a subset Sonew of OBS’ of observations that have become obsolete and contradict onew . The process
of identifying the set of obsolete observations Sonew takes place in 3 steps: (1) restrict the -Contradictory set OBS’ into
Sonew ; (2) decompose the -Contradictory set Sonew looking for obsolete observations; and (3) compose the AND-OR
tree from the set of obsolete observations Sonew .

3.1. Restricting the -Contradictory Set


When an information base is inconsistent, it does not mean that all observations therein are contradictory. Let us
consider an elderly about whom we know the following information: he is a 60-year-old man, he has no illness that
somehow causes a motor disorder, and he has good visual acuity. A moment later, we receive new information about
this person indicating that he has a severe gait disorder. This new information contradicts what we know about this
elderly person, but, it is clear that the contradiction occurs regardless of the elderly’s sex. Hence, information about
the elderly’s sex is not a necessary part of the set Sonew . The set Sonew of obsolete observations can be restricted as
follows: Sonew ⊆ OBS’ such that there exists an active path between Onew and each element of Sonew . The active path
is defined based on the concept of d-separation in CBNs [8]. In the particular case where all nodes of the Markov
Blanket [12], MB(Onew ), of the node Onew are observed, Onew becomes independent of any other node according to
the Bayes assumption. The gray area in Fig. 2 represents the MB(Onew ). Thus the set Sonew can be restricted as follows:
Sonew ⊆ OBS’ ∩ MB(Onew ). We conclude this section with the following proposition:
Proposition 1. Given a CBN, a new observation onew on a variable Onew , a set OBS’ that is -Contradictory to onew
and a non-empty subset Sonew ⊆ OBS’ such that Sonew contains all the nodes of OBS’ for which there exists an active
path with Onew , the following statements are true:
(1) Any observation of Sonew may belong to the set of obsolete observations after acquiring onew ; and
(2) Any observation of OBS’ \ Sonew does not belong to the set of obsolete observations after acquiring onew .
Proof. This proposition comes from the fact that (1) any variable X in Sonew is dependent on Onew through the existence
of an active path from X to Onew and hence may be obsolete; (2) any variable in OBS’ \ Sonew is independent of Onew
by using the notion of active path and hence cannot be obsolete.

3.2. Decompose the -Contradictory Set


Up till now, we have shown that OBS’ can be restricted to Onew ’s dependent variables in Sonew . As a first step, we
can decompose this set further into subsets, Si , by bringing the observations on dependent variables together.
Depending on the causal structure of the CBN, the subset Si can take one of the following three forms shown in
Fig. 2: (1) Si contains all the causes of the common effect Onew (2) Si only contains an effect of Onew , when there are
no other causes of this effect (3) Si contains an effect of Onew and the other variables that cause this effect if they exist.
Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179 1175
S. Chaieb et al. / 00 (2021) 000–000

Fig. 2. An example of decomposition.

Then, we explore all subsets Si resulting from the decomposition, and for each one, we check whether it is consis-
tent with onew or not. If so, it will be ignored and will not be considered in the rest of the process. So the result of the
decomposition phase is a set Sd ={S1 , S2 , ..., Sk } of subsets, each containing observations and their related variables.
Each subset obeys these two decomposition properties:
Property 1. Each subset Si resulting from the decomposition phase contains dependent variables.
Property 2. Each subset Si resulting from the decomposition phase is -Contradictory to onew , given the CBN.
At this point, the main obsolete information detection process will revolve around each -Contradictory subset Si .
Following the example shown in Fig. 2, and respecting the 3 possible forms of the subset Si , the observations into Si
are organized so that in case of contradiction, three explanations are possible:
Explanation 1: in the case where Si contains all causes of onew (group 1 in Fig. 2), at least x1,1 or x2,1 or both
cannot cause onew . In such a case, we distinguish two aspects of each observation xi, j contained in the set Si . Either xi, j
is individually -Contradictory to onew (i.e., P(onew |xi, j ) ≤ ) or not. Thereby, we can decompose Si further into two
disjoint -Contradictory subsets:

• The AND-Set, SiAND , containing each observation in Si that is individually -Contradictory to onew , given the
CBN.
• The OR-Set, SOR
i , containing each observation in Si that is not individually -Contradictory to onew , given the
CBN, but the entire OR-Set is -Contradictory to onew , given the CBN.

