Powered Wearable Technologies for Dementia Care: Evaluating

Activity Recognition Models and Dataset Challenges

Mariana Carvalho

, Inês C. Rocha

, Marcelo Arantes

, Ricardo Linhares

José Soares

, António Moreira

1,2 f

, João L. Vilaça

1,2 g

, Demétrio Matos

Pedro Morais

1,2 i

and Vítor Carvalho

1,2 j

2Ai−School of Technology, Polytechnic University of Cávado and Ave (IPCA),

Campus of IPCA, 4750-810 Barcelos, Portugal

LASI−Associate Laboratory of Intelligent Systems, 4800-058 Guimarães, Portugal

Research Institute for Design, Media and Culture (ID+), School of Design, Polytechnic of Cávado and Ave,

4750-810 Barcelos, Portugal

Keywords: Wearable Technology, Activity Recognition, AI, Elderly, Dementia.

Abstract: Dementia is a progressive neurological condition affecting millions worldwide, posing significant challenges

for patients and caregivers. Wearable technologies integrated with artificial intelligence (AI) provide

promising solutions for continuous activity monitoring, supporting dementia care. This study evaluates the

performance of various AI models, including tree-based methods and deep learning approaches, in

recognizing activities relevant to dementia care. While the first excelled in handling class imbalances and

recognizing common activities, deep learning models demonstrated superior capabilities in capturing complex

temporal and spatial patterns. Additionally, a comprehensive analysis of 30 datasets revealed significant gaps,

including limited representation of elderly participants, insufficient activity coverage, short recording

durations, and a lack of real-world environmental data. To address these gaps, future work should focus on

developing datasets tailored to dementia care, incorporating long-duration recordings, diverse activities, and

realistic contexts. This study highlights the potential of AI-powered wearable systems to transform dementia

management, enabling accurate activity recognition, early anomaly detection, and improved quality of life for

patients and caregivers.

1 INTRODUCTION

Dementia encompasses a range of neurological

disorders characterized by memory loss and cognitive

decline (Winblad et al., 2016). Currently, over 55

million people worldwide live with dementia, and this

number is projected to double by 2050, posing

significant challenges for healthcare systems and

https://orcid.org/0009-0004-1818-8213

https://orcid.org/0009-0009-7960-7748

https://orcid.org/0000-0003-3611-3974

https://orcid.org/0009-0001-6108-2463

https://orcid.org/0009-0003-3958-7742

https://orcid.org/0000-0002-2148-9146

https://orcid.org/0000-0002-4196-5357

https://orcid.org/0000-0003-4417-6115

https://orcid.org/0000-0002-1995-7879

https://orcid.org/0000-0003-4658-5844

families (World Health Organization, 2023). With

aging populations and no cure available, the

prevalence of dementia continues to rise.

As the condition progresses, symptoms may

include disorientation, mood swings, confusion,

severe memory loss, behavioural changes, and

difficulties with speaking, swallowing, or walking

(Lindeza et al., 2024). These challenges place a

Carvalho, M., Rocha, I. C., Arantes, M., Linhares, R., Soares, J., Moreira, A., Vilaça, J. L., Matos, D., Morais, P. and Carvalho, V.

Powered Wearable Technologies for Dementia Care: Evaluating Activity Recognition Models and Dataset Challenges.

DOI: 10.5220/0013396600003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 2: HEALTHINF, pages 995-1006

ISBN: 978-989-758-731-3; ISSN: 2184-4305

995

significant emotional and physical burden on both

individuals and their caregivers, requiring substantial

support from an early stage.

The role of caregivers is fundamental in

supporting elderly individuals with dementia.

However, it often presents significant challenges,

including high rates of depression and stress, physical

strain, and social isolation (Lavretsky, 2005). To

address these challenges, enhancing caregiver

support and implementing effective dementia

monitoring are crucial.

Dementia monitoring offers numerous benefits. It

helps prevent accidents by tracking movements and

reducing risks associated with wandering. Enables

long-term health tracking, providing valuable data for

caregivers and healthcare professionals to make

informed care and treatment decisions. Furthermore,

monitoring reduces family anxiety by promptly

alerting caregivers to potential issues and facilitates

patient independence, allowing individuals with

dementia to engage safely in activities both indoors

and outdoors (Lin et al., 2008).

Wearable technology plays a pivotal role in

monitoring individuals with dementia by providing

non-invasive, continuous, and objective data on

various physiological and behavioral parameters.

These devices are generally well-accepted by both

patients and caregivers, making them a practical

solution for continuous monitoring (Husebo et al.,

2020).

One notable application is GPS-enabled

wearables, which help monitor mobility patterns and

locate missing individuals with dementia, offering a

non-intrusive way to track movements and prevent

wandering (Cullen et al., 2022). Additionally, these

devices can report detailed mobility outcomes, such as

activity duration, out-of-home movements, and

trajectory patterns (Cullen et al., 2022). They also

provide insights into health indicators specific to

dementia, including lower daily activity levels,

decreased sleep efficiency, and greater circadian

rhythm variability compared to controls(Cote et al.,

2021).

For patients and caregivers, the comfort,

convenience, and affordability of wearable devices

are key priorities. Essential features include long

battery life, water resistance, and an emergency

button, which enhance usability and reliability

(Stavropoulos et al., 2021).

The work presented in this paper is part of a larger

project focused on developing a wearable device

tailored to the unique needs and challenges of

individuals with dementia. A study from this project

highlights a significant gap in the availability of

comprehensive devices, as most existing wearables

fail to provide an integrated solution that includes

activity monitoring (daily activities, daytime and

nighttime patterns, activity and movement trends),

real-time location tracking, fall detection, and SOS

alert systems (Rocha et al., 2024).

The primary objective of this paper is to describe

the available datasets obtained from wrist-worn

wearables and evaluate the best AI architectures for

predicting activities based on this data. This analysis

provides critical insights into selecting and

developing effective solutions for activity

monitoring, which is an essential step toward

enhancing the functionality of wearable devices for

dementia care.

