SMGR-BS: Stacking Multiple Gated Recurrent Butterfly Search Model-Based Innovative AAL for Aging and Disabled Individuals

Abstract

Ambient assisted living (AAL) for aging and disabled people involves creating supportive environments that leverage technology to improve the quality of life and independence of these individuals. Traditional methods for developing AAL solutions for aging and disabled people face several challenges, such as scalability, high costs, and privacy concerns. To tackle these complexities, this article proposed a novel method named stacking multiple gated recurrent-based butterfly search (SMGR-BS) for the development of AAL for aging and disabled people. In this study, stacking multiple gated recurrent units are utilized to capture intricate temporal dependencies in sensor data, and the deep recurrent neural network extracts the features from the variety of sensor inputs. Also, the butterfly optimization algorithm with a local search strategy is employed to fine-tune the parameters and enhance the effectiveness of the SMGR-BS method. In this work, the experiments are conducted on the Mobile HEALTH dataset, and the performance evaluation of the SMGR-BS method involves analyzing its effectiveness based on evaluation metrics, namely specificity, F1-score, recall, precision, and accuracy, and comparing its performance against existing methodologies to assess its effectiveness. The experimental results illustrate the effectiveness of the SMGR-BS method for developing AAL for aging and disabled people.

Main article text

INTRODUCTION

In recent years, education, assistance, rehabilitation, and many other contexts have discovered widespread use in smart environments. Artificial intelligence (AI), machine learning, and advanced sensors make human life more comfortable and easier in several aspects of life ( Abidi et al., 2020b, 2022b). These advanced technologies offer numerous ways to assist persons with disabilities or old age people. Constructing classroom environments with innovative technologies provides support for teaching staff with disabilities and creates more impressive learning ( Thakur and Han, 2022). Correspondingly, this device supports the rehabilitation of impaired people and presents more particular tools. In home automation circumstances, many functions can be conveniently accessed remotely, and the deployment of innovative technologies enhances the quality of everyday life or allows the usage of natural interfaces ( Cicirelli et al., 2021), such as voice and gestures, to control climate entertainment systems, lighting, and appliances. The population of older people in the world is currently 964 million, and it is expected to grow at an unprecedented ratio and attain 2 billion by 2050. Due to the behavior of various rates of reduction in the mental, emotional, social, motor, and psychological skills and more obstacles such as behavioral, neurological disorders, disabilities, cognitive, and visual impairments, depending on the diversity, the aging population is corresponding with multiple and various effects on the older people ( Correia et al., 2021). The population of aging people in the world is increasing, and the population of caregivers is decreasing in looking after older people, which creates several difficulties ( Thakur and Han, 2021).

To enable the next industrial revolution and the future of healthcare, Internet of Things (IoT) technologies are considered. IoT trust is a strong infrastructure and the driving force behind its design principles, and the boundary conditions for building IoT ecosystems are limited ( Abidi et al., 2020a, 2022a). No amount of assessment can reject the value IoT will bring to industries across several platforms and the ability to maximize the growth of the global economy ( Maskeliūnas et al., 2019). This familiar prediction of growing technology obtains an exact determination of crucial and novel IoT technology’s ability. One of the available resources is ambient assisted living (AAL), which focuses on products and services embedded in the physical environment, intelligent technologies, and assistance of older adults to improve their independence, safety, autonomy, well-being, and community participation ( Queirós et al., 2017). The view of everyday circumstances enables the characterization of human–computer interaction as classified by restrained, anticipatory communications, and pervasive. The safety, health, and hygiene of the elderly can be effectively improved by AAL; early detection of potential health impairment, continuous monitoring of the condition of the elderly, and detection of hazardous events are supported by wireless sensor network architectures ( Taramasco et al., 2022). AAL technologies are used in mobile emergency response systems, video surveillance, reminders, mobility and automation assistance tools, fall detection systems, and more.

The major role of AAL is to reduce hospital expenses and enhance the living conditions of older people ( Dasios et al., 2015). Multiple types of sensors are being integrated into the AAL environment to gather broader information. The activities of the daily living dataset are difficult to identify and establish in an AAL context. It is possible to initially understand what activities the user is doing, how they are progressing, and how they are performing. Activities of daily living (ADL) control aims to differentiate between medical problems and problems caused by insufficient exercise ( Alluhaibi et al., 2023). A key contribution of AAL is to increase the survival rate of the elderly in selected environments through personalized health-monitoring devices, communication technologies, and information. It encourages research into more flexible lifestyles that evolve into creative ways of aging and looking at how care is given ( Guerra et al., 2022). This can be thought-provoking research, and searching for multiple ADLs and self-categorization can be a huge obstacle. The motivation for the study is given below.

