Classifying Alzheimer’s wandering and apathetic behaviours using indoor and outdoor localization data

Abstract

Dementia, especially Alzheimer’s disease, is increasing due to population ageing and is projected to nearly triple by 2050. Understanding and identifying common behaviours such as apathy, depression, and wandering can enable caregivers and medical professionals to enhance the quality of care for patients and implement appropriate interventions to mitigate cognitive decline. This paper analyses localization data collected from patients with mild cognitive impairment and Alzheimer’s disease residing in a healthcare facility. In particular, metrics and behavioural indices are computed to classify the behaviour of individuals. The proposed method was initially applied to the localization data of seven residents, each exhibiting distinct behaviours. Subsequently, validation was conducted using data from 46 patients monitored over three months. Classification results were compared with the observations made by healthcare operators and physicians. The study demonstrated a high level of accuracy in classifying patients into different behavioural domains, such as wandering and depressed–apathetic, with 42 out of 46 cases correctly classified.

Keywords

Alzheimer’s disease elderly patient monitoring localization wandering behaviour detection behavioural classification

1. Introduction

Alzheimer’s disease (AD) is the most common form of dementia, accounting for about 60% to 70% of dementia cases, as reported by the World Health Organization (2023). It is the fifth leading cause of death among Americans age 65 and older, with approximately one in nine people (10.8%) in this age group suffering from Alzheimer’s dementia (Better, 2023). Dementia manifests through a spectrum of symptoms, including memory deficits, challenges in problem-solving, language impairments, and other cognitive limitations, affecting an individual’s daily activities.

Common behaviours frequently observed in Alzheimer’s patients include apathy, depression, and wandering. Apathy is characterized by a lack of feeling, emotion, interest, concern, or motivation unrelated to decreased consciousness, cognitive impairment, or emotional distress. Depression, on the other hand, often results in disruptions to sleep and appetite, increased anxiety, and, in the case of AD patients, fatigue and slower movements (American Psychiatric Association, 2013). Wandering is defined as a syndrome of dementia-related locomotion behaviour characterized by frequent, repetitive, temporally disordered, and/or spatially disoriented patterns, such as lapping, random wandering, or pacing (Algase et al., 2007). It impacts patients’ quality of life and poses significant challenges for caregivers and healthcare providers.

In this work, we use a tracking technology with coarse-grained localization within a healthcare facility dedicated to AD patients where they can move freely indoors and outdoors within delimited areas. We propose a data-driven method to detect, classify, and analyse their behaviour patterns. Monitoring AD patients’ movements can offer better insights into the progression of the disease (Levine et al., 2020), understand their social behaviour (Bellini et al., 2021), and enable the implementation of targeted interventions and enhanced care practices (Kleine Deters et al., 2024).

The tracking system utilizes antennas with a received signal strength indicator (RSSI) to detect positions at a coarse level, such as room level (indoors) or specific areas (outdoors) (Masciadri et al., 2018). By analysing the movements over several days and starting from the classification of pattern-based movements (i.e. direct, pacing, lapping, and random) introduced by Martino-Saltzman et al. (1991), we describe a method that extends previous approaches to classify depressed/apathetic, wandering, or normal behaviours. Initially, the method was applied to localization data from seven guests at a residence for mild Alzheimer’s patients, each exhibiting a distinct behaviour. Subsequently, it was validated using data from 46 residents with AD monitored for three months. Comparing the results with direct observations made by the facility’s staff, the experimental outcomes demonstrate the method’s ability to effectively identify the main traits of the behaviours of Alzheimer’s patients.

The paper is structured as follows: Section 2 reports background works and reviews the most common approaches for the analysis of the behaviour of Alzheimer’s patients; Section 3 describes the localization system and the devices used to track the movements. Section 4 describes the proposed method, while Section 5 presents and discusses the results of our experiments. Finally, Section 6 summarizes the findings and concludes.

2. State of the art

Most studies monitoring AD patients’ behaviours in the literature aim at (a) localizing and tracing individuals with AD in case of wandering or getting lost (e.g. Wojtusiak and Mogharab Nia, 2021), (b) comparing the patterns of AD patients to healthy people in outdoor scenarios (e.g. Ghosh et al., 2022; Puthusseryppady et al., 2022; Bayat et al., 2021), or (c) identifying the violation of specific geographical perimeters, either indoors or outdoors (e.g. Yuce and Gulkesen, 2013; Mendoza et al., 2017).

A group of studies focus on the identification of wandering behaviours. An accepted model for their identification is that of Martino-Saltzman et al. (1991). They analysed movements observed in an elderly nursing home and identified four distinct travel patterns: direct, that is, a straightforward path from a location to another; random, that is, a path where points do not follow a particular order; pacing, that is, a path that moves back and forth between two locations; lapping, that is, a loop path involving at least three points. The latter three behaviours are considered potential forms of wandering.

The proposed algorithm has been further refined in Vuong et al. (2011) to consider also temporal information of the movement like the time spent at the destination and in Vuong et al. (2014) to incorporate machine-learning-based algorithms for feature extraction and classification (where random forests yielded the best results).

In Vuong et al. (2015), the approach has been adapted to utilize a wearable device with inertial sensors around the waist, wrist, and ankle. They capture orientation changes to detect the four travel patterns and demonstrate that their data-driven algorithm performs better than other machine-learning algorithms. A similar approach was also experimented more recently by Kamil et al. (2021) with a lightweight device with inertial sensors: 12 older adults with mild cognitive impairment or AD wore the wearable device for four days, enabling continuous monitoring of their movements.

Several solutions exploit global positioning system (GPS) data (e.g. Lin et al., 2012; Sposaro et al., 2010). The main disadvantage is that GPS-based tracking techniques only apply in open environments and are affected by factors such as weather conditions or crowded areas. It would be essential to monitor indoor movements to ensure continuous monitoring and achieve a comprehensive behavioural analysis across diverse environments and activities.

In the context of behavioural and psychological symptoms of dementia (BPSD), standardized assessment instruments such as the neuropsychiatric inventory (NPI) and the behavioural pathology in AD rating scale (BEHAVE-AD) have been introduced (Ismail et al., 2013). These instruments are designed to evaluate a wide range of symptoms associated with dementia, including but not limited to wandering, depression, apathy, hallucinations, and agitation/aggression. However, despite their widespread use in research settings, their practical application in clinical settings is still limited due to their time-intensive nature, often requiring structured interviews with caregivers (Cloak and Al Khalili, 2021; Kales et al., 2015). As a result, healthcare professionals do not incorporate these scales into routine clinical practice.

More efficient and user-friendly assessment tools are needed to provide timely insights into behavioural patterns without burdening caregivers and/or clinicians. To this end, this paper proposes a data-driven methodology for extracting behavioural indices from patients’ localization data.