This paper is organized in seven sections. Section

2 presents the state-of-the-art advancements in AI and

wearable technologies for dementia care, focusing on

activity recognition and the challenges of developing

effective models. Section 3 outlines the methodology

employed, including dataset selection, preprocessing

steps, and the AI models evaluated. Section 4

discusses the datasets analysed in this study,

emphasizing sensor types, participant demographics,

recording durations, and recorded activities. Section

5 provides a detailed analysis of model performance

across various datasets, highlighting the strengths and

limitations of different AI approaches. Section 6

discusses the implications of the findings, challenges

encountered, and recommendations for future

research. Finally, Section 7 concludes the paper,

summarizing key insights and proposing directions

for advancing AI-powered wearable technologies in

dementia care.

2 STATE OF THE ART

In the field of dementia care, wearables and Artificial

Intelligence (AI) are becoming increasingly

significant, offering solutions for monitoring (Husebo

et al., 2020), early diagnosis (Godfrey et al., 2019;

Sashima, 2022), and improved quality of life

(Wilmink et al., 2020).

2.1 Activity Recognition in Dementia

Care

Activity Recognition are crucial in improving care for

individuals with dementia. Through the use of

wearable sensors and machine learning algorithms,

these systems provide valuable insights into patients’

daily activities, supporting caregivers in addressing

deficits and improving care delivery (K. J. Kim et al.,

WHC 2025 - Special Session on Wearable HealthCare

996

2009). For instance, deviations from typical behaviour

can be identified by analysing parameters such as

time, location, and activity duration (Gayathri et al.,

2015).

2.2 AI in Activity Recognition

AI plays a pivotal role in processing sensor data,

identifying patterns, and recognizing activities. By

extracting both spatial and temporal features, AI

enhances the accuracy and efficiency of activity

recognition systems (Khan et al., 2022).

AI algorithms can process streaming data in real-

time, enabling dynamic recognition of human

activities. For example, sliding window-based

approaches combined with time-decay factors have

been shown to improve recognition accuracy in

dynamic environments, ensuring reliability even in

complex, real-world scenarios (Krishnan & Cook,

2014).

One key advantage of AI-driven systems is their

ability to identify new or unexpected activities not

encountered during training. This adaptability

enhances the system’s relevance to real-world

conditions, making it better suited for the

unpredictable nature of dementia care (Leite et al.,

2021).

Different AI techniques offer unique benefits, and

their applicability depends on factors like the

complexity of activities, the nature of sensor data, and

the amount of labelled data available.

2.2.1 Machine Learning Techniques

Machine learning (ML), a subset of AI, is widely used

in activity recognition systems for its ability to learn

patterns from data and generalize them to new

scenarios.

Among traditional ML approaches, Support

Vector Machines (SVMs) are particularly effective

for tasks involving well-separated classes, achieving

high accuracy in identifying activities such as

walking, running, and sitting from wearable data (L.

Cheng et al., 2017). Similarly, Random Forest (RF) is

known for its resilience to noise and ability to classify

multiple activities (Badawi et al., 2019), while K-

Nearest Neighbors (KNN) is most suitable for

datasets with fewer classes, working by comparing

feature similarity (Murad & Pyun, 2017). Logistic

Regression, with its computational efficiency and

interpretability, is commonly used in binary

classification tasks such as distinguishing between

active and inactive states in wearable systems(Rabbi

et al., 2021).

When it comes to Deep Learning approaches,

Convolutional Neural Networks (CNNs) are effective

in learning complex patterns from raw data, ideally

for special feature extraction from sensor data (Khan

et al., 2022). Recurrent Neural Networks (RNNs)

such as Long Short-Term Memory (LSTM) networks

can capture temporal dependencies in sequential

activity data (Murad & Pyun, 2017). Gated Recurrent

Units (GRUs), a variation of RNNs, are particularly

effective for wearable time-series data, as they can

predict transitions between complex activities, such

as alternating sitting and standing, based on

accelerometer and gyroscope readings (Erdaş &

Güney, 2021).

Boosting algorithms like Extreme Gradient

Boosting (XGBoost) and Light Gradient Boosting

Machine (LightGBM) have also proven effective for

wearable data. XGBoost is optimized for speed and

scalability, making it suitable for identifying key

sensor contributions and managing missing data in

activity monitoring applications (Ge, 2023). On the

other hand, LightGBM is particularly advantageous

for processing large datasets and handling real-time

data streams, making it an excellent choice for

latency-critical tasks like fall detection and abnormal

movement tracking (S. T. Cheng, 2017).

Each of these techniques offers unique benefits,

and their applicability depends on factors like the

complexity of activities, the nature of sensor data, and

the volume of labelled data available.

2.2.2 Challenges

Wearable devices collected data is often noisy due to

movement artifacts, environmental interference, or

device misplacement. To address this, techniques

such as feature disentanglement are employed to

separate meaningful activity patterns from irrelevant

noise (Su et al., 2022).

While deep learning methods like CNNs and

LSTMs networks have proven effective for activity

recognition, the integration of data from multiple

sensors, such as accelerometers, gyroscopes, and

heart rate monitors, introduces significant

complexity. This fusion increases computational

demands, posing challenges for both model training

and deployment (Nweke et al., 2018).

Another limitation is the difficulty in generalizing

activity recognition models across users with varying

physical characteristics or across different

environmental contexts. This often leads to reduced

performance in real-world applications, highlighting

the need for models that are adaptable to diverse

scenarios (Lara & Labrador, 2013).

Powered Wearable Technologies for Dementia Care: Evaluating Activity Recognition Models and Dataset Challenges

997

Wearables are limited by battery life and

processing power, making energy-efficient AI

models essential for real-time activity recognition,

hybrid ensemble models and feedback-based

adaptive sampling have been proposed to balance

energy efficiency with recognition accuracy (Min &

Cho, 2011).