AAL is focused on enabling people with any kind of impairment to stay independent in their own houses for as long as possible. To achieve this, data and communication technologies are utilized in various ways. Currently, with the rapid growth of the population, elderly people face more difficulties in the environment. AAL focusing on the help of elderly adults enhances their independence, safety, autonomy, and well-being. Further conditions of elderly people can be enhanced effectively by AAL, early detection of potential health impairment, and continuous monitoring of their condition supported by wireless sensor network structures. AAL helps caretakers easily aid disabled people, reduces clinical expenses, and provides long life for elderly people. This system’s merits are staying active longer and securing data. However, this method did not detect the automated healthcare, high cost, and lack of training. The study contributes to the advancement of AAL for aging and disabled people by using various techniques. The major contribution of the proposed stacking multiple gated recurrent-based butterfly search (SMGR-BS) method is explained below.

Novel method

This is a novel approach that combines stacking multiple gated recurrent unit (GRU)-based deep recurrent neural networks (RNNs) with the integration of the butterfly optimization algorithm (BOA) with a local search strategy. This unique combination of techniques offers a new perspective on the development of AAL for aging and disabled people.

Sequential data processing

AAL systems collect time-series data from different devices and sensors. Stacking GRU-based RNNs can effectively analyze these sequential data to identify anomalies and patterns in the daily activities and health status of aging and disabled people.

Efficient hyperparameter optimization

Incorporating a BOA with a local search strategy optimizes the parameters and enhances the SMGR-BS method’s performance in developing AAL for aging and disabled people.

Real-time monitoring

By incorporating multiple layers of GRUs, the SMGR-BS method can capture complex dependencies in data and make real-time predictions based on sensor inputs, ensuring timely responses to emergencies or changes in residents’ conditions.

The SMGR-BS method stands out due to its unique combination of stacking multiple GRU-based deep RNNs and integrating the BOA for parameter fine-tuning. This hybrid approach enhances the model’s ability to extract features from heterogeneous data sources while optimizing its performance, thus improving the accuracy and effectiveness of AAL systems for aging and disabled individuals.

The remaining sections of the article are arranged as follows: the existing works related to the development of AAL for aging and disabled people are described in the Literature Survey section. The proposed methodology for developing AAL for aging and disabled people is detailed in the Proposed Methodology section. The results and discussions of the SMGR-BS method are illustrated in the Experimental Results and Discussion section. Also, the conclusion and future scope of the paper are portrayed in the Discussion and Conclusion section.

LITERATURE SURVEY

This section presents the related research work in the area of AAL. Gulati and Kaur (2022) introduced an alert generation method for AAL based on a strong social IoT for aging FriendCare-AAL. Aging people face more difficulties in the environment, so this method was developed. The main aim of this method was to help the aging person living in the smart infrastructure and also naturally alert the server when any difficult situation arises. Further, the system activity was experimentally analyzed, and smart house AAL infrastructure for an aging person utilized a people action simulator. The name of the simulator was Home Sensor Simulator, and a typical dataset of humans was created. On the contrary, predicting the healthy lives of aging persons was using the naive Bayes (NB) and random forest (RF), and the accuracy was 9.2% and 83.9% for NB and RF, respectively. As a result, the system was accurately active in emergencies and lacked scalability.

Cascone et al. (2022) developed a digital twinning pepper and ambient assisted living (DTPAAL) pepper as a human–robot figure that expresses language and communicates in the environment. This method centered on relating the virtual twins with the replicas of the smart house in smart products; thus, the experiences are detailed by using the VPepper. Pepper robots contain upper limbs and palms; however, motors and actuators do not help the training to learn how to feel the product, so the digital twin metaphor was important at that time. On the contrary, the virtual robot machine learning moves freely for the digital twins. In addition, digital twins provide a caretaker for any emergencies. As a result, the limitation of the DTPAAL method was more expensive. Patro et al. (2021) established the supervised learning used in AAL for cardiovascular disease. Currently, the quick growth of the elderly population affects healthcare in the environment. Heart attacks can suddenly occur in people without any prior symptoms, leaving them without immediate first aid in critical situations. To address this issue, the patient’s real-time information is essential. Now, IoT plays a main role in the healthcare system, and it also helps patients easily detect heart disease. The aim of these methods was to make it easy for AAL to predict the disease using some classification, namely K-nearest neighbor, NB, and support vector machine (SVM). Further monitoring of the aged people’s heartbeat and prediction occur accurately. Further, the accuracy was 92%, and this method does not detect the automated smart healthcare and high cost.