3. Tracking system

The tracking system consists of Bluetooth-receiving fixed antennas distributed throughout the facility, covering buildings and open spaces, and wearable bracelets equipped with TAGs.

Antennas. The antennas are Raspberry Pi 3 Model B v1.2. The devices are configured for network boot without a micro SD card. The Raspberry Pi is powered by a PoE power supply (model LS-POE B0-525). Network boot was chosen to have a single centralized image; this choice simplifies maintenance and updates, and eliminates the need for SD cards, reducing the risk of failure and costs. However, it is necessary to manage the boot phase by segmenting the network to limit bandwidth requests and prevent network saturation.

In Indoor spaces, the antennas are placed in the false ceiling, about 2.7 m high, approximately in the centre of the rooms and corridors, spaced 6 m apart from each other, or in strategic positions for larger rooms, always keeping about 6 m between the centre of the antennas (with a density of about 28 $m^{2}$ per antenna). In outdoor spaces, since a precise location is no longer required and there are structural constraints in open areas, the antennas are spaced variably, ranging from 10 to 30 m apart. The height varies from 1.5 m (on outdoor lamps) to about 2.5 m (on buildings). The position is constrained by the availability of a path for network cable routing and the architecture of the area. Further details about the strategies for positioning the antennas can be found in previous work Masciadri et al. (2019).

The location of the antennas is therefore a trade-off between the need for accurate TAG location estimation and the availability/possibility of placing antennas. In our setup, consisting of 207 antennas, we have ensured 100% coverage of the entire structure.

An excerpt of the facility’s coverage map is depicted in Figure 1, showing various densities in the open spaces and indoors. A nearly complete indoor coverage has been achieved within the areas of interest.

Figure 1.

Excerpt of the facility coverage map with outdoor and indoor areas.

TAGs. Individuals wear active TAGs. The TAGs are primarily wristbands resembling watches. Figure 2 shows the wearable device. Specifically, the bracelet (upper part of the image) and the TAG holder are separable. The TAG holder is visible in the lower part of the photograph, along with the actual TAG. The battery, a CR2032, is not visible and is positioned underneath the TAG, which defines the size constraint of the TAG. For users’ resistant to wearing wristbands, we experimented with pins, placing the central unit in small, discreet pockets sewn into clothing.

Figure 2.

The wearable Bluetooth wristband with its separable TAG holder and the TAG itself (in the lower part).

The TAG devices used in the localization system need to have their Universally Unique Identifier identification codes modified to make each TAG distinguishable from the others. Each TAG is then associated with a guest within the localization system. This process requires a mobile app to manage the Bluetooth TAGs: we used BeaconFlyer, available on both the Apple App Store and Google Play Store. Moreover, all necessary data protection measures, including pseudonymization and disk encryption, have been implemented in compliance with General Data Protection Regulation.

Data transmission and pre-processing. TAG transmits a unique identifier every 300 ms. This interval was chosen to maximize battery life, considering the system’s goal of long-term monitoring. Antennas intercept messages broadcasted by the TAGs. The RSSI of received packets from all visible TAGs is used to estimate the TAG position. To reduce possible noise due to signal interference, reflections, diffraction, and environmental factors such as walls and furniture, or human bodies attenuating Bluetooth signals, RSSI data from each tag undergo pre-processing using a moving median filter. The filter is implemented on the antenna. The window size for the filter varies based on the specific location of the antenna. For instance, indoor antennas (e.g. rooms and corridors) use a 15-sample window; outdoor antennas, where noise levels tend to be higher due to longer distances and fewer obstructions, use a larger window of 65 samples to smooth out the signal variations. Moreover, on each antenna, only RSSI readings above a predefined threshold are considered during localization, which helps eliminate low-strength signals that could be caused by distance, reflections, or other sources of interference. This filtering process reduces network load by ensuring that only meaningful values are considered. Data collected from detected TAGs are periodically (every 10 s) transmitted via HTTP to a server, which estimates the TAG’s location.

Position estimation. Proximity is determined using the winner-takes-all method: the estimated position is assigned to the antenna showing the highest coupling to the TAG, which may not necessarily coincide with the nearest physical antenna.

Several estimation errors may occur. On the server, two main filters are implemented. One filter is used to reduce bouncing effects between adjacent rooms. The second filter implements a heuristic-based approach to check the consistency of location transitions. If an estimated TAG location suddenly jumps to a distant area without passing through adjacent antennas, the system retains the previous valid position until a plausible transition is detected. Or, if wall crossing occurs, constraint mechanisms apply to allow only admissible transitions Masciadri et al. (2019). This helps to prevent erroneous location estimates due to temporary signal losses or extreme noise.

More in detail, the algorithm for position estimation is the following. For each tag ${TAG}_{i}$ at time $t$ , the server analyses the list of messages $L_{i} (t)$ . The proximity detection method selects the message with the highest RSSI as the candidate position ${CP}_{i} (t)$ . The selected position ${SP}_{i} (t)$ is then determined by applying constraints to prevent incorrect detections of wall crossing. The selected position ${SP}_{i} (t)$ is evaluated by considering both the current candidate position ${CP}_{i} (t)$ and the previously selected position ${SP}_{i} (t - 1)$ as follows: –

If ${CP}_{i} (t) = {SP}_{i} (t - 1)$ : ${SP}_{i} (t) = {CP}_{i} (t)$ , and the likelihood that ${TAG}_{i}$ remains in ${SP}_{i} (t - 1)$ increases by $P_{δ}$ .

–

If ${CP}_{i} (t) \neq {SP}_{i} (t - 1)$ : The likelihood that ${TAG}_{i}$ remains in ${SP}_{i} (t - 1)$ decreases by $P_{δ}$ .

If the probability

P ({SP}_{i} (t))

falls below a minimum threshold

P_{\min}

, wall crossing is permitted, and the selected position

{SP}_{i} (t)

is updated to the candidate position

{CP}_{i} (t)

Notice that this methodology allows for coarse-grained localization (i.e. a person is always located under an antenna) as opposed to more complex methods such as trilateration. Therefore, the estimated location represents an area rather than the precise TAG position and the estimated paths are those determined by the connections between adjacent antennas.

System performance. The system’s localization and tracking performances were thoroughly analysed through various tests. Here, we report only the main results, being the main focus of the paper on the methodology for detecting individual behaviours using coarse-grained localization.

Two dynamic tests that are important to conduct are settling time and oscillation analysis. The settling time refers to the duration it takes for a TAG to be detected accurately as it moves to a new position. In our experiments, the settling time is less than 5 s, even with TAG transmission times of 300 ms, thus the system ensures real-time tracking of TAG movements. The oscillation analysis occurs when a TAG is in an intermediate position between two antennas or when masking conditions arise. It is essential since it can adversely impact path length estimation and the identification of different movement behaviours. In our setting, the probability of oscillation is less than 5%.