3 METHODOLOGY

To train AI models for activity recognition, suitable

datasets are essential. For our study, the ideal dataset

includes data from accelerometers, gyroscopes, and

heart rate sensors, as these sensors provide crucial

insights into movement, orientation and physiological

responses. The dataset should feature labelled activity

data to facilitate supervised learning and encompass

range of activities – such as walking, running, laying,

sleeping, eating, and hygiene – that are particularly

relevant to dementia care. Additionally, it is crucial

for the dataset to include diverse participants,

specifically older adults of both genders, as dementia

predominantly affects this demographic. The selected

datasets were used to train various AI models,

including Support Vector Machines (SVM), Random

Forest (RF), K-Nearest Neighbors (KNN),

Convolutional Neural Networks (CNN), Recurrent

Neural Networks (RNN), Long Short-Term Memory

(LSTM), Extreme Gradient Boosting (XGBoost),

Logistic Regression, Light Gradient Boosting

Machine (LightGBM), and Gated Recurrent Units

(GRU). Model performance was evaluated using

metrics such as precision, recall, and F1-score for

each class, along with overall accuracy, macro-

average, weighted-average, and a confusion matrix to

analyse classification outcomes.

4 DATASETS

To develop and evaluate AI models for activity

recognition in dementia care, this study analysed 30

publicly available datasets commonly used in

wearable activity and health monitoring research.

These datasets were selected to explore their

applicability in detecting activities relevant to

dementia, such as walking, eating, sleeping, and fall-

related movements.

4.1 Sensors

In this study, a total of 30 datasets were analysed to

examine the types of sensors utilized for wearable

activity and health monitoring systems.

Among these datasets, the most used sensors were

accelerometers (ACC), which appeared in 22 datasets

when combining data from wrist-mounted, chest-

mounted, and general-purpose accelerometers.

Accelerometers are foundational in wearable systems

due to their ability to capture motion data, making

them versatile for applications such as activity

detection, fall monitoring, and posture analysis.

Gyroscopes (GYR) were the second most frequent

sensor type, featured in 16 datasets. These sensors

provide rotational motion data, complementing

accelerometers in capturing more detailed movement

patterns, especially for activities involving complex or

rotational motions.

Heart rate (HR) sensors were present in 8 datasets,

often used for applications requiring cardiovascular

activity tracking.

Other sensors, such as temperature (TEMP)

sensors and electrocardiograms (ECG), were found in

6 datasets each, highlighting their role in physiological

and health monitoring. Electrodermal activity (EDA)

sensors, which measure skin conductance changes and

are used for stress detection, were utilized in 5

datasets. Additionally, respiration (RESP), oxygen

saturation (SpO2), and photoplethysmography (PPG)

sensors were included in a smaller number of datasets,

primarily focusing on health monitoring and specific

physiological applications. Figure 1 provides a visual

representation of the number of datasets utilizing each

sensor type. This analysis underscores the importance

Figure 1: Frequency of Sensor Types Used in the Datasets

- Accelerometer (ACC, m/s²); Gyroscope (GYR, rad/s);

Heart Rate (HR, bpm); Temperature (TEMP, °C);

Electrocardiography (ECG, mV); Electrodermal Activity

(EDA, μS); Photoplethysmogram (PPG); Respiration

(RESP, bpm); Oxygen Saturation (SpO2, %). The vertical

axis represents the count of datasets containing each sensor

type.

WHC 2025 - Special Session on Wearable HealthCare

998

of accelerometers and gyroscopes as fundamental

components in wearable systems for activity

detection. However, the integration of physiological

sensors, such as HR combines physical and health data

to develop more comprehensive monitoring systems.

4.2 Participants Demographics

The demographic composition of participants in

wearable datasets is crucial for developing activity

recognition models tailored to elderly individuals

with dementia. Since dementia predominantly affects

older populations, datasets used for model training

must reflect the physiological and behavioural

characteristics of this demographic. Discrepancies in

age, gender, or participant diversity can lead to

models that fail to generalize effectively to real-world

applications in dementia care.

The total number of participants across datasets

varies significantly. Larger datasets such as the

Parkinsons Disease Smartwatch dataset (PADS)

(Julian et al., 2024), with 469 participants, and the

Sleep Health and Lifestyle Dataset (Tharmalingam,

2023), with 374 participants. In contrast, smaller

datasets like the OPPORTUNITY Activity

Recognition (Roggen, Alberto, et al., 2010), and the

Smartwatch Heart Rate Data Dataset (Biswas &

Ashili, 2023), involve only a single participant.

The age range of participants varies, with most of

the datasets focus on adults with a median age of 20-

30 years, some of them being the Physical Activity

Monitoring Dataset PAMAP2 (Attila, 2012),

Objectively Recognizing Eating Behaviour and

Associated Intake (OREBA) (Rouast et al., 2020),

and Annotated Wearable Multimodal Biosignals

recorded during Everyday Life Activities in

Naturalistic Environments (ScientISST MOVE)

(Saraiva et al., 2024), Figure 2.

Datasets targeting specific populations, like the

elderly, include an older demographic. For example,

the Wrist Elderly Daily Activity and Fall Dataset

(WEDA-FALL) (Marques, 2022) has participants

with a mean age of 50.48 years, while the Long-Term

Movement Monitoring Database (Ihlen et al., 2015)

includes participants aged 65–78 years.

Several datasets report near-equal gender

representation. For example, the Sleep Health and

Lifestyle Dataset (Tharmalingam, 2023) has a 51%

male and 49% female split, enhancing model fairness

and applicability across both genders. While others,

such as the Wearable Stress and Affect Detection

(WESAD) (Schmidt et al., 2018), are male

dominated, with only 20% female participants, such

biases may lead to models that underperform for

underrepresented groups.

4.3 Duration of Recordings

To accurately model and monitor daily routines,

datasets must capture a representative snapshot of an

individual's activities throughout the day. Short

recordings may only provide fragmented insights,

while longer recordings enable a view of patterns,

deviations, and anomalies. Extended datasets are

particularly important for identifying changes in

routines, such as prolonged inactivity, increased

wandering, or disruptions in sleep patterns, which are

critical indicators for dementia care.

For example, the Long-Term Movement

Monitoring Database (Ihlen et al., 2015) provides 3

days of continuous data, the Smartwatch heart rate

data (Biswas & Ashili, 2023), includes 1 month of

data, and the Clemson All-day Dataset (CAD)

(Hoover, 2020) spans for 354 days, making these

datasets ideal for tracking routine behaviours over

multiple days.