Ardito et al. (2022) illustrated edge AAL methods based on clinical pathways. These methods aim to avoid medical status aggravations and also to cure patients using edge computing as well as AI methods. In the Internet of Medical Things (IoMT) approach, active logs were compiled using clinical and mobile information and mainly focused on edge construction. Several merits were presently developed by the edge construction in the house infrastructure, such as monitoring the environment and detecting the patient at all times. On the contrary, this architecture expands the boundaries of the clinical ward by transforming the local region into its branches. As a result, the drawback data were not secure, and the robotic process automation was not developed. Oguntala et al. (2021) discussed the long short-term memory (LSTM) networks RNN activity classification method and the AAL-based passive radiofrequency identification (RFID) module. The moving people action structure was developed to leverage the received signal strength indicator data of compliant RFID to obtain specific actions. This method develops LSTM networks and RNNs as separate electronic product code as well as RSS great-level aspects. Further, various sizes of information were presented in the LSTM-RNN, and RNN hyperparameters were performed exactly for the categorization. In addition, it achieves the highest categorization accuracy of 98.18%. On the contrary, RFID was utilized to identify an aged person’s health environment. As a result, this method did not discover the three-dimensional information.

Liyakathunisa et al. (2022) implemented AAL for aging care using the IoMT and smart sensors along with the GRU deep learning method. With the fast growth of the elderly population, they face various challenges in day-to-day life, such as social problems, weak walking speeds, medical problems, etc. This method was developed to help monitor an aging person’s activity to overcome these issues. Also, bidirectional GRU and GRU deep learning were used to select robust features to detect targets and anomalies. On the contrary, two datasets were utilized: recorded AAL data and Mobile HEALTH (MHEALTH) benchmark data. In addition, the accuracy was 98.14% and 99.26% for AAI and MHEALTH data, respectively. The advantage was accurately monitoring the aging person’s health status all the time, and the limitations were more expensive and lacked transparency. Madhusanka and Ramadass (2021) discussed communication intention-based activities for daily living for elderly or disabled humans to enhance well-being. The main goal of these methods was to enhance the communication between the curator and the earless people. Earless humans face many challenges in the society, and it is hard to find a qualified curator. To overcome these issues, these models were introduced. Every activity was about discovering a vision-based design structure. Further convolution neural network-based SVMs utilized conversation methods employing computer gaze. As a result, these methods help the caretakers easily interact with earless humans.

Srinivasu et al. (2022) elaborated on AAL to identify the physical activity of diabetic adults based on body area networks. AAL was the social context that enhanced the quality of life in all periods of life. It helps the patient identify real-time activities and provide a robust life. Sometimes, diabetic patients stay in remote areas, so they need to monitor their activity and provide correct time medication. Further, the gold oxide sensor was utilized to calculate the patient’s glucose level; this sensor was placed inside the human body. These methods categorize the high and low glucose levels by using spectrogram images. In addition, it helps to alert the curator to provide efficient medication for the patients. As a result, the limitation was that a small number of datasets were used.

Research gap

The usage of deep learning in the process is AAL for aging and disabled people. The development of AAL involves elderly adults enhancing their independence, safety, and early detection of potential health impairment. In the current work, a few existing models, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL, contain some drawbacks such as high cost, lack of scalability, and small dataset used. To overcome these problems, a strong model is needed in this context, and the SMGR-BS method is developed. The research gaps for this work are as follows.

Scalability

The scalability of the existing methods to handle a huge number of devices and data points remains a challenge. Ensuring that the proposed SMGR-BS method can efficiently process and analyze such extensive datasets is essential for practical deployment.

Privacy concern

The existing methods do not preserve the data and are easy to cause potential damage, ensuring that the SMGR-BS method secures more information in AAL.

Cost

The existing method is costlier, ensuring that the proposed SMGR-BS method can handle the low cost for aging and disabled people.