All the observations in the AND-Set are obsolete and need to be updated, which is not always the case for those in
the OR-Set. Let’s explain this tricky situation with the following example: if we suppose that leaving home (Onew ) is
conditioned by both events: driving a car (X1 ) or shopping (X2 ). We know that an older adult does not drive his car (X1 ,
no) and does not go shopping (X2 , no). A new observation given by an external sensor indicating that he has left his
house (Onew , yes). Certainly, this new observation raises a contradiction with what we already know. However, if we
take a closer look at each of these observations separately, we find that each of them is not individually -Contradictory
to (Onew , yes). Indeed, since these observations are conditionally dependent, P(Onew = yes|X2 = no) >  comes from
the fact that the elderly may have driven her car. The same applies to P(Onew = yes|X1 = no) > . So, updating one
of these two old observations is enough to remove the -Contradiction. However, with the available knowledge at our
disposal, we are not able to accurately infer which one(s) should be updated. Consider this scenario now replacing the
observations (no, no, yes) respectively on the variables (X1 , X2 , Onew ) with (yes, yes, no). The observations on the two
variables X1 and X2 are individually -Contradictory to (Onew , no) and will be placed into the AND-Set.
Explanation 2: in the case where Si contains the effect of onew (group 2 in Fig. 2), the contradiction is explained
by the fact that onew cannot cause x3,2 , i.e., that x3,2 is individually -Contradictory to onew and is therefore classified
in an AND-Set of Si .
1176 Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179
S. Chaieb et al. / 00 (2021) 000–000

Explanation 3: in the case where Si contains the effect of onew and its other causes (group 3 in Fig. 2), the con-
tradiction is explained by the fact that at least onew or x5,2 or both cannot cause x4,1 . As we previously assumed that
the newly acquired observation onew is certain, meaning it cannot be objected, then the contradiction is explained by
the fact that (1) either x5,2 cannot cause x4,1 given onew and then need to be updated, or x4,1 cannot be the effect of
both x5,2 and onew and then need to be updated. In this case, both x5,2 and x4,1 are not individually -Contradictory
to onew and then are classified in an OR-Set of Si . (2) x4,1 cannot be the effect of onew independently of x5,2 , meaning
that x4,1 is individually -Contradictory to onew and then is classified in an AND-Set. An example of (1) is (Onew , X4 ,
X5 ) = (leaveHome, NbOfExitsGPS, GPSState), (onew , x4,1 , x5,2 ) = (yes, 0, OK). An example of (2) is (Onew , X4 , X5 ) =
(diabetes, drugsNb, cardiovascularDrugs), (onew , x4,1 , x5,2 ) = (yes, 0, no).

3.3. Composition of the explanation AND-OR tree

As a result of the decomposition phase, we obtain the set Sonew of obsolete observations, which satisfies the propo-
sition 1. The main aim of the current phase is to combine the results of the subsets of Sonew to create an explanation
AND-OR tree whose internal nodes are labeled either AND or OR and whose leaves represent all possible obsolete
observations contained in each subset SiAND and SORi .
The explanation AND-OR tree is constructed as follows. We introduce a root node labeled AND. Then, for each
subset Si , we introduce an AND node whose parent is the root node. Next, for each AND-Set (resp. OR-Set) of Si ,
we introduce an AND (resp. OR) node whose parent is the corresponding node of Si and a child leaf node for each
observation in SiAND (resp. SOR
i ). Each leaf node is labeled with the obsolete observation. Fig. 3 shows the hierarchical
structure of the resulted explanation AND-OR tree.