The system’s estimation of the residents’ movements exhibits high accuracy, with a precision of 0.75 (this metric measures the system’s ability to correctly associate the TAG with relevant antennas along the actual route during movement) and a recall of 0.85 (this metric evaluates the system’s capability to identify the predefined route accurately, connecting only to antennas that are relevant to the programmed path).

4. Method

4.1. Overview of the process

Starting from the data collected by the monitoring system for each user, a series of steps are applied to analyse individuals’ movement patterns and behaviours. Figure 3 outlines the main steps of the process, grouped into two main stages: (a) classification of walking patterns and (b) classification of individuals into three behavioural domains (wandering, apathetic/depressed, and normal). The output of the process consists of three percentages, representing the estimated likelihood that the resident’s state is classified as normal, wandering, or apathetic/depressed. The steps are detailed in the following subsections.

Figure 3.

Overview of the data-driven process for evaluating Alzheimer’s disease (AD) behaviour.

4.2. Walking patterns classification

The first phase consists of identifying the various pattern-based movements of each patient. We rely on the accepted classification model proposed by Martino-Saltzman et al. (1991). We adapted it to accommodate the varying granularity and unique characteristics of the data collected in our study, consisting of trajectories represented by the sequences of positions of antennas with their timestamps. In particular, we redefined key concepts such as location, stop location, and travel episode.

4.2.1. Stop locations identification

In our model, we determine the resident’s location based on the antenna’s position that receives the highest RSSI signal. To avoid wall crossing, we force the resident’s movement to follow existing paths in the structure. The antennas are placed at a certain distance from each other, which means that the location data changes only when the resident is in range of the next antenna. To enhance the accuracy, we implement a moving average filter on the location data. This process mitigates position fluctuations and minimizes false localizations.

The stop location is a location where the person stops. Due to the relatively low granularity of localization data, we consider two factors when identifying stop locations. Firstly, a location is classified as a stop if the individual spends at least 15 s there, adhering to the standard protocol utilized by gerontologists (Algase et al., 2008). Secondly, we also factor in the time the individual takes to transition to one of the subsequent locations. In some cases, an individual may appear stationary to the same antenna for some time, even when they are simply passing through a location. To address this scenario, we compute the time needed for the person to traverse the space between two adjacent nodes, considering their average speed and the distance between them.

In our model, we compute the distance between two antennas using the Euclidean and Manhattan distances. We reasonably assume that the distance covered by a person is the average of these two distances. Individuals typically prefer to take the shorter path (Euclidean distance) when moving between locations. However, they are often constrained by obstacles such as doors and corners, which make the Manhattan distance also relevant in their movement decisions.

To compute the person’s average speed, we can monitor their movements over a period and calculate the time taken to traverse the distance between adjacent antennas. Alternatively, we can consider the average walking speed of hospitalized older adults (Graham et al., 2010), which is 0.5 m/s. For our experiments, we have adopted the latter value.

A stop location is therefore defined as the location where the duration spent there surpasses the threshold $StopTimeThreshold = 15 + TimeToNextLocation [seconds]$ , where $TimeToNextLocaation$ is the time needed to reach the subsequent location in the trajectory and is computed as $0.5 [m/s] \times distanceBetweenAntennas [m]$ .

4.2.2. Travel episode identification and classification

The movement between two consecutive stop locations is defined as a travel episode: this is an aggregation of movements consisting of consecutive movements from one location to another, ending in a stop location. For example, a trip from A to D (where D is the destination where the person stops) composed of three consecutive movements $A \to B, B \to C, C \to D$ among the adjacent antennas A, B, C, and D is a travel episode.

Each travel episode is then classified as a direct, random, pacing, or lapping pattern movement according to the definitions proposed by Vuong et al. (2011).

Notice that multiple patterns may occur within a travel episode. For instance, in a sequence such as ABABCABC, a pacing pattern ABAB can be detected first, where the person moves forth and back; a lapping pattern ABCABC can also be identified, where the person travels twice in a loop: in such a case, the classification is determined by the most frequent sub-pattern. In case of a tie, priority is given to random, followed by pacing and, finally, lapping patterns. The reason behind considering random patterns first is that there is a higher correlation between cognitive impairment and the amount of random wandering compared to the other two patterns (Algase et al., 2001).

After applying the two aforementioned steps, the result is a list of travel episodes, each classified into direct, random, pacing, or lapping patterns. These travel episodes are interleaved with stop locations, and the corresponding timestamps of their occurrences are recorded as well. In the case of travel episodes, the sequence of locations (i.e. antennas) traversed during the journey is also stored for further analysis. An example of such a travel episode is presented in Table 1, depicting a movement from location 15 to location 11, classified as random, followed by a stop in location 11.

Table 1.
Example of output of the first steps of the algorithm with a travel episode and its walking pattern classification, followed by a stop location.

User Day Travel episode Start time Duration Type

5 12 June 2021 [15, 9, 8, 2, 4, 9, 11] 19:54:30 3 min 20 s Random

5 12 June 2021 [11] 19:57:50 20 min 10 s Rest place

User	Day	Travel episode	Start time	Duration	Type
5	12 June 2021	[15, 9, 8, 2, 4, 9, 11]	19:54:30	3 min 20 s	Random
5	12 June 2021	[11]	19:57:50	20 min 10 s	Rest place

4.3. Domain classification

The main goal of this step is the classification of the individual as a possible wanderer, apathetic/depressed, or normal individual. A “wanderer” spends a significant amount of time walking. This class encompasses the three walking patterns – random, pacing, or lapping – identified by Martino-Saltzman et al. (1991). On the other hand, a depressed/apathetic person spends less time walking and typically exhibits a direct walking pattern Algase et al. (2001). If a person is neither wandering nor displaying apathetic behaviours, they are classified as normal. To assign individuals to one of three classes (normal, wandering, or apathetic/depressed), a fuzzy-based approach is used alongside proposed statistical metrics.

The classification process involves several steps. After classifying the travel episodes over a period of time, their duration and distance are calculated to generate key behavioural metrics (see Section 4.3.1). These metrics are then normalized to allow comparison across patients (see Section 4.3.2), and two aggregated indices are computed (see Section 4.3.3). These indices are used in the final classification, to categorize the individual behaviour as a set of percentages in the three domains: wandering, apathetic, or normal (see Section 4.3.4).

