Monitoring the activities and routines of the elderly requires continuous work and efort on the part of the care staf. It is vital to follow up on each user since abnormal behaviour could reveal the presence of a medical condition and, in these cases, early intervention is of utmost importance. In this paper, we deploy a room-level symbolic location system based on Bluetooth Low Energy (BLE) technology and present the results obtained from applying several methods capable of determining daily and weekly habits, non-invasive, using association rules and decision trees.
PID2021-122642OBPID2021-122642OB
In today’s age of technology, monitoring and tracking users’ daily behaviours and activities
are of great importance for medical professionals and care staf. In particular, monitoring the
habits of the elderly is crucial to identify any changes in their behaviour and prevent possible
diseases [
This work introduces a novel approach focused on this methodology using association rules and decision trees that allow an automatic and non-invasive determination of the daily and weekly habits of each user. In this study, a set of emitting beacons using BLE technology is used, distributed in the diferent rooms where the presence of people needs to be detected. Subsequently, localisation algorithms are used to obtain the symbolic location of the participants. The information obtained is acquired through the smart watches worn by the participants in this trial.
To evaluate the proposal presented in this work, we use data from a nursing home, where a
set of residents wore a smartwatch throughout the study[
The rest of the article is organised as follows: Section II provides an exploration of the prior research and related work; Section III presents the methods used in this work and the proposed methodology; Section IV deals with the data used, the experimental phase and the results obtained; finally, Section V describes the conclusions reached and future research lines.
Numerous works and approaches have utilized BLE (Bluetooth Low Energy) for various
applications, such as signal intensity monitoring [
Regarding the work presented in [
On the other hand, the authors in [
Finally, in the work presented in [
Our proposed approach leverages BLE technology to estimate room-level positions of elderly individuals residing in a nursing home. The primary goal is to discern and illustrate weeklylevel routines in a more intelligible and visually descriptive manner, employing logical rules or ifrst-order axioms. To achieve this, we utilize association rule algorithms such as Apriori and decision trees to generate these descriptive rules
This project was initiated in the year 2020 within the following publication [
The foremost advantages of utilizing BLE technology over WiFi are as follows: 1. Enhanced battery life: BLE technology ofers superior battery longevity compared to WiFi, with an approximate duration of 2-3 months before requiring a recharge or replacement. 2. Flexible beacon placement: Unlike WiFi routers, BLE beacons can be positioned freely without restrictions, allowing for more versatile and adaptable deployment options In our approach, we employ a passive beacon methodology wherein the smartwatch autonomously captures the Received Signal Strength Indication (RSSI) of the beacons
Some indoor positioning algorithms aim at obtaining the symbolic location of a subject in an indoor space. In this work, a localisation method based on fingerprinting techniques is used, which is explained below. First of all, it is necessary to create the radio map = {ℱ , ℒ} which consists of capturing, at diferent indoor positions, the received signal strength (RSSI) by each beacon available in the environment. In our case, the beacons are BLE. The set ℱ is defined as follows: ℱ = { 1, 2, . . . , } (1)
This set is composed of vectors or fingerprints, stored as vectors of RSSI measurements ( = { 1,, 2,, . . . , ,}, ∈ [1, ]), where represents the number of beacons used in the experiment, and , represents the RSSI measurement of beacon associated with sample . On the other hand, there exists the set of ℒ annotations, which consists of an -dimensional vector associating each fingerprint to a location. We define it as:
ℒ = { 1, 2, . . . , }
The main objective of the positioning is to estimate the location of a user using the radio map as a training database. For this purpose, it is very common to use the -NN algorithm which, in its most elementary version, given a test fingerprint, obtains the most similar fingerprints, based on a distance function, among those included in the radio map and estimates the user’s location ( ) as the centroid of the locations associated to the closest fingerprints. (2)
Association rules is an artificial intelligence technique widely used in Data Mining, which consists of detecting and extracting intrinsic structures or patterns in the data, in an unsupervised way, by generating first-order logic. The rules are of the form {1, 2, 3, . . . , } Z⇒ , where each rule has an antecedent and a consequent. The antecedent consists of a series of predicates that are disjunctively interpreted: 1 ∧ 2 ∧ 3 . . . ∧ . On the other hand, the consequent is the predicate that is concluded when the antecedent is true.