Fig. 3. Hierarchical structure of the explanation AND-OR tree.

The AND-OR tree represents precisely the set of obsolete observations Sonew and explains the logical relationships
among its AND-Sets and OR-Sets. Indeed, given a new observation onew and an AND-OR tree T that represents the
set of obsolete observations relative to onew , the following three propositions are true:
Proposition 2. For each observations x ∈ OBS’, x  T if and only if x is not obsolete.
Proof. ”=⇒” if x  T then x is not part of the set of obsolete observations since T reflects this set. So x is not
obsolete.
”⇐=” x is not obsolete means that x is not part of the set of obsolete observations and since T represents this set then
x  T.
Proposition 3. All observations of the AND-Set are obsolete.
Proof. Follows immediately from the fact that each observation in the AND-Set is individually -Contradictory to the
new observation given the CBN.
Proposition 4. At least one observation of the OR-Set is obsolete.
Proof. Follows immediately from: (1) the fact that none of the observations is individually -Contradictory to the
new observation given the CBN; (2) the OR-Set is -Contradictory to the new observation; and (3) removing any
observation from the OR-Set restores its consistency with onew , but we can not know exactly which one is obsolete.
Thus, each of the observations contained in the OR-Set is likely to be involved in the contradiction
Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179 1177
S. Chaieb et al. / 00 (2021) 000–000

3.4. Obsolete Information Detection Algorithm

Algorithm 1 Obsolete Information Detection Algorithm (OIDA)


Input: (Onew , onew ), B
Output: AND-OR Tree
Parameters: : a real number, 0 ≤  ≤ 1
1: let OBS’ be the set of observations and their associated variables in B except (Onew , onew ).
2: Sonew = Prune(OBS’, Onew , B).
3: if I sContradictory(Sonew , (Onew , onew ), B, ) then
4: Sonew = Decompose(Sonew , B, ).
5: let Sonew = {S1 , ..., Si , ..., S p } such that each Si is -Contradictory to onew given B.
6: AND-OR-Tree ← Compose(S1 ) ∧ ... ∧ Compose(Si ) ∧ ... ∧ Compose(S p ).
7: return AND-OR-Tree
8: else
9: return T rue
10: end if

We now define the main steps of Algorithm 1 for detecting obsolete information. As a first step onew arrives. In
a second step, our algorithm tries to find the candidate variables that may be involved in case of contradiction by
computing Sonew using the function Prune (line 2 in Algorithm 1). This function takes as input the set OBS’ and
restricts it to Sonew . So, instead of processing the entire CBN, we are only interested in variables that depend on Onew ,
which makes a considerable time gain especially since the Prune function is linear in the number of directed edges
and nodes in the CBN [4]. Then, OIDA checks if there is an -Contradiction between onew and Sonew (line 3). If so,
then our system has to look for obsolete observations. Indeed, considering the dependency relations between the CBN
variables, we decompose Sonew (line 4) as explained in Section 3.2 into subsets Si of dependent variables. Then, this
function checks the consistency of each observation given by the set Si to onew and places it appropriately in either
the AND-Set or the OR-Set so as to obtain the new decomposed set Sonew . The Decompose function takes O(E × N 2 ).
Line 6 of OIDA traverses all elements of Sonew and for each one we call the main function : Compose(Si ). As we have
explained in Section 3.3, the Compose function takes as input the set Si subdivided into SiAND and SOR i , and returns
a sub-tree of obsolete observations relating to each Si . This function takes O(N s2 ) time complexity where N s is the
number of observations in Si . The result of OIDA is an AND-OR Tree of all possible obsolete observations.
The OIDA runs in O(E × N 2 ) where N is the number of CBN variables and E is the number of edges in the CBN.
The algorithm is sound and complete.
Proof. Soundness. follows from proposition 3 since each observation individually -Contradictory to onew , given the
CBN, must be updated and from proposition 4 since at least one observation of the OR-Set must be updated.
Proof. Completeness. follows from proposition 1 since obsolete observations are found only among Onew dependent
variables.