4.3.1. Metrics computation

The classified travel episodes for each resident are collected on a daily basis. Then, for a given observation period over weeks/months, three metrics are calculated for each individual, for each single walking pattern:

–
Rhythm of walking: This metric represents the average daily frequency of each type of pattern. It is a pure number. For example, a person may exhibit, on average, 15 direct, 13 random, two lapping, and three pacing patterns per day. The underlying concept is that a higher number of occurrences of random, lapping, or pacing patterns indicate a higher probability of wandering behaviour.
–
Duration of walking: This metric denotes the average daily time the person walks for each pattern type. It is expressed in seconds. For instance, on average, a person may walk 1922 s in direct patterns, 1543 s in random patterns, 248 s in lapping, and 294 s in pacing patterns per day. Once again, a longer duration spent walking with random, lapping, or pacing patterns suggests a higher likelihood of wandering behaviour.
–
Movement (space): This metric represents the average daily space covered by the person for each type of pattern. It is expressed in metres. For example, a person may walk, on average, 451 m in direct patterns, 387 m in random patterns, 100 m in lapping, and 130 m in pacing patterns per day. Similarly, a larger space covered by random, lapping, or pacing patterns implies a higher probability of wandering behaviour.

4.3.2. Metrics normalization and aggregation

Notice that the number of direct and non-direct patterns may vary depending on the specific structure or residence under analysis. Factors such as the organization of spaces and the types of AD can strongly influence the values of the computed metrics. To address this variability and enable meaningful comparisons, we introduce a normalization process that considers an interval defined by the maximum and minimum values for each metric.

Using the data collected from all guests, global daily maximum and minimum values are calculated for each metric. The daily metrics for each guest are normalized to the overall daily trend of the entire guest population (resulting in a normalized daily value). It is important to note that movement patterns may vary due to factors such as weather, but this variation is minimized through the normalization process described.

The formula for normalization is expressed as follows: $(val - min) / (max - min)$ , where $val$ represents the metric value for the person, and $min$ and $max$ denote minimum and maximum values of reference, respectively, for that metric concerning the considered pattern. Being difficult to set specific minimum and maximum values, we decided to consider a relative evaluation: we compare the metric value of each individual with the metric values of the entire group of monitored people within the structure. In this way, we can implicitly take into account the characteristics of the structure (e.g. the layout of the spaces that may induce lapping/pacing behaviours) and the kinds of monitored AD patients/residents (e.g. all the patients have mild cognitive impairments, or they are at a preliminary stage of the disease, etc.).

Moreover, we applied a further simplification. Considering that the locomotion behaviour of AD patients manifests in lapping, random, and/or pacing patterns, we combine these patterns into a single aggregated metric as they can be considered estimators of the same (wandering) condition.

For instance, considering the rhythm data for a specific individual (15 direct, 13 random, two lapping, and three pacing patterns per day) and assuming the ranges for the rhythm metric for the entire group of monitored people are as follows:

–
Random/pacing/lapping (RPL) rhythm range: [min 2, max 40] paths per day
–
Direct (D) rhythm range: [min 5, max 50] paths per day;
the normalized aggregated metrics for the rhythm of this individual become:

–
% ${RPL}_{rhythm} = 42 % ((13 + 2 + 3) - 2) / (40 - 2))$ ).
–
% $D_{rhythm} = 22 % ((15 - 5) / (50 - 5)$ ).
The percentages, %RPL and %D (direct behaviour), are computed for all three metrics: rhythm, duration, and movement. Consequently, we obtain a total of six percentages.
4.3.3. Characterization of the behaviour: Behavioural indices computation

The probability of a person being in a domain (normal, wandering, and depressed/apathetic) is determined by using two $behavioural indices$ :

–
${BI}_{RPL}$ : The behavioural index for the RPL domain is calculated as the average of the three percentages related to rhythm, duration, and movement patterns for RPL: ${BI}_{RPL} = (% {RPL}_{rhythm} + % {RPL}_{duration} + % {RPL}_{movement}) / 3$ .
–
${BI}_{D}$ : The behavioural index for the direct (D) domain is obtained similarly: ${BI}_{D} = (% D_{rhythm} + % D_{duration} + % D_{movement}) / 3$ .
Notice that their average will also fall within this range, being %RPL and %D for the three metrics values between 0 and 1.

These two indicators characterize the subject’s behaviour and enable the identification of the domain(s) (e.g. wandering, normal, and depressed/apathetic) to which the person belongs. It is worth noting that a person’s behaviour may exhibit characteristics that could be predominant in a specific domain, or the behaviour could display features that overlap with more than one domain.

For example, consider the following characterization of a resident: –
${%RPL}_{rhythm} = 42 %$ , ${%D}_{rhythm} = 22 %$
–
${%RPL}_{duration} = 35 %$ , ${%D}_{duration} = 40 %$
–
${%RPL}_{movement} = 10 %$ , ${\%D}_{movement} = 32 %$
The behavioural indices are: –
${BI}_{RPL}$ : $(42 % + 35 % + 10 %) / 3 \approx 29 %$
–
${BI}_{D}$ : $(22 % + 40 % + 32 %) / 3 \approx 31 %$
Each person exhibits unique behavioural indices. Intuitively, low values of ${BI}_{RPL}$ associated with high values of ${BI}_{D}$ denote a healthy individual; conversely, if the opposite pattern emerges, it suggests characteristics more typical of an Alzheimer’s patient.
4.3.4. Behaviour classification with fuzzy matrices

Utilizing the two behavioural indices ${BI}_{RPL}$ and ${BI}_{D}$ , a fuzzy classification method is employed to categorize the individual’s behaviour into three domains: normal, wandering, and apathetic/depressed. These indices capture movement patterns and their characteristics, but for the final classification, we rely on the healthcare staff’s experience. In particular, healthcare operators contributed by providing expert qualitative evaluations based on their daily interactions with the patients. The operators classify individuals as ‘normal’, ‘wandering’, or ‘apathetic/depressed’ based on their direct observations. Their input allowed the building of three fuzzy matrices, one for each domain.

4.3.5. Fuzzy-based matrices construction

A subset of patients with different characteristics (wandering, normal, and apathetic) was selected and a questionnaire was administered to healthcare operators. Questionnaires were requested from the operators multiple times over different periods to reduce recall bias and increase reliability. The operators were asked to assess each patient’s behaviour on a continuum from wandering to apathy/depression, with a midpoint of normal. Note that not all operators are familiar with every guest in detail; furthermore, operators work shifts that cover 24 hours a day, 7 days a week, and may have different perspectives on the guests. Due to the qualitative nature of the assessment, there was inherent uncertainty in the responses.