In contrast, the remaining 25 datasets in the

review capture data for durations shorter than 24

hours, limiting their applicability for in-depth routine

analysis, Figure 3.

Figure 2: Age Distributions in Research Databases.

Powered Wearable Technologies for Dementia Care: Evaluating Activity Recognition Models and Dataset Challenges

999

Figure 3: Duration Breakdown of Data Recordings.

4.4 Recorded Activities

Monitoring daily routines requires datasets that

contain a wide range of activities typically performed

throughout the day. This includes basic activities like

walking, sitting, sleeping and eating, as well as more

complex or irregular behaviours such as hygiene

routines, wandering, or falling movements, as seen on

Table 1.

Walking is the most frequently recorded activity

across the datasets, appearing in 40% (12 datasets) of

the reviewed datasets. Eating activities, essential for

monitoring nutritional health and independence, are

labelled in 20% (6 datasets). Sitting and sleeping

activities are recorded in 13.3% (4 datasets) each,

highlighting a focus on sedentary and rest-related

behaviours.

In addition to these activities several datasets (11)

include labels for miscellaneous activities that

provide unique insights into daily routines and

specific behaviours. For instance, the

OPPORTUNITY dataset (Roggen, Alberto, et al.,

2010) includes activities like "opening a door" and

"drinking water," used for recognizing fine-grained

motor skills. The PAMAP2 dataset (Attila, 2012)

features labelled activities such as "ascending stairs",

“descending stairs”, “watching TV”, “standing” and

"house cleaning," capturing more dynamic and

context-specific movements, particularly useful for

training models that aim to recognize household

activities. The WEDA-FALL dataset (Marques,

2022) focuses on fall-related activities and recovery

movements, critical for fall detection systems,

similarly, the Long-Term Movement Monitoring

Database (Ihlen et al., 2015)focuses on prolonged

activity tracking, offering continuous movement data

collected over several days from older adults. Other

datasets, like the ScientISST MOVE dataset (Saraiva

et al., 2024), include transitions between activities

such as "standing-to-sitting" and "sitting-to-lying,"

relevant for understanding changes in posture or

transitions that may indicate health issues. The

OREBA dataset (Rouast et al., 2020) targets eating

behaviours by providing multimodal data for

recognizing eating gestures and associated intake,

contributing to dietary monitoring. The Sleep Health

and Lifestyle Dataset (Tharmalingam, 2023) on the

other hand, focuses on sleep patterns and lifestyle

habits capturing detailed sleep metrics such as

duration, efficiency, and disruptions, which are vital

for understanding circadian rhythm irregularities

often observed in dementia patients.

The WESAD (Schmidt et al., 2018), are focused

on stress recognition, providing labelled data for

different emotional states, including stress,

amusement, and neutral conditions. These datasets

often integrate physiological signals, such as heart

rate variability (HRV), electrodermal activity (EDA),

and respiratory patterns, alongside motion data.

Each of these datasets provides unique insights

and data characteristics that enrich the development

of AI models, enabling more comprehensive and

accurate activity recognition systems tailored to

dementia care.

5 MODEL PERFORMANCE

ANALYSIS

To evaluate the effectiveness of machine learning

models in activity recognition, we analysed the

performance of multiple algorithms across various

datasets. This section summarizes the results obtained

for each dataset. Performance metrics, including

precision, recall, and F1-score, were evaluated for

models like Random Forest, K-Nearest Neighbors

(KNN), and Gradient Boosting.

5.1 MMASH Dataset

The MMASH (Multimodal Activity Recognition in

Smart Home Environments) (Rossi et al., 2020)

dataset is a comprehensive dataset designed for

activity recognition research. It includes data from

multiple sensor types such as accelerometers,

gyroscopes, magnetometers, and physiological

sensors. Covering a wide range of activities, including

basic actions like sitting, walking, and lying down, as

well as complex activities such as eating or

performing household tasks.

WHC 2025 - Special Session on Wearable HealthCare

1000

Table 1: Activity Types Captured in Wearable Activity Recognition Datasets.

Ref./ Activity Walk Sit Sleep Eat Fall Miscellaneous

(Saraiva et al., 2024) x x

(Guy et al., 2024) x

(Nicoomanesh, 2024) x x

(Julian et al., 2024)

(Godzwon, 2024)

(Tharmalingam, 2023) x x x

(Grimaldi et al., 2023) x x x

(Mehrgardt et al., 2022)

(Amin et al., 2022)

(Marques, 2022) x x x

(Karas et al., 2021) x x

(Rossi et al., 2020) x x x x x

(Fuller, 2020) x x x

(Hoover, 2020) x

(Rouast et al., 2020) x

(J. Kim, 2020) x x

(Walch, 2019) x x

(Kyritsis et al., 2019) x

(Schmidt et al., 2018)

(Jafarnejad, 2018)

(Bhogal & Mani, 2017)

(Jarchi & Casson, 2017) x x

(Kyritsis et al., 2017) x

(Ihlen et al., 2015) x

(Banos et al., 2014) x x x x x

(Attila, 2012) x x

(Roggen, Calatroni, et al., 2010)

(Jager et al., 2003)

(Moody & Mark, 2001)

Both the XGBoost and LigthGBM models

consistently achieved higher accuracy, and F1-scores

compared to other models. For instance, the

XGBoost, demonstrated strong generalization with

higher overall accuracy and consistently balanced

precision and recall across activities, including

underrepresented classes. With the LightGBM

outperforming in handling imbalanced data,

particularly for rare activities.

The Random Forest and Gradient Boosting

models performed best for the generalized activities

with large support values, such as sitting. However,

the model struggled with specific or underrepresented

activities such as large screen usage and sleeping.

Comparing to other models, the KNN

underperformed, presenting low precision and recall

values for most activities, due to the class imbalance.

5.2 ScientISST Dataset

The ScientISST Dataset (Saraiva et al., 2024) is a

comprehensive and multimodal dataset designed for

human activity and gesture recognition. It is

particularly valuable for developing and evaluating

machine learning models in scenarios requiring high

Powered Wearable Technologies for Dementia Care: Evaluating Activity Recognition Models and Dataset Challenges

1001

precision and robustness, such as healthcare,

wearable technology, and human-computer

interaction.