4. Experimental results
In this section, we validate our OIDS on a case study related to the prevention of falls among older adults. Our
experiments are carried out using a set of 224 scenarios with 112 contradictory scenarios and 112 non-contradictory
scenarios, generated from the CBN and validated by two medical experts, an orthopedist and a neurologist, who were
not involved in the construction and the validation of the CBN. The generation of scenarios is done using an automatic
generation process, which consists of (1) randomly selecting variables from the given CBN, (2) assigning random
observations to the selected variables, and (3) arbitrarily choosing a pair (variable, observation) that represents the
newly acquired information. Thus each scenario Ci is represented by the pair of newly acquired information (variable,
new observation) accompanied by a sequence of pairs of some previously acquired observations that are consistent
(variable, observed value). For each scenario, Ci , a label ci given by the experts is associated such as: ci = 1 if Ci is
declared contradictory by the experts, ci = 0 otherwise. Then, for each scenario labeled as contradictory, the experts
provided a list of subsets of all possible obsolete observations, such that the withdrawal of some observations of these
1178 Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179
S. Chaieb et al. / 00 (2021) 000–000

subsets restores the consistency of the remaining observations with the newly acquired one. The resulting subsets lists
were then organized into AND-OR trees.
Table 1. An example of a contradictory scenario.

Onew = onew OBS’ Dependency P(Onew |Xi ) P(Onew |OBS’) AND-OR tree Execution time

age = 60 direct cause 0.4993


gaitDisorder = no direct cause 0.0443
{gaitDisorder ∨
gender = female d-separated —
balanceDisorder ∨
balanceDisorder = no direct cause 0.1756
hypotension ∨
hypotension = no direct cause 0.2200
environmentalRisk
fallsNb ≥ 7 environmentalRisk = no direct cause 0.3585 1.247e−05 8.96 ms
∨ strokeTIA} ∧
strokeTIA = no direct cause 0.4575
{livesAlone ∨
livesAlone = yes V-structure 0.4109
nursingHome} ∧
hearingPb = no d-separated —
{fearFalling}
nursingHome = no direct effect 0.3211
fearFalling = no direct effect 0.0039