For each guest in the selected group, the operators’ qualitative responses were converted into three percentages representing apathy/depression, normal, and wandering behaviours. For example, if a patient’s behaviour was rated halfway between wandering and normal, it was quantified as 50% normal and 50% wandering. This allowed the representation of uncertain qualitative data with quantitative metrics. The percentage scores from multiple operators were averaged and then normalized to multiples of 5% ensuring that they summed to 100%.

Notice that the operators’ observations are inherently subjective and influenced by individual interpretations and biases, but collectively, their opinions tend to be accurate, reflecting the principle of the wisdom of crowds. Consequently, the values delineating the characteristics of AD patients tend to be generally valid.

For the selected patients, the two indices ${BI}_{RPL}$ and ${BI}_{D}$ are computed to identify a specific point in the matrix. For example, if a patient obtains values of 90% for ${BI}_{RPL}$ and 10% for ${BI}_{D}$ , these percentages determine coordinates within a matrix where the values of ${BI}_{RPL}$ and ${BI}_{D}$ form the two dimensions, each divided into 10% steps. This enables the precise positioning of each individual’s percentages of wandering, normal, or apathetic states across the three corresponding matrices.

On the three matrixes, corner cases are defined according to Table 2.

Table 2.
Values associated with the corner cases.

(0,0) (9,9) (9,0) (0,9)

Apathetic 100% 0% 0% 0%

Normal 0% 0% 0% 0%

Wandering 0% 100% 100% 100%

	(0,0)	(9,9)	(9,0)	(0,9)
Apathetic	100%	0%	0%	0%
Normal	0%	0%	0%	0%
Wandering	0%	100%	100%	100%

Given the corner values and the values for the selected individuals, the remaining values are computed using Matlab’s griddata() function. This function interpolates scattered data points in an $N$ -dimensional space (typically two-dimensional or three-dimensional) and returns the interpolated values on a regular grid in the same space. The apathetic/depressed matrix and the wandering matrix are treated independently. To ensure the total sums to 100%, some values have been adjusted. Finally, the normal matrix is calculated by taking the difference.

Figure 4 presents the values obtained from this process, rounded to multiples of 5.

Figure 4.

Degree of belonging for every single domain ( ${BI}_{D}$ on the $y$ -axis, ${BI}_{RPL}$ on the $x$ -axis).

4.3.6. Classification with fuzzy matrices

The computed matrices are utilized for classifying any new individual. For instance, for an individual with computed indices of ${BI}_{RPL}$ and ${BI}_{D}$ , such as ${BI}_{RPL}$ : $29 %$ and ${BI}_{D}$ : $31 %$ , three probabilities for each domain can be derived from the matrices:

–
Depressed/apathetic = 5%
–
Normal = 80%
–
Wandering = 15%
The result shows that the person’s behaviour is predominantly normal, with a small tendency towards wandering. Monitoring these values over an extended period allows for the estimation of potential changes in behaviour, cognitive decline, or other relevant patterns.

It is worth noting that these values serve as indicators, providing an objective assessment of each guest’s condition and progress over time (behavioural drift). They assist operators and doctors in evaluating the effectiveness of interventions, such as gymnastics, group activities, and pharmacological support, by enabling them to strengthen their perception of the individual’s state accordingly.
5. Results and discussion

Localization data were collected in a protected structure, where the residents have the freedom to move within the facility but are under the close observation of specialized operators at all times. To ensure privacy and data ownership, personal information has not been provided to avoid any possibility of tracing back to the monitored individuals.

5.1. First data collection campaign

We started with the first data collection campaign in April–May 2021. During this period, we interacted and collaborated with the medical and nursing staff to fine-tune our approach. A total of seven residents, all in the fifth stage (late fourth to early sixth stage) of the global deterioration scale (GDS) (Reisberg et al., 1982) with different characteristics were observed and identified by letters from $a$ to $g$ . At this stage, the person requires more consistent assistance and may experience severe confusion about recent events, places, or people, difficulty with daily tasks such as dressing or meal preparation, increased agitation or delusions, trouble recognizing family and friends, language and communication issues, disrupted sleep patterns, and loss of control over bodily functions, such as incontinence.

The metrics for the two groups of walking patterns (random, pacing, and lapping on one side, and direct on the other side) for the seven guests are reported in Table 3. Residents $d$ and $g$ exhibit values above the average for all three metrics, indicating a higher likelihood of belonging to the wandering class. On the contrary, guests $c$ and $f$ have values below the average in all three metrics, suggesting a higher probability of belonging to the normal or depressed/apathetic class. However, resident $c$ has higher values in the RPL component compared to the direct values, indicating the need for further investigation. Patient $a$ shows low values for the RPL component and relatively high values for the D component in rhythm but with low values in movement. This pattern suggests that the individual might belong to a domain between normal and depressed/apathetic. Patient $b$ is characterized by high duration values and low movement, with a majority of RPL movements compared to D, implying a higher likelihood of wandering behaviour. Finally, resident $e$ has a high movement value with low values for the other two metrics.

Table 3.
Metrics values for each resident: Average number of random/pacing/lapping (RPL) and D walking patterns (rhythm), average duration in seconds, and average movement in metres.

Rhythm [number] Duration [s] Movement [m]

Resident Valid days (observed period) RPL D RPL D RPL D

a 45 days (2 April–16 May 21) 16.4 37.2 2620.7 1056.1 585.3 503.4

b 13 days (7–25 February 21) 26.6 35.8 4097.5 1339.2 908.8 621.3

c 30 days (10 April–16 May 21) 24.2 23.9 3138.3 747.6 847.3 414.1

d 19 days (10 April–1 May 21) 36.8 45.7 5115.0 1318.3 1477.6 896.6

e 23 days (10 April–9 May 21) 20.9 22.2 3454.1 938.2 1677.9 887.7

f 21 days (12 April–8 May 21) 28.7 21.6 3598.0 483.0 873.1 300.0

g 23 days (22 April–16 May 21) 37.0 35.3 6116.9 1661.0 1469.2 663.5

Average 27.2 31.7 4020.1 1077.6 1119.9 612.4

		Rhythm [number]	Duration [s]	Movement [m]
a	45 days (2 April–16 May 21)	16.4	37.2	2620.7	1056.1	585.3	503.4
b	13 days (7–25 February 21)	26.6	35.8	4097.5	1339.2	908.8	621.3
c	30 days (10 April–16 May 21)	24.2	23.9	3138.3	747.6	847.3	414.1
d	19 days (10 April–1 May 21)	36.8	45.7	5115.0	1318.3	1477.6	896.6
e	23 days (10 April–9 May 21)	20.9	22.2	3454.1	938.2	1677.9	887.7
f	21 days (12 April–8 May 21)	28.7	21.6	3598.0	483.0	873.1	300.0
g	23 days (22 April–16 May 21)	37.0	35.3	6116.9	1661.0	1469.2	663.5
Average		27.2	31.7	4020.1	1077.6	1119.9	612.4

Starting from data in Table 3, we can compute the maximum and minimum values for each metric as follows: –

Rhythm range:

RPL = [16.4–37], D = [21.6–45.7]

–

Duration range:

RPL = [2620.7–6116.9], D = [483–1661]

–

Movement range:

RPL = [585.3–1685], D = [300–896.5]

By using the max and min, each single value for each resident can be normalized as (

(val--min)/(max--min)

). Results are reported in Table 4.