The KNN and Random Forest models

outperformed the other models, achieving nearly

perfect results across most activities for accuracy,

precision and F1-score, apart from the less frequent

activities.

The CNN and GRU models performed well in

recognizing frequent and sequential activities. CNN

excelled at extracting spatial features, achieving high

F1-scores for structured tasks like "Run" (0.95) and

dynamic movements like "Jumps" (0.74). GRU

effectively captured temporal dependencies, making

it ideal for activities with transitions, such as "Lift"

(F1-score: 0.83) and "Run" (F1-score: 0.93). The

SVM model showed strong performance for well-

separated and frequent activities. The MLP model

demonstrated consistent performance for frequent

and distinct activities, achieving an F1-score of 0.97

for "Run."

All models, however, struggled with nuanced

gestures like "Greetings" and “jump” presenting

reduced precision and recall.

5.3 PAMAP2 Dataset

The PAMAP2 (Physical Activity Monitoring for

Aging People 2) (Attila, 2012) dataset is a collection

of data designed to facilitate the development and

evaluation of activity recognition algorithms. This

dataset is widely utilized in the field of wearable

computing and health monitoring, particularly for

applications involving elderly care and fitness

tracking.

The Random Forest and XGBoost models exhibit

stellar performance, with nearly perfect precision,

recall, and F1-scores close to 100% across a wide

range of activities. This performance indicates their

robust predictive capabilities and adaptability in

managing diverse data types, particularly in complex

activity scenarios such as 'nordic walking' and

'cycling'.

In a similar way, LightGBM, a gradient boosting

model optimized for speed and reduced memory

usage, offers substantial advantages for real-time

activity recognition applications. Combining the

robust framework of gradient boosting with

enhancements designed to improve processing speed

and efficiency, making it competitive for applications

where quick response times are crucial.

The KNN model showed moderate performance

with an overall accuracy of 91%. While it performed

well on frequent activities like "Sitting" (F1-score:

0.98) and "Cycling" (F1-score: 0.96), its performance

dropped for more complex and underrepresented

activities, such as "Ascending Stairs" (F1-score: 0.74)

and "Descending Stairs" (F1-score: 0.72).

The Logistic Regression model shows varying

performance across different activities, reflecting

some fundamental limitations in handling complex,

multiclass problems. While it performs commendably

in activities with clear distinctions, such as 'lying' and

'ironing', it faces challenges in activities requiring

nuanced differentiation, such as 'standing' versus

'sitting'. This variation highlights the need for

sophisticated feature engineering or advanced data

preprocessing to bolster its effectiveness in more

complex scenarios.

6 DISCUSSION

The findings of this study provide important insights

into the development and optimization of AI models

and wearable technologies for activity recognition in

dementia care. However, a significant challenge is the

lack of comprehensive datasets tailored to the unique

requirements of this domain. Current datasets

predominantly feature younger adults, offering

limited representation of older individuals who are

most affected by dementia, thereby reducing the

applicability of AI models to the intended population.

Additionally, existing datasets often fail to cover the

full range of activities relevant to dementia care, such

as hygiene routines, eating behaviors, wandering, and

fall-related movements. This lack of comprehensive

activity coverage limits the ability of AI systems to

monitor the complex behaviors associated with

dementia effectively.

Another limitation is the prevalence of short-

duration recordings, which are insufficient for

analyzing long-term activity patterns and deviations

that are critical for dementia monitoring and detecting

changes in routine or health status. Furthermore, most

datasets are collected in controlled environments,

which fail to capture the complexity and variability of

real-world settings, such as homes or assisted living

facilities where dementia patients typically reside.

This discrepancy reduces the robustness and

generalizability of models trained on such data.

Additionally, many datasets suffer from significant

class imbalances, with underrepresented activities

leading to poor model performance for these specific

behaviors. Addressing these limitations is essential to

develop AI-driven wearable solutions that are

accurate, robust, and capable of meeting the practical

needs of dementia care.

WHC 2025 - Special Session on Wearable HealthCare

1002

Future datasets should prioritize the inclusion of

elderly participants representing diverse genders and

cognitive stages, ensuring that the data accurately

reflects the population most affected by dementia.

These datasets should aim to capture a comprehensive

range of activities, including eating, hygiene routines,

wandering, fall recovery, and activity transitions, as

well as nighttime behaviours. To enable a deeper

understanding of daily routines and their variations, it

is essential to include long-duration recordings

spanning multiple days or even weeks. Collecting

data in naturalistic environments, such as homes or

care facilities, will further enhance the validity of the

datasets and significantly improve the robustness and

generalizability of AI models developed for dementia

care.

The performance evaluation of the AI models in

this study highlights the strengths and limitations of

different approaches for activity recognition in

dementia care. Models such as Random Forest (RF),

XGBoost, and LightGBM consistently demonstrated

robust performance, excelling in handling class

imbalances and accurately recognizing well-defined

activities like walking, running, and sitting. Their

resilience to noisy data and ability to generalize

across common activity classes make them reliable

choices for general activity monitoring.

However, despite these promising results, all

models faced challenges in identifying less frequent

and more nuanced activities, such as eating or

transitions, due to limitations in dataset quality and

class imbalances. The underrepresentation of these

critical activities in the datasets hindered model

performance, leading to reduced precision and recall

for these classes. Moreover, the prevalence of short-

duration recordings further constrained the models'

ability to analyze long-term activity patterns, limiting

their effectiveness in detecting behavioral trends and

anomalies essential for dementia care.

These findings underscore the necessity of

selecting and tailoring models based on specific

application requirements. For general activity

recognition tasks, tree-based models like XGBoost

and LightGBM offer strong performance and

efficiency. In contrast, deep learning approaches,

such as CNNs and GRUs, are better suited for tasks

that require detailed temporal and spatial analysis,

particularly when handling complex or transitional

activities. Addressing dataset limitations, including

activity coverage, class balance, and recording

duration, will be critical for enhancing model

performance and ensuring their practical applicability

in real-world dementia care scenarios.