The validation of our OIDS consists of two parts: measuring the accuracy of the system in detecting contradictions
and evaluating the quality of the AND-OR trees resulting from our OIDS. The first part consists of measuring the
false positive and the false negative rates of our system. The second part is to compare the AND-OR trees given by
our system with the results given by the experts. The comparison is made in two levels. At the first level, we check the
number of AND nodes (the subsets Si ) that appear in level 2 of each tree as shown in Fig. 3. Then, for each Si node, we
compare the number of leaves that compose each of the SiAND and the SOR i nodes with those provided by the experts.
We have requested the assistance of domain experts who validated the CBN to adjust the threshold value required
to detect contradictions. They evaluated the resulting marginal distributions of the CBN and how the distributions
change after observing certain variables and provided us with the most suitable threshold value, which is set at 1e−2.
We apply our OIDA on the 224 scenarios. The results are summarized as follow: out of the 112 scenarios labeled
by the expert as contradictory, 104 scenarios are estimated with our algorithm as -Contradictory, and out of the 112
scenarios labeled by the expert as non-contradictory, 97 are considered -non-contradictory by our algorithm.
Then, for the 104 AND-OR trees resulted from our systems which are associated with the scenarios classified
as -Contradictory by our system, the resulting explanation AND-OR trees are perfectly in line with those provided
by the experts. The credibility and accuracy of the resulting trees are theoretically justified by the propositions and
properties given in Section 3. Owing to space limitations, we cannot display all the results issuing from this step.
However, an example that was treated by our algorithm is shown in Table 1. Column 1 (resp. 2) of the table given
in Table 1 contains the newly (resp. previously) acquired observation(s). Column 3 refers to Section 3.1 in which
we seek to find among the observations previously acquired those that depend on Onew . The conditional probability
given in column 5 means that the scenario is contradictory since the resulting value ≤ . As explained in Section
3.2, the decomposition phase gives the set Sonew = {{(age, 60), (gaitDisorder, no), (hypotension,no), (balanceDisorder,
no), (strokeTIA, no), (environmentalRisk, no)}, {(livesAlone, yes), (nursingHome, no)}, { (fearFalling, no)} To find out
which observations in Sonew are part of the AND-Set and the OR-Set, we calculate the probability of Onew given Xi for
each Xi ∈ Sonew dependent on Onew as shown in column 4 of Table 1. In our example, we notice that the observation
( f earFalling, no) is individually -Contradictory to f allsNb ≥ 7 since its associated probability is less than .
Each observations in the first (resp. second) subset, S 1 (resp. S 2 ), of Sonew is individually -non-contradictory, but
the whole subset {(age, 60), (gaitDisorder, no), (hypotension,no), (balanceDisorder, no), (strokeTIA, no), (environ-
mentalRisk, no)} (resp. {(livesAlone, yes), (nursingHome, no) } is -Contradictory to f allsNb ≥ 7. As explained in
Section 3.2, we proceed by elimination. This process leads to ignore (age, 60) from S 1 since the withdrawal of this
observation does not restore the consistency of S 1 with onew . At the end of treatment and as shown in column 7
of Table 1, the OIDA returns the explanation AND-OR tree of all possible obsolete observations. It means that to
integrate onew , we must check the observation on the variable fearFalling, verify either (gaitDisorder, no) or (hypoten-
sion,no) or (balanceDisorder, no) or (strokeTIA, no) or (environmentalRisk, no), and check either (livesAlone, yes) or
(nursingHome, no) depending on the user’s choice. The average execution time of a scenario is about 10ms.
One possible use of the resulting AND-OR trees is to generate a list of relevant questions for physicians (or related
personnel) in order to update values of obsolete observations and get the newer ones if they exist. Furthermore, we
Salma Chaieb et al. / Procedia Computer Science 192 (2021) 1170–1179 1179
S. Chaieb et al. / 00 (2021) 000–000

can suggest, among the observations contained in the OR-Sets, the most likely to be updated based on some priority
measures. In addition to this, we can help the target user make the right decisions by suggesting the most likely values
which can replace the obsolete ones with some prediction precision. This may be of great interest in further research.

5. Conclusion
This study focuses on managing a set of elderly patient-specific observations in complement to the generic knowl-
edge embedded in a CBN, taking into account the fact that some observations may have become obsolete over time
and/or contradict other information. Our contribution includes (1) the proposal and validation of a 25-node probabilis-
tic causal model representing some elderly features (2) the definition of a new concept -Contradiction required for
detecting contradictory scenarios underlying a dataset (3) the proposal of an entirely new causal-based approach to
accurately identify all possible obsolete observations. Our approach runs in a polynomial time and returns results in
an original way, in the form of an explanation AND-OR tree. It encodes all possible subsets of obsolete observations
and can be effectively used to maintain a consistent database. Our approach efficiency is confirmed experimentally
since our simulations on a real-life database in the elderly fall prevention context are very encouraging, reaching an
accuracy of 89%. In the future, we will try to answer the following questions: how the resulting trees can be used to
give reliable recommendations to restore the database consistency, and through which strategy the observations of the
OR-Sets will be removed? In this paper, and for reasons of clarity, we present our model assuming that our CBN is
perfect and fixed during processing, that newly acquired observations are certain and consistent. However, this may
not be in real life and thus we will try to extend our model to handle such practical situations.