Table 4.

Normalized metrics of Table 3.

	Rhythm [number]		Duration [s]		Movement [m]
Resident	RPL	D	RPL	D	RPL	D
a	0.000	0.647	0.000	0.487	0.000	0.341
b	0.495	0.589	0.422	0.727	0.296	0.539
c	0.379	0.095	0.148	0.225	0.240	0.191
d	0.990	1.000	0.713	0.709	0.817	1.000
e	0.218	0.025	0.238	0.386	1.000	0.985
f	0.597	0.000	0.280	0.000	0.263	0.000
g	1.000	0.568	1.000	1.000	0.809	0.609

RPL: random/pacing/lapping.

Finally, the behavioural indices for all persons can be computed as the average of the three RPL values (for ${BI}_{RPL}$ ) and the three D ( ${BI}_{D}$ ). Results are reported in Table 5.

Table 5.

Behavioural indices for the seven residents.

Resident	${BI}_{RPL}$	${BI}_{D}$	Row, Col
a	00%	49%	0.4
b	40%	62%	4.6
c	25%	17%	2.1
d	84%	90%	8.9
e	48%	46%	4.4
f	38%	00%	3.0
g	94%	73%	9.7

The obtained behavioural indices (reported in Table 5) allow identifying the probability of a person belonging to a specific domain; these probabilities are summarized in Table 6 and in Figure 5, and are detailed in Figure 6.

Figure 5.

Probability of being depressed (blue – dark), normal (orange) or wandering (yellow – light) for the seven residents.

Figure 6.

Mapping of the behavioural indices of the seven residents on the matrices.

Table 6.

Behavioural domain characterization for the seven residents and dominant characterization according to the staff observations.

Resident	Wandering	Normal	Depress/apathy	Staff observation
a	0%	90%	10%	Normal
b	70%	30%	0%	Wanderer
c	0%	20%	80%	Depressed/apathetic
d	100%	0%	0%	Wanderer
e	30%	70%	0%	Normal
f	20%	30%	50%	Depressed/apathetic
g	100%	0%	0%	Wanderer

These results match with the qualitative analysis described earlier and have been validated by the operators of the residence, who have direct observation and knowledge of the residents. The team members involved in the evaluation consist of a geriatrician, physiotherapist, psychologist, educator, and social health worker. They administer the NPI test and rely on direct observation to define the prevailing characteristics of the residents with respect to the presence of behaviour disorders (BPSD). According to their analysis, residents $c$ and $f$ are classified as depressed/apathetic, while $a$ and $e$ are considered normal. Residents $b$ , $d$ , and $g$ are identified as wandering. These results are reported in Table 6. We noticed that, compared to their observations, the analysis of our proposed approach provides more characterization details, which has caught their attention and highlighted certain aspects that were not as evident through direct observation only. Indeed, continuous monitoring offers the capability to detect subtle changes in residents’ behaviours and quantify various patterns that may not be noticed through mere visual observation.

5.2. Validation with a second dataset

We applied the same technique to a dataset of 46 residents of the facility collected during the period from May 2022 to July 2022. Also in this case, they are classified in the fourth stage (late third to early fifth stage) of the GDS scale. Of the 92 monitored days, on average, approximately 89% of the days were deemed significant. In the remaining cases, data collection was incomplete due to issues such as battery life limitations and occasional difficulties with residents wearing the tracking devices. The analysis is limited by the availability of patients in the healthcare facility, thus constraining the number of users and days included in the study. Nevertheless, these numbers are comparable to or even exceed those observed in similar studies in the literature, such as those investigating the behaviour of AD patients through GPS trajectory analysis (Bayat et al., 2021; Puthusseryppady et al., 2022; Ghosh et al., 2022).

During this second campaign of experiments, similar to the first one, the classification provided by our method showed a very good correspondence with the observations made by the operators and physicians at the facility. Specifically, when a resident’s behaviour was strongly characterized by a particular pattern (usually with values between 80% and 100% for a specific behaviour), the results matched the assessments made by the operators. Instead, when there were percentages between 35% and 65% the operators sometimes were unable to give specific, detailed descriptions or associate the residents with normal behaviours. For all the cases, we asked for an opinion on the indices with an answer ‘I agree–I disagree’. Among the 46 cases, only four cases were not aligned with the obtained results.

Considering the method used for normalizing the values based on the minimum and maximum values of each metric, we tried to understand the impact of such a choice, by using the normalization values obtained in the second survey campaign to calculate the indices of the first campaign survey; the result is reported in Table 7. It can be noticed that the indices (W, N and A/D) with the new normalization values are not identical but are significantly similar considering the discretization of the indices and the entries in the tables.

A continuous update of the minimum to maximum values used for normalization allows the method to converge towards a state of stability as long as the following hypothesis holds: the structure has guests with all the ‘behaviour models’, which is the typical situation in most of the residences, especially for diseases such as Alzheimer’s where different persons show different behaviours, with common characteristics.

Table 7.
Behavioural domain characterization for the seven residents of the first campaign with the original normalization values of Table 5 and with the normalization values of the second campaign.

Original normalization values Normalization values second campaign

Resident Wandering Normal Depr./apathy Wandering Normal Depr./apathy

a 0% 90% 10% 0% 75% 25%

b 70% 30% 0% 30% 70% 0%

c 0% 20% 80% 0% 60% 40%

d 100% 0% 0% 100% 0% 0%

e 30% 70% 0% 45% 55% 0%

f 20% 30% 50% 20% 30% 50%

g 100% 0% 0% 100% 0% 0%

	Original normalization values	Normalization values second campaign
a	0%	90%	10%	0%	75%	25%
b	70%	30%	0%	30%	70%	0%
c	0%	20%	80%	0%	60%	40%
d	100%	0%	0%	100%	0%	0%
e	30%	70%	0%	45%	55%	0%
f	20%	30%	50%	20%	30%	50%
g	100%	0%	0%	100%	0%	0%

5.3. Comparison of fuzzy logic with advanced classification methods

In this study, we chose fuzzy logic to classify Alzheimer’s patients’ behaviour, opting for it over other advanced classification techniques such as support vector machines (SVMs) and neural networks (NNs).