7 CONCLUSION

This study highlights the potential of AI models and

wearable technologies for activity recognition in

dementia care, demonstrating the strengths of tree-

based models like Random Forest, XGBoost, and

LightGBM in handling class imbalances and

recognizing common activities, as well as the

capabilities of deep learning models such as CNNs

and GRUs in capturing complex patterns and

transitions. However, significant challenges remain,

including the lack of comprehensive datasets that

adequately represent the elderly population,

encompass a diverse range of activities, and provide

long-duration recordings in real-world environments.

These limitations reduce the generalizability and

effectiveness of AI models in detecting nuanced

behaviors and long-term activity patterns critical for

dementia monitoring.

To address these gaps, future research should

focus on developing tailored datasets with enhanced

demographic diversity, extended recordings, and

realistic environmental contexts. Combining

traditional and deep learning models into hybrid

approaches can further optimize performance, while

energy-efficient AI solutions will ensure real-time

monitoring capabilities for wearable devices. By

overcoming these challenges, AI-powered wearable

technologies can play a transformative role in

dementia care, enabling accurate activity recognition,

early intervention, and improved quality of life for

patients while reducing the burden on caregivers.

ACKNOWLEDGEMENTS

This project was funded through the Foundation for

Science and Technology (FCT) under the projects

UIDB/05549/2020(DOI:10.54499/UIDB/05549/202

0), UIDP/05549/2020 (DOI:10.54499/UIDP/05549/

2020), LASILA/P/0104/2020, and CEECINST/

00039/2021. This work was also funded by the

Innovation Pact HfFP–Health From Portugal, co-

funded from the” Mobilizing Agendas for Business

Innovation” of the ”Next Generation EU” program of

Component 5 of the Recovery and Resilience Plan

(RRP), concerning ”Capitalization and Business

Innovation”, under the Regulation of the Incentive

System ”Agendas for Business Innovation”.

Powered Wearable Technologies for Dementia Care: Evaluating Activity Recognition Models and Dataset Challenges

1003

REFERENCES

Amin, M. R., Wickramasuriya, D., & Faghih, R. T. (2022).

A Wearable Exam Stress Dataset for Predicting

Cognitive Performance in Real-World Settings

[Dataset]. PhysioNe. https://doi.org/https://doi.org/

10.13026/kvkb-aj90

Attila, R. (2012). PAMAP2 Physical Activity Monitoring

[Dataset]. UCI Machine Learning Repository.

https://doi.org/https://doi.org/10.24432/C5NW2H.

Badawi, A. A., Al-Kabbany, A., & Shaban, H. (2019).

Multimodal Human Activity Recognition from

Wearable Inertial Sensors Using Machine Learning.

402–407. https://doi.org/10.1109/iecbes.2018.8626737

Banos, O., Garcia, R., Holgado-Terriza, J. A., Damas, M.,

Pomares, H., Rojas, I., Saez, A., & Villalonga, C.

(2014). mHealthDroid: A Novel Framework for Agile

Development of Mobile Health Applications.

Bhogal, A. S., & Mani, A. R. (2017). Pattern analysis of

oxygen saturation variability in healthy individuals:

Entropy of pulse oximetry signals carries information

about mean oxygen saturation. Frontiers in Physiology,

8(AUG). https://doi.org/10.3389/fphys.2017.00555

Biswas, N., & Ashili, S. (2023). Smartwatch heart rate

data. IEEE Dataport. https://doi.org/https://dx.doi.org/

10.21227/y4ek-d516

Cheng, L., Guan, Y., Zhu, K., & Li, Y. (2017). Recognition

of Human Activities using Machine Learning Methods

with Wearable Sensors. IEEE.

Cheng, S. T. (2017). Dementia Caregiver Burden: a

Research Update and Critical Analysis. In Current

Psychiatry Reports (Vol. 19, Issue 9). Current Medicine

Group LLC 1. https://doi.org/10.1007/s11920-017-

0818-2

Cote, A. C., Phelps, R. J., Kabiri, N. S., Bhangu, J. S., &

Thomas, K. (2021). Evaluation of Wearable

Technology in Dementia: A Systematic Review and

Meta-Analysis. In Frontiers in Medicine (Vol. 7).

Frontiers Media S.A. https://doi.org/10.3389/

fmed.2020.501104

Cullen, A., Mazhar, M. K. A., Smith, M. D., Lithander, F.

E., Breasail, M., & Henderson, E. J. (2022). Wearable

and Portable GPS Solutions for Monitoring Mobility in

Dementia: A Systematic Review. Sensors, 22(9).

https://doi.org/10.3390/s22093336

Erdaş, Ç. B., & Güney, S. (2021). Human Activity

Recognition by Using Different Deep Learning

Approaches for Wearable Sensors. In Neural

Processing Letters (Vol. 53, Issue 3, pp. 1795–1809).

Springer. https://doi.org/10.1007/s11063-021-10448-3

Fuller, D. (2020). Replication Data for: Using machine

learning methods to predict physical activity types with

Apple Watch and Fitbit data using indirect calorimetry

as the criterion. Harvard Dataverse.

https://doi.org/https://doi.org/10.7910/DVN/ZS2Z2J

Gayathri, K. S., Elias, S., & Ravindran, B. (2015).

Hierarchical activity recognition for dementia care

using Markov Logic Network. Personal and

Ubiquitous Computing, 19(2), 271–285.

https://doi.org/10.1007/s00779-014-0827-7

Ge, R. (2023). XGBoost-Based Human Activity

Recognition Algorithm using Wearable Smart Devices.

Applied and Computational Engineering, 2(1), 352–

358. https://doi.org/10.54254/2755-2721/2/20220514

Godfrey, A., Brodie, M., van Schooten, K. S.,

Nouredanesh, M., Stuart, S., & Robinson, L. (2019).

Inertial wearables as pragmatic tools in dementia. In

Maturitas (Vol. 127, pp. 12–17). Elsevier Ireland Ltd.

https://doi.org/10.1016/j.maturitas.2019.05.010

Godzwon, I. (2024). FitBit Heart Rate [Dataset]. OpenML.

https://www.openml.org/search?type=data&status=act

ive&id=46103&sort=runs

Grimaldi, E., Vigneri, D., Di Poce, G., & Grieco, N. (2023).