A decision tree is a predictive model, quite well known within the field of machine learning, which consists in performing a series of divisions of the data, based on a measure of impurity [11]. It aims to generate a tree-like structure, where the nodes are partitions of the data, and the branches are the rules that are applied to split the data. The potential of the decision tree lies in the interpretation of the results. This allows, within the field of data mining, to detect intrinsic data patterns.
Figure 2 shows a schematic of our proposal to obtain a set of rules to define the behaviour of users during their daily life in the nursing home. The diferent elements are explained below.
Given a user, and from the RSSI measurements obtained during the data collection campaign, we obtain one fingerprint per minute. The set of all test fingerprints is denoted by ℱ . The objective is to obtain, for each fingerprint, the location of the user (symbolic in our case), i.e. the set ℒ^ . For this purpose, the -NN algorithm with = 1 has been used. By means of the above process, we obtain the user’s location in each of the 1440 minutes of the day.
We call the above data Labelmap and it can be expressed as shown below: = { 1, 2, . . . , }
Where d is the total number of days on which data has been collected from a user and is a vector of 1440 location annotations estimated ( 1^,) on a given day such that: = {^1,, . . . , ^,}, ∈ [1, ], = 1440
Once the Labelmap set ( ) is defined, feature extraction is applied by partitioning temporary windows according to the chosen requirements. For each partition, the number of occurrences of each symbolic location for each is counted.
(0,1) is an occurrence function such that: • (0, 1) is the time interval (in minutes of the day) at which we calculate the occurrence of a symbolic location.
, • (0,1) the number of minutes the user is in room , on day , for the time window (0, 1). • is the number of diferent locations in this study.
The dataset resulting from feature extraction, , can be obtained as follows: where: (3) (4) (5) (6) (0,1) : → {(,01,1), (,02,1), . . . , (0,1), . . . , (0,1)}
, ,
∀0,1 ∈ [
= ⋃=︁1 = ⋃=︁1 (0 ,1 )( ) = {(,101,11), (,102,11), . . . , (,10,11), (,201,21), (,202,21), . . . , (,20,21),
,1 ,2 ,
. . . , (0,1), (0,1), . . . , (0,1)}
{∀0,1 ∈ [
0, ) (see Eq. 5) is the occurrence function in time partition .
• (,0,1 ) is the total number of minutes in which the user is in the room.
Finally, we discretise the set by applying the following: (′0,,1 ) = {︃1 ,
(0 ,1 ) ≥ 0 ℎ
With as the threshold time, in minutes, for which we consider whether a subject is at a location. We apply this discretisation to the set and obtain a data set ′ . This set contains rows, where each row ′ is a vector of size as described by the equation (8).
Where: ′ = {(10,11), (10,11), . . . , (′10,,11), ′,1 ′,2 ′,1 ′,2 ′, . . . , (0,1), (0,1), . . . , (0,1)} ∈ [1, ], ∈ [1, ], ∈ [1, ], ′, (0 ,1 ) ∈ {0, 1} (7) (8) (9) • ′ ∈ R x is the dataset resulting from feature extraction (see Eq. 6) and discretisation (see Eq. 7), with rows and columns. • number of partitions. • (′0,,1 ) is a dichotomous variable indicating whether on day , in the time interval (0 , 1 ), the user is in room , (′0,,1 ) = 1, or he/she is not, (′0,,1 ) = 0.
Finally, we obtain the annotated feature extraction set , where for each day we have annotated the corresponding day of the week ∈ : ′ = { , } Being ∈ {Monday, Tuesday, . . . , Sunday}, a vector of annotations of dimension Rx1.
The dataset obtained with the extraction of annotated features is the one we will use to extract the association rules later on.