Fuzzy logic, NNs, and SVMs are three distinct approaches used in artificial intelligence to address classification, prediction, and control problems, each with its characteristics. In particular, fuzzy logic stands out for its ability to handle imprecise and uncertain data, making it particularly advantageous in contexts where available information is limited. Due to its inherently interpretable nature, fuzzy logic-based systems can easily translate complex rules into natural language, facilitating the understanding of the system’s decisions. In the situation under consideration, where variables are not clearly defined, fuzzy logic allows for the incorporation of uncertainties in evaluation and better management of ambiguities. In this context, data are incomplete and influenced by individual perception, making the fuzzy approach more robust compared to other methods. Furthermore, the problem at hand is not complex, does not involve large datasets, but allows for the definition of some rules based on human perception and experience. For this reason, the fuzzy approach is particularly well-suited to the problem under examination. From our perspective, NNs and SVMs are not appropriate for the given context. Specifically, NNs are capable of learning from training data, making them powerful for recognizing complex patterns but requiring large datasets. Conversely, SVMs are effective with small datasets but are primarily aimed at binary classification problems. In our opinion, other approaches not mentioned (e.g. logic regression) suffer from the same issues reported for NNs and SVM, making fuzzy logic the best way to transform experience-based interpretations into information.

5.4. Limitations

While the decision to focus on individuals residing in a healthcare facility provides interesting results, it also imposes certain limitations on the generalizability of the findings. The relatively small sample size limits the statistical power of the findings and may not fully represent the broader population of Alzheimer’s patients, particularly those at different stages of the disease. Further experiments should therefore include people at different stages of the disease and involve diverse healthcare facilities including home-based care, assisted living facilities, and hospitals to validate the adaptability of the proposed approach to different settings.

The behavioural indicators depend on the presence of all cases ranging from apathy to wandering. Wandering is crucial as it defines distances (which vary based on the facility’s characteristics), frequency, and duration. This is because normalization values depend on what is detected. Consequently, the method becomes more reliable with a larger population in the healthcare residence.

Tracking residents requires their consistent use of a tracker (such as a bracelet, brooch, pendant, or pocket) and diligent cooperation from staff for maintenance and intervention in case a resident is missing their TAG. This aspect is far from simple and obvious. This research achieved the reported results through various strategies, including multi-tagging residents, to mitigate hardware failures and human failure, such as forgetfulness in maintenance or the voluntary abandonment of the tracker by the resident.

6. Conclusions

This study presents a novel method for continuous monitoring and characterization of movement patterns in individuals affected by AD. The goal is to detect and classify behaviours as normal, wandering, or apathetic/depressed, along with assessing the degree of each behaviour. In previous research, specific movement patterns, such as random, pacing, and lapping have been identified. However, understanding the complex behaviour of AD patients over time poses a significant challenge for staff. Building on this knowledge, our paper proposes a data-driven approach to derive relevant indices that enable the classification of individual behaviour over an extended monitoring period, thereby returning a fuzzy result corresponding to the complexity inherent in the characteristics of the disease.

We conducted two survey campaigns using this methodology. In the first campaign, we observed seven mild cognitive Alzheimer’s patients with distinct behaviours. The monitoring was carried out continuously for approximately one month (up to 45 days). In the second campaign, we extended the analysis to 46 patients monitored for three months. The results obtained from our proposed method closely align with the observations made by healthcare operators and physicians at the facility; performance was consistent also when applied to the second population.

In the last few years, there has been growing interest towards the utilization of real-time location systems within residential care environments for locating residents (Haslam-Larmer et al., 2022) and, more in general, for Information and Communication Technology-based solutions for AD care Ali et al. (2024). Collecting data on movements and patterns over time offers the advantage of considering objective and continuous measures, which may not be feasible with direct observations alone. By accurately identifying and classifying the behaviours exhibited by AD patients, caregivers can gain valuable insights into the residents living in nursing homes and provide targeted interventions and personalized care plans to address their specific needs. Tools for continuous monitoring can optimize not only care delivery but also resource allocation; objective measures can increase families’ confidence in the quality of care provided by the facility.

Future developments of this work will focus on the evolution of behavioural characteristics and the detection of the onset of behavioural changes and drifts. Moreover, we will explore questions such as: Are there any paths, or spaces that accentuate or reduce wandering? Does the presence of other residents or operators affect the wandering/depression/apathetic conditions? Finally, a more longitudinal study, across different healthcare facilities, with continuous monitoring and characterization of AD patients’ behaviours could contribute to a more comprehensive understanding of the disease and improve care practices.

Footnotes

ORCID iD

Sara Comai

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Conflict of interest

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

References

Algase

Antonakos

Yao

, et al. (2008) Are wandering and physically nonaggressive agitation equivalent? The American Journal of Geriatric Psychiatry 16(4): 293–299.

Algase

Beattie

Therrien

(2001) Impact of cognitive impairment on wandering behavior. Western Journal of Nursing Research 23(3): 283–295.

Algase

Moore

Vandeweerd

, et al. (2007) Mapping the maze of terms and definitions in dementia-related wandering. Aging & Mental Health 11(6): 686–698.

Ali

Turetta

Demrozi

, et al. (2024) ICT-based solutions for Alzheimer’s disease care: A systematic review. IEEE Access 12: 13944–13961.

American Psychiatric Association (2013) Diagnostic and Statistical Manual of Mental Disorders (DSM-5). Arlington, Virginia, USA: American Psychiatric Publishing.

Bayat

Naglie

Rapoport

, et al. (2021) A GPS-based framework for understanding outdoor mobility patterns of older adults with dementia: An exploratory study. Geron 68: 1–15.

Bellini

Cipriano

Comai

, et al. (2021) Understanding social behaviour in a health-care facility from localization data: A case study. Sensors 21(6): 2147.

Better

(2023) Alzheimer’s disease facts and figures. Alzheimer’s & Dementia 19(4): 1598–1695.

Cloak

Al Khalili

(2021) Behavioral and Psychological Symptoms in Dementia. Treasure Island, FL: StatPearls Publishing.

10.

Ghosh

Puthusseryppady

Chan

, et al. (2022) Machine learning detects altered spatial navigation features in outdoor behaviour of Alzheimer’s disease patients. Scientific Reports 12(1): 3160.

11.

Graham

Fisher

Bergés

I-M

, et al. (2010) Walking speed threshold for classifying walking independence in hospitalized older adults. Physical Therapy 90(11): 1591–1597.

12.

Haslam-Larmer

Shum

Chu

, et al. (2022) Real-time location systems technology in the care of older adults with cognitive impairment living in residential care: A scoping review. Frontiers in Psychiatry 13: 1038008.

13.