PROPOSED METHODOLOGY

A novel SMGR-BS algorithm has been proposed in this article for developing AAL for aging and disabled people. Figure 1 provides an overview of the SMGR-BS algorithm. The developed method comprises three major phases, namely the perception layer, the processing layer, and the visualization layer. The perception layer collects data from sensors, cameras, and more to gather information about aging and disabled people. The processing layer has data preprocessing, feature extraction, and classification phases. For analysis purposes, the MHEALTH dataset is used to develop AAL. The data preprocessing phase includes the elimination of erroneous data, noise signal removal, and data cleaning. After preprocessing, AAL is established using stacking multiple GRUs with deep RNNs, and the butterfly optimization with a local search strategy is employed to enhance the hyperparameters of the deep RNN. Finally, the visualization layer gathers details via smartphones, tablets, PCs, etc., such as body temperature, sugar level, and blood pressure; these are sent to the medical practitioner.

Figure 1:

Workflow of the proposed system.

Perception layer

This layer addresses the difficulty of data collection and understanding the goal of comprehensive data perception. The data are gathered from the real world through sensors, cameras, robots, and more.

Processing layer

It is a middleware layer that typically enables multi-linked devices concurrently in the structure of cloud computing to provide better storage, computing, security, and network performance. This layer applies three types of techniques to forward these data to the visualization layer, namely data preprocessing, feature extraction, and classification.

Data preprocessing

Data preprocessing performs a main role in deep learning algorithms, and suitable preprocessing data are mandatory for obtaining performance ( Mishra et al., 2020). In the signal, it clears unnecessary effects, prevents issues, and improves accuracy. In this stage, a dataset, namely MHEALTH, and three types of operations, namely elimination of erroneous data, noise signal removal, and data cleaning, are performed to develop AAL for aging people.

Elimination of erroneous data

Erroneous data are of no use and only increase the amount of data and time to train the method. Therefore, erroneous data should be removed, as duplicate data interfere with the analysis process, and giving high importance to repeated values does not yield accurate results. So, it is necessary to remove the incorrect data.

Noise signal removal

Noise can be introduced due to noisy data, errors in data collection, errors during data entry, or data transmission errors. Denoising filters can be used to reduce noise.

Data cleaning

It refers to correcting or removing incorrect, duplicate, or corrupted data within a dataset, namely the MHEALTH dataset ( Joshi and Patel, 2020). Several data cleaning methods can be applied, such as imputation, transformation, and more. Data cleaning eliminates unnecessary rows and columns, and also, if data are incorrect, consequences and algorithms are untruth and may seem correct.

Feature extraction and classification

This section presents the classification and feature extraction process using the proposed SMGR-BS approach. It effectively extracts relevant data features and accurately classifies them depending on their features. The parameters of stacking multiple GRUs with the deep RNN are optimized using the BOA to improve the prediction accuracy and decrease the loss function and classification performance, and the developed method is referred to as SMGR-BS. The methods utilized for AAL are explained in the subsections below.

Stacking multiple GRUs with the deep RNN

Applying a gating mechanism is the primary concept of GRU, at every time step, which particularly optimizes the network’s hidden state ( Pavithra et al., 2019). This mechanism manages the information flow outside and inside the network, including resetting and updating gates. The update and reset gates demonstrate how to update the hidden state, how many new inputs should be applied, and how the previous hidden state should be forgotten. Depending on the updated hidden state, GRU output is computed. The below-mentioned equations are employed to evaluate the reset gate, the update gate, the new candidate’s hidden state, and the hidden state of a bottom layer GRU: the update gate ( y_ s^1) is given by

(1) $$(Y\text{\_}{s}^{\wedge }1)=\sigma (X\text{\_}{y}^{\wedge }1\ast [G\text{\_}\{s-1\},w\text{\_}s])$$

The reset gate ( q_ s^1) is given below:

(2) $$(q\text{\_}{s}^{\wedge }1)=\sigma (X\text{\_}{q}^{\wedge }1\ast [G\text{\_}\{s-1\},w\text{\_}s])$$

The new candidate’s hidden state and the hidden state update are expressed as

(3) $$(g\text{~\_}{s}^{\wedge }1)=\tan g(X\text{\_}{g}^{\wedge }1\ast [q\text{\_}{s}^{\wedge }1\ast G\text{\_}\{s-1\},W\text{\_}s])$$

(4) $$G\text{\_}{s}^{\wedge }1=(1-y\text{\_}{s}^{\wedge }1)\ast G\text{\_}{\{s-1\}}^{\wedge }1+y\text{\_}{s}^{\wedge }1\ast g~\text{\_}{s}^{\wedge }1.$$