In order to use the association rule algorithm, a priori, we have to transform (see Eq. 9) into a set of ℳ transactions. For this, we define ℐ as the set of distinct attributes (or items) and as the transactions containing a set of ℐ items, such that ⊆ ℐ [12]. We define ℳ to be the set of all transactions such that: ℳ = {1, 2, . . . , } (10)
We apply a transformation to obtain the set of itemsets ℐ. Each row of is represented by a vector of size + 1, where each attribute is a dichotomous variable (′0 ,1 ) ∈ {0, 1} indicating presence in a room, and (day of the week) is a categorical variable of seven possible values. For each dichotomous variable, two items (dichotomous complementary attributes)(′0 ,1 ) → {1, 2} are generated. The transformation generates the vector of attributes that compose the itemset ℐ ∈ R2+7, a set of 2 + 7 items that represent all the possible combinations that a user can make in a transaction.
Finally, each transaction ⊆ ℐ is a set of items observed on day . If we put all the transactions together we obtain the set ℳ that we can see represented in Table 1.
For automatic pattern determination, a rule-based descriptive method is used. To describe a subject, we have to obtain the rule extraction file ℛ = {1, 2, . . . , }, in which each rule has the form shown in equation (11).
1 : {1, 2, . . . , } Z⇒
There are three metrics used in data mining to evaluate the performance of a rule extracted from data: • Support. Proportion of transactions containing ∪ versus the total number of transactions: |∪|
|ℳ| • Confidence . Proportion of transactions containing ∪ against the support of the antecedent: |∪|
|| • Lift [13]. It measures the distance between the observed and expected support under the (∪) assumption of independence between the elements: ()· ()
For the automatic extraction of rules we have considered two alternatives. The first one consists of using a priori association rule algorithms. The second one consists in training a decision tree and, with the partitions made by the tree, converting them into rules and calculating the three metrics mentioned above. (11)
4.1. Data In order to evaluate the proposal presented in this paper, the sixth campaign of the database published in [14] has been used. During this campaign, six volunteers were provided with a smartwatch. These volunteers were required to wear the watch at all times, except when the watch was removed by the care staf to be recharged (usually before going to bed). During the day the watch is continuously acquiring the RSSI of the diferent beacons deployed throughout the nursing home as shown in Figure 3. This campaign lasted 8 weeks.
To adjust the hyperparameter of the -NN algorithm, we used a validation set included in the database [14] in which a set of fingerprints was obtained in each of the locations included in the nursing home (room, dining room, gym,...). By means of cross-validation, the performance of the -NN algorithm was evaluated for diferent values of , being = 1 the value for which the best result was obtained, where the accuracy of the classifier, for the validation set, is found to be very high (100%).
It is with this trained model that we will obtain all the symbolic annotations of the users of the residence in order to obtain the labelmap shown in Figure 4. As this figure shows, it can be seen that practically all users go to the dining room (orange colour) at the same time. However, the activities before or after the meals vary greatly between users.
After obtaining , we divide the time windows into two partitions: morning (9:00-13:30) and afternoon (13:31-16:50). To do so, we use the values shown in Table 2.
We apply these values to the occurrence function ( 0, ) (see Eq. 5) and obtain the feature 1 extraction file . Finally, we discretise it with = 5 minutes, join it to the weekly annotations and obtain the discretised and annotated feature extraction file .
The next stage is the clustering of locations with similar activities and at a close distance, to define three groups: • Physical activity: Gym, corridor, therapy room. • Passive activity: TV room, terrace.
• Room: Room.
In this work, we have excluded the locations living room and dining room as we want to focus on the activities they carry out in the morning (before lunch) and in the afternoon.
One way to visualise volunteers’ routines is by using frequency tables. If we group by day of the week, we can observe for a specific volunteer, the percentage of times that on a Monday, for example, they are in the room, doing some kind of physical activity or doing a passive activity. Similarly, we can determine for the entire campaign, the percentage of times that each volunteer has been found doing each of the activities. In this way, we can determine which volunteers are more similar or if any volunteer has unusual behaviour.