Ismail

Emeremni

Houck

, et al. (2013) A comparison of the E-BEHAVE-AD, NBRS, and NPI in quantifying clinical improvement in the treatment of agitation and psychosis associated with dementia. The American Journal of Geriatric Psychiatry 21(1): 78–87.

14.

Kales

Gitlin

Lyketsos

(2015) Assessment and management of behavioral and psychological symptoms of dementia. The BMJ 350: h369 .

15.

Kamil

Bakar

Ehrenburg

, et al. (2021) Detection of wandering behaviors using a body-worn inertial sensor in patients with cognitive impairment: A feasibility study. Frontiers in Neurology 12: 529661.

16.

Kleine Deters

Janus

Silva

JAL

, et al. (2024) Sensor-based agitation prediction in institutionalized people with dementia A systematic review. Pervasive and Mobile Computing 98: 101876.

17.

Levine

Allison

Stojanovic

, et al. (2020) Spatial navigation ability predicts progression of dementia symptomatology. Alzheimer’s & Dementia 16(3): 491–500.

18.

Lin

Zhang

Huang

, et al. (2012) Detecting wandering behavior based on GPS traces for elders with dementia. In: 2012 12th international conference on control automation robotics & vision (ICARCV), Guangzhou, China, 5–7 December 2012, pp.672–677. IEEE.

19.

Martino-Saltzman

Blasch

Morris

, et al. (1991) Travel behavior of nursing home residents perceived as wanderers and nonwanderers. The Gerontologist 31(5): 666–672.

20.

Masciadri

Carlini

Comai

, et al. (2018) Supporting Alzheimer’s residential care – a novel indoor localization system. In: Proceedings of the 15th international joint conference on e-business and telecommunications – ICETE, INSTICC, Porto, Portugal, pp.272–278. SciTePress.

21.

Masciadri

Comai

Salice

(2019) Wellness assessment of Alzheimer’s patients in an instrumented health-care facility. Sensors 19(17): 3658.

22.

Mendoza

Bergado

De Castro

JLB

, et al. (2017) Tracking system for patients with Alzheimer’s disease in a nursing home. In: TENCON 2017-2017 IEEE region 10 conference, pp.2566–2570. IEEE

23.

Puthusseryppady

Morrissey

Aung

, et al. (2022) Using GPS tracking to investigate outdoor navigation patterns in patients with Alzheimer disease: Cross-sectional study. JMIR Aging 5(2): e28222.

24.

Reisberg

Ferris

de Leon

, et al. (1982) The global deterioration scale for assessment of primary degenerative dementia. American Journal of Psychiatry 139(9): 1136–1139.

25.

Sposaro

Danielson

Tyson

(2010) iWander: An Android application for dementia patients. In: 2010 annual international conference of the IEEE engineering in medicine and biology, Buenos Aires, Argentina, 11 November 2010, pp.3875–3878. IEEE.

26.

Vuong

Chan

Lau

, et al. (2011) Feasibility study of a real-time wandering detection algorithm for dementia patients. In: Proceedings of the first ACM MobiHoc workshop on pervasive wireless healthcare, Paris, France, 16 May 2011, pp.1–4. New York, NY: ACM.

27.

Vuong

Chan

Lau

(2014) Automated detection of wandering patterns in people with dementia. Gerontechnology 12(3): 127–147.

28.

Vuong

Chan

Lau

, et al. (2015) Preliminary results of using inertial sensors to detect dementia-related wandering patterns. In: 2015 37th annual international conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, Italy, 5 November 2015, pp.3703–3706. IEEE.

29.

Wojtusiak

Mogharab Nia

(2021) Location prediction using GPS trackers: Can machine learning help locate the missing people with dementia? Internet of Things 13: 100035.

30.

World Health Organization (2023) Dementia Report, updated March.

31.

Yuce

Gulkesen

(2013) CaregiverNet: A novel social support intervention for locating and securing wandering Alzheimer’s patients as soon as possible. In: 2013 9th international wireless communications and mobile computing conference (IWCMC), pp.1405–1411. IEEE.

		Rhythm [number]		Duration [s]		Movement [m]
Resident	Valid days (observed period)	RPL	D	RPL	D	RPL	D
a	45 days (2 April–16 May 21)	16.4	37.2	2620.7	1056.1	585.3	503.4
b	13 days (7–25 February 21)	26.6	35.8	4097.5	1339.2	908.8	621.3
c	30 days (10 April–16 May 21)	24.2	23.9	3138.3	747.6	847.3	414.1
d	19 days (10 April–1 May 21)	36.8	45.7	5115.0	1318.3	1477.6	896.6
e	23 days (10 April–9 May 21)	20.9	22.2	3454.1	938.2	1677.9	887.7
f	21 days (12 April–8 May 21)	28.7	21.6	3598.0	483.0	873.1	300.0
g	23 days (22 April–16 May 21)	37.0	35.3	6116.9	1661.0	1469.2	663.5
Average		27.2	31.7	4020.1	1077.6	1119.9	612.4

Resident	${BI}_{RPL}$	${BI}_{D}$	Row, Col
a	00%	49%	0.4
b	40%	62%	4.6
c	25%	17%	2.1
d	84%	90%	8.9
e	48%	46%	4.4
f	38%	00%	3.0
g	94%	73%	9.7

Resident	${BI}_{RPL}$	${BI}_{D}$	Row, Col
a	00%	49%	0.4
b	40%	62%	4.6
c	25%	17%	2.1
d	84%	90%	8.9
e	48%	46%	4.4
f	38%	00%	3.0
g	94%	73%	9.7

Classifying Alzheimer’s wandering and apathetic behaviours using indoor and outdoor localization data

Abstract

Keywords

1. Introduction

2. State of the art

3. Tracking system

4.1. Overview of the process

4.2.1. Stop locations identification

4.2.2. Travel episode identification and classification

4.3.1. Metrics computation

4.3.5. Fuzzy-based matrices construction

Table 2. Values associated with the corner cases. (0,0) (9,9) (9,0) (0,9) Apathetic 100% 0% 0% 0% Normal 0% 0% 0% 0% Wandering 0% 100% 100% 100%

5.1. First data collection campaign

5.4. Limitations

6. Conclusions

Footnotes

ORCID iD

Funding

Conflict of interest

References

Table 2.
Values associated with the corner cases.

(0,0) (9,9) (9,0) (0,9)

Apathetic 100% 0% 0% 0%

Normal 0% 0% 0% 0%

Wandering 0% 100% 100% 100%

Resident	${BI}_{RPL}$	${BI}_{D}$	Row, Col
a	00%	49%	0.4
b	40%	62%	4.6
c	25%	17%	2.1
d	84%	90%	8.9
e	48%	46%	4.4
f	38%	00%	3.0
g	94%	73%	9.7