References

[1] Aidemark, J., Askenäs, L., 2019. Fall prevention as personal learning and changing behaviors: Systems and technologies. Procedia Computer
Science 164, 498–507.
[2] Alchourrón, C.E., Gärdenfors, P., Makinson, D., 1985. On the logic of theory change: Partial meet contraction and revision functions. The
journal of symbolic logic 50, 510–530.
[3] Baker, D.I., Leo-Summers, L., Murphy, T.E., Katz, B., Capobianco, B.A., 2019. Intervention to prevent falls: community-based clinics. Journal
of applied gerontology 38, 999–1010.
[4] Butz, C.J., dos Santos, A.E., Oliveira, J.S., 2016. Relevant path separation: A faster method for testing independencies in bayesian networks,
in: Conference on Probabilistic Graphical Models, pp. 74–85.
[5] Carpenter, M.G., Campos, J.L., 2020. The effects of hearing loss on balance: A critical review. Ear and Hearing 41, 107S–119S.
[6] Choi, S.h., Lim, C.g., 2020. Immediate effects of ankle non-elastic taping on balance and gait ability in patients with chronic stroke: A
randomized, controlled trial. Journal of Manipulative and Physiological Therapeutics 43, 922–929.
[7] Gärdenfors, P., Rott, H., Gabbay, D., Hogger, C., Robinson, J., 1995. Belief revision. Computational Complexity 63.
[8] Geiger, D., Verma, T., Pearl, J., 1990. d-separation: From theorems to algorithms, in: Machine Intelligence and Pattern Recognition. Elsevier.
volume 10, pp. 139–148.
[9] Jiang, S.y., Li, X., Zheng, Q., Wang, L.x., 2009. Approximate equal frequency discretization method, in: 2009 WRI Global Congress on
Intelligent Systems, IEEE. pp. 514–518.
[10] Mostafa, S.A., Mustapha, A., Mohammed, M.A., Ahmad, M.S., Mahmoud, M.A., 2018. A fuzzy logic control in adjustable autonomy of a
multi-agent system for an automated elderly movement monitoring application. International journal of medical informatics 112, 173–184.
[11] Mozaffari, N., Rezazadeh, J., Farahbakhsh, R., Yazdani, S., Sandrasegaran, K., 2019. Practical fall detection based on iot technologies: A
survey. Internet of Things 8, 100124.
[12] Pearl, J., 1988. Bayesian inference. Probabilistic reasoning in intelligent systems: networks of plausible inference. 2nd ed. San Francisco:
Morgan Kaufmann Publisher , 29–75.
[13] Pearl, J., 2009. Causality. Cambridge university press.
[14] Pearl, J., Mackenzie, D., 2018. The book of why: the new science of cause and effect. Basic books.
[15] Sharif, S.I., Al-Harbi, A.B., Al-Shihabi, A.M., Al-Daour, D.S., Sharif, R.S., 2018. Falls in the elderly: assessment of prevalence and risk
factors. Pharmacy Practice (Granada) 16.
[16] Sharma, M., Hart, R.G., Connolly, S.J., Bosch, J., Shestakovska, O., Ng, K.K., Catanese, L., Keltai, K., Aboyans, V., Alings, M., et al., 2019.
Stroke outcomes in the compass trial. Circulation 139, 1134–1145.
[17] Sorrentino, G., Sale, P., Solaro, C., Rabini, A., Cerri, C.G., Ferriero, G., 2018. Clinical measurement tools to assess trunk performance after
stroke: a systematic review. European journal of physical and rehabilitation medicine 54, 772–784.
[18] Tsai, C.F., Chen, Y.C., 2019. The optimal combination of feature selection and data discretization: An empirical study. Information Sciences
505, 282–293.
[19] Yacchirema, D., de Puga, J.S., Palau, C., Esteve, M., 2018. Fall detection system for elderly people using iot and big data. Procedia computer
science 130, 603–610.

You might also like