We have applied the a priori algorithm with the package arules in R [15]. For this problem, we start with transaction data. The first step is to extract daily rules and remove them and then extract weekly rules. To do the former, we use a function itemFrequency, which returns the individual support of each item belonging to ℐ, and we eliminate the items whose support is higher than a certain threshold which, in our case, we have set at 90%.
For the rule generation with a priori we have defined the following parameters: • maxlen = 4 → The number of items making up the antecedent does not exceed four • support = 1e-5 → Minimum support to generate the rule • confidence = 1e-5 → Minimum confidence to generate the rule
For the decision trees, we have used the rpart and tidyrules libraries to train the tree and extract association rules from the discretised and annotated dataset. First, we extract the daily rules with itemFrequency and then we fit a decision tree for each subject using the CART algorithm [16] with a maximum depth of 3 and a minimum of 4 samples per node. Finally, we extract rules from the tree and obtain the metrics with tidyRules. In our case, we have limited the depth of the tree to maintain interpretation without having an excessive number of antecedents in a rule.
To determine the accuracy of the algorithms used for automatic rule extraction, we have an annotation file in which the care staf detailed the daily and weekly behaviours of each of the participants in the study. In this way, we can extract rules and check whether there is a contradiction with those noted in the file. In this case, we will perform this analysis with user 9FE9 from campaign 6. The rules provided by the care staf that we can extract from the annotation file are the following: • Every day – In the morning: the volunteer does physical activity. – In the afternoon: the volunteer usually does physical or passive activities and goes to his/her room when finished. • Tuesdays: the volunteer does physical or passive activity in the morning. • Twice a week: the volunteer spends the whole morning doing physical activity.
To find patterns, we first use a daily frequency table (see Table 3). From user 52EA we noticed that he/she does not do physical activity and he/she is not in leisure areas, spending most of the time in his/her room. In contrast, user 9FE9 is physically active and is usually found engaging in passive activities during the afternoon, such as spending time on the terrace or watching TV. We also use a weekly frequency table (see Table 4) to see which activities are performed on each day of the week. From user 02A8, we can see that he/she engages in less physical activity on Thursdays and spends most of his/her time doing passive activities on Tuesday afternoons. On the other hand, user 9FE9 is not usually in his/her room on Thursdays and Fridays. Instead, he/she is frequently found doing some physical activity or any other type of activity.
Whether we use an association rules algorithm or decision trees, we have first obtained the absolute rules with the itemFrequency. In the case of volunteer 9FE9, we extract as a daily rule PhysicalMorning = 1, as we can see in Figure 5. This rule agrees with the annotations provided by the care staf.
The next step is to apply the a priori algorithm and a decision tree to volunteer 9FE9. First, we remove the PhysicalMorning attribute, the daily rule obtained with itemFrequency, and we obtain a weekly rule set of 315 rules with a priori. We select the seven best rules according to the gain and obtain the seven strongest rules that we see in Table 5. Finally, we train a decision tree, as we see in Figure 6 and extract its weekly rules with tidyRules and obtain a set of 5 weekly rules. Table 6 shows these rules. 1: {Passivemorning=0, Physicalafternoon =0} Z⇒ 2: {Passivemorning=1, Roommorning=1} Z⇒ 3: {Passivemorning=0, Physicalafternoon =1, Passiveafternoon =0} Z⇒ 4: {Passivemorning=1, Roommorning=0} Z⇒ 5: {Passivemorning=0, Physicalafternoon =1, Passiveafternoon =1} Z⇒
This paper discusses the technology used to obtain data, the steps involved in the estimation of symbolic location and the methods applied for the descriptive analysis of people’s rule-based habits. The data used in the paper are limited due to time mismatches, excessive missing data and small sample sizes. Rule association algorithms have potential, but require a large amount of sample information and may generate redundant rules. On the other hand, decision trees provide robust rules and a clear visualisation and interpretation of them. As possible lines of future work, we propose to improve the sample quantity by generating synthetic data and the possibility of using Bayesian networks. Alternatively, we are working on a new longer data campaign in order to better verify this proposal.