Falls vs Normal Activities [Dataset]. Kaggle.

https://www.kaggle.com/datasets/enricogrimaldi/falls-

vs-normal-activities

Guy, E. F. S., Isaac, F., Jaimey Anne, C., Trudy, C. der K.,

Rongqing, C., Jennifer, K., Knut, M., & James

Geoffrey, C. (2024). Respiratory and heart rate

monitoring dataset from aeration study [Dataset]].

PhysioNet. https://doi.org/https://doi.org/10.13026/

e4dt-f689

Hoover, A. (2020). Clemson All-day Dataset (CAD)

[Dataset]. Clemson University. https://cecas.

clemson.edu/~ahoover/allday/

Husebo, B. S., Heintz, H. L., Berge, L. I., Owoyemi, P.,

Rahman, A. T., & Vahia, I. V. (2020). Sensing

technology to facilitate behavioral and psychological

symptoms and to monitor treatment response in people

with dementia: A systematic review. In Frontiers in

Pharmacology (Vol. 10). Frontiers Media S.A.

https://doi.org/10.3389/fphar.2019.01699

Ihlen, E. A. F., Weiss, A., Helbostad, J. L., & Hausdorff, J.

M. (2015). The Discriminant Value of Phase-

Dependent Local Dynamic Stability of Daily Life

Walking in Older Adult Community-Dwelling Fallers

and Nonfallers. BioMed Research International, 2015,

1–11. https://doi.org/10.1155/2015/402596

Jafarnejad, S. (2018). An Open Dataset for Human Activity

Analysis [Dataset]. Kaggle. https://www.kaggle.com/

datasets/sasanj/human-activity-smart-devices

Jager, F., Taddei, A., Moody, G. B., Emdin, M., Antolic,

G., Dorn, R., Smrdl, A., Marchesi, C., & Mark, R. G.

(2003). Long Term ST Database . PhysioNet.

Jarchi, D., & Casson, A. J. (2017). Description of a database

containing wrist PPG signals recorded during physical

exercise with both accelerometer and gyroscope

measures of motion. Data, 2(1). https://doi.org/

10.3390/data2010001

Julian, V., Alexander, B., Lucas, P., Catharina, van A.,

Michael, F., & Tobias, W. (2024). PADS - Parkinsons

Disease Smartwatch dataset. PhysioNet.

Karas, M., Urbanek, J., Crainiceanu, C., Harezlak, J., &

Fadel, W. (2021). Labeled raw accelerometry data

captured during walking, stair climbing and driving

[Dataset]. PhysioNet. https://doi.org/10.13026/51h0-

a262

Khan, I. U., Afzal, S., & Lee, J. W. (2022). Human activity

recognition via hybrid deep learning based model.

Sensors, 22(1). https://doi.org/10.3390/s22010323

WHC 2025 - Special Session on Wearable HealthCare

1004

Kim, J. (2020). UCI-HAR [Dataset]. Kaggle.

https://kaggle.com/competitions/uci-har

Kim, K. J., Hassan, M. M., Na, S., & Huh, E. N. (2009).

Dementia wandering detection and activity recognition

algorithm using tri-axial accelerometer sensors.

Proceedings of the 4th International Conference on

Ubiquitous Information Technologies and

Applications, ICUT 2009. https://doi.org/

10.1109/ICUT.2009.5405672

Krishnan, N. C., & Cook, D. J. (2014). Activity recognition

on streaming sensor data. Pervasive and Mobile

Computing, 10(PART B), 138–154.

https://doi.org/10.1016/j.pmcj.2012.07.003

Kyritsis, K., Diou, C., & Delopous, A. (2017). Wrist-

mounted IMU data towards the investigation of in-meal

human eating behavior - the Food Intake Cycle (FIC)

dataset. IEEE Journal of Biomedical and Health

Informatics.

Kyritsis, K., Diou, C., & Delopous, A. (2019). Wrist-

mounted IMU data towards the investigation of free-

living human eating behavior - the Free-living Food

Intake Cycle (FreeFIC) dataset. NIAID Data

Ecosystem. https://doi.org/10.5281/zenodo.4421860

Lara, Ó. D., & Labrador, M. A. (2013). A survey on human

activity recognition using wearable sensors. IEEE

Communications Surveys and Tutorials, 15(3), 1192–

1209. https://doi.org/10.1109/

SURV.2012.110112.00192

Lavretsky, H. (2005). Stress and Depression in Informal

Family Caregivers of Patients with Alzheimer’s

Disease. Aging Health, 1(1), 117–133.

https://doi.org/10.2217/1745509x.1.1.117

Leite, B., Abdalrahman, A., Castro, J., Frade, J., Moreira,

J., & Soares, C. (2021). Novelty detection in physical

activity. ICAART 2021 - Proceedings of the 13th

International Conference on Agents and Artificial

Intelligence, 2, 859–865. https://doi.org/10.5220/

0010254908590865

Lin, C. C., Lin, P. Y., Lu, P. K., Hsieh, G. Y., Lee, W. L.,

& Lee, R. G. (2008). A healthcare integration system

for disease assessment and safety monitoring of

dementia patients. IEEE Transactions on Information

Technology in Biomedicine, 12(5), 579–586.

https://doi.org/10.1109/TITB.2008.917914

Lindeza, P., Rodrigues, M., Costa, J., Guerreiro, M., &

Rosa, M. M. (2024). Impact of dementia on informal

care: a systematic review of family caregivers’

perceptions. In BMJ Supportive and Palliative Care

(Vol. 14, Issue e1, pp. E38–E49). BMJ Publishing

Group. https://doi.org/10.1136/bmjspcare-2020-

002242

Marques, J. (2022). Wrist Elderly Daily Activity and Fall

Dataset (WEDA-FALL).

Mehrgardt, P., Khushi, M., Poon, S., & Withana, A. (2022).

Pulse Transit Time PPG Dataset. PhysioNet.

https://doi.org/https://doi.org/10.13026/jpan-6n92

Min, J.-K., & Cho, S.-B. (2011). Activity Recognition based

on Wearable Sensors Using Selection/Fusion Hybrid

Ensemble. IEEE.

Moody, G. B., & Mark, R. G. (2001). The Impact of the

MIT-BIH Arrhythmia Database History, Lessons

Learned, and Its Influence on Current and Future

Databases.

Murad, A., & Pyun, J. Y. (2017). Deep recurrent neural

networks for human activity recognition. Sensors

(Switzerland), 17(11). https://doi.org/10.3390/

s17112556

Nicoomanesh, A. (2024). FitBit Fitness Tracker Data

[Dataset]. Kaggle. https://www.kaggle.com/datasets/

arashnic/fitbit/data

Nweke, H. F., Teh, Y. W., Al-garadi, M. A., & Alo, U. R.

(2018). Deep learning algorithms for human activity

recognition using mobile and wearable sensor

networks: State of the art and research challenges. In

Expert Systems with Applications (Vol. 105, pp. 233–

261). Elsevier Ltd.

https://doi.org/10.1016/j.eswa.2018.03.056

Rabbi, J., Fuad, Md. T. H., & Awal, Md. A. (2021). Human

Activity Analysis and Recognition from Smartphones

using Machine Learning Techniques. http://arxiv.org/

abs/2103.16490

Rocha, I. C., Arantes, M., Moreira, A., Vilaça, J. L., Morais,

P., Matos, D., & Carvalho, V. (2024). Monitoring

Wearable Devices for Elderly People with Dementia: A

Review. In Designs (Vol. 8, Issue 4). Multidisciplinary

Digital Publishing Institute (MDPI).

https://doi.org/10.3390/designs8040075

Roggen, D., Alberto, C., Long-Van, N.-D., Ricardo, C., &

Hesam, S. (2010). OPPORTUNITY Activity

Recognition [Dataset]. UCI Machine Learning

Repository. https://doi.org/10.24432/C5M027.

Roggen, D., Calatroni, A., Rossi, M., Holleczek Thomas,

Forster, K., Troster, G., Lukowicz, P., Bannach, D., &

Pirkl, G. (2010). Collecting complex activity datasets in

highly rich networked sensor environments.

Rossi, A., Da Pozzo, E., Menicagli, D., Tremolanti, C.,

Priami, C., Sirbu, A., Clifton, D., Martini, C., &

Morelli, D. (2020). Multilevel Monitoring of Activity

and Sleep in Healthy People [Database]]. PhysioNet.

https://doi.org/https://doi.org/10.13026/cerq-fc86

Rouast, P. V., Heydarian, H., Adam, M. T. P., & Rollo, M.

E. (2020). OReBA: A dataset for objectively

recognizing eating behavior and associated intake.

IEEE Access, 8, 181955–181963.

https://doi.org/10.1109/ACCESS.2020.3026965

Saraiva, J. A., Abreu, M., Carmo, A. S., Silva, H. P., &

Fred, A. (2024). ScientISST MOVE: Annotated

Wearable Multimodal Biosignals recorded during

Everyday Life Activities in Naturalistic Environments.

PhysioNet. https://doi.org/10.13026/hyxq-r919.

Sashima, A. (2022). Machine-learning-driven Wearable

Healthcare for Dementia: A Review of Emerging

Technologies and Challenges. 864–871.

https://doi.org/10.5220/0010973900003123

Schmidt, P., Reiss, A., Duerichen, R., & Van Laerhoven, K.

(2018). Introducing WeSAD, a multimodal dataset for

wearable stress and affect detection. ICMI 2018 -

Proceedings of the 2018 International Conference on

Powered Wearable Technologies for Dementia Care: Evaluating Activity Recognition Models and Dataset Challenges

1005

Multimodal Interaction, 400–408. https://doi.org/

10.1145/3242969.3242985

Stavropoulos, T. G., Lazarou, I., Diaz, A., Gove, D.,

Georges, J., Manyakov, N. V., Pich, E. M., Hinds, C.,

Tsolaki, M., Nikolopoulos, S., & Kompatsiaris, I.

(2021). Wearable Devices for Assessing Function in

Alzheimer’s Disease: A European Public Involvement

Activity About the Features and Preferences of Patients

and Caregivers. Frontiers in Aging Neuroscience, 13.

https://doi.org/10.3389/fnagi.2021.643135

Su, J., Wen, Z., Lin, T., & Guan, Y. (2022). Learning

disentangled behaviour patterns for wearable-based

human activity recognition. Proceedings of the ACM on

Interactive, Mobile, Wearable and Ubiquitous

Technologies, 6(1). https://doi.org/10.1145/3517252

Tharmalingam, L. (2023). Sleep Health and Lifestyle

Dataset. Kaggle. https://www.kaggle.com/datasets/

uom190346a/sleep-health-and-lifestyle-dataset/data

Walch, O. (2019). Motion and heart rate from a wrist-worn

wearable and labeled sleep from polysomnography

[Database]. PhysioNet. https://doi.org/https://doi.org/

10.13026/hmhs-py35

Wilmink, G., Dupey, K., Alkire, S., Grote, J., Zobel, G.,

Fillit, H. M., & Movva, S. (2020). Artificial

intelligence-powered digital health platform and

wearable devices improve outcomes for older adults in

assisted living communities: Pilot intervention study.

JMIR Aging, 3(2). https://doi.org/10.2196/19554

Winblad, B., Amouyel, P., Andrieu, S., Ballard, C., Brayne,

C., Brodaty, H., Cedazo-Minguez, A., Dubois, B.,

Edvardsson, D., Feldman, H., Fratiglioni, L., Frisoni,

G. B., Gauthier, S., Georges, J., Graff, C., Iqbal, K.,

Jessen, F., Johansson, G., Jönsson, L., Zetterberg, H.

(2016). Defeating Alzheimer’s disease and other

dementias: A priority for European science and society.

In The Lancet Neurology (Vol. 15, Issue 5, pp. 455–

532). Lancet Publishing Group. https://doi.org/

10.1016/S1474-4422(16)00062-4

WHC 2025 - Special Session on Wearable HealthCare

1006