The below equations are employed to evaluate the update gate, the reset gate, the new candidate’s hidden state, and the hidden state of a top Layer GRU. The update gate and the reset gate are expressed as

(5) $$(Y\text{\_}{s}^{\wedge }2)=\sigma (X\text{\_}{y}^{\wedge }2\ast [G\text{\_}{\{s-1\}}^{\wedge }2,G\text{\_}{s}^{\wedge }1])$$

(6) $$(r\text{\_}{s}^{\wedge }2)=\sigma (X\text{\_}{q}^{\wedge }2*[G\text{\_}{\{s-1\}}^{\wedge }2,G\text{\_}{s}^{\wedge }1])$$

The new candidate’s hidden state ( g∼_s^2) is given by

(7) $$(g\text{~\_}{s}^{\wedge }2)+\tan g(X\_g^{\wedge }2\ast [q\_s^{\wedge }2\ast G\_{\{s-1\}}^{\wedge }2,G\_s^{\wedge }1])$$

The hidden state update is

(8) $$G\text{\_}{s}^{\wedge }2=(1-y\text{\_}{s}^{\wedge }2)\ast G\text{\_}{\{s-1\}}^{\wedge }2+y\text{\_}{s}^{\wedge }2\ast g\text{~\_}{s}^{\wedge }2$$

From the above equations, ( G_ s^1), ( G_ s^2) denotes the hidden state of the bottom, and the top GRU layer at the time step s and w_ s represents the input at the time step s. The weight metrics for the candidate’s hidden state, the update gate, and the reset gate are X_ g^1, X_ y^1, and X_ q^1 of the bottom layer, and X_ g^2, X_ y^2, and X_ q^2 of the top layer. The output of the bottom GRU layer, G_ s^1, serves as a portion of the input to the top GRU layer at each time step in a stacked configuration, allowing the network to seize difficult patterns in sequential data.

Deep RNN

Through the operation of temporal storage, this network can encode long-term information. The backpropagation algorithm is trained to utilize it in the recurrent layer at the time of training the deep RNN ( Shang et al., 2021). We propagate the previous layer sent to the current layer and the previous time sent in two directions for error generation:

(9) $$t_k(u)=g(NE{T}_k(u))$$

(10) $$z_l(u)=h(NE{T}_l(u))$$

where g and h are the output and activation function layers of the hidden state, respectively. Equation 11 represents the loss function of training the neural network.

(11) $$NE{T}_l(u)={\displaystyle \sum _k^n{t}_k(u)}{x}_{lk}+c_l$$

The mathematical expression of the loss function of the training network can be represented as

(12) $$D=\frac{1}{2}{\displaystyle \sum _q^o{\displaystyle \sum _l^p{(e_{ql}-z_{ql})}^2}}$$

The backward propagation of error is

(13) $$\delta =-\frac{\partial (D)}{\partial (NET)}$$

(14) $${\delta }_{qk}(u-1)={\displaystyle \sum _i^n{\delta }_{qi}}(u)v_{ik}{g}^{\prime }(t_{qk}(u-1))$$

If the time step is u, i is the hidden layer node and when the time step is u−1, k is the hidden layer node index. Then, the output error layer is

(15) $$f_0(u)=e(u)-z(u)$$

The output layer weights are denoted as

(16) $$X(u+1)=X(u)+\eta t(u)f_0{(u)}^U$$

The output layer to the hidden layer error gradient propagation is

(17) $$f_i(u)=e_i(f_0{(u)}^{U}W,u)$$

Due to the gradient function, the deep RNN is difficult to train and has a slow and complex training process. A fixed dropout rate of 0.5 was adopted after the feed-forward layer, given the compact size of the dataset, to avert overfitting issues.

Butterfly search optimization algorithm

This algorithm imitates the food-consuming habits of butterflies. The sections below discuss some biological certainties and how to model them in this optimization algorithm. To discover food and mating partners, butterflies employ their vision, smell, sight, touch, taste, and listening ( Arora and Singh, 2019). These senses help butterflies escape from enemies, move from one place to another, and lay their eggs in the right places. Among all these senses, the most important sense is smell, which helps in detecting honey from long distances. On the one hand, the global search in the algorithm is that if a butterfly can sense scent from an alternate butterfly, it will go toward it. On the other hand, the local search is referred to if a butterfly is not capable of sensing scent from the environment it goes irregularly.

Fragrance

Every scent has its personal touch and unique fragrance, which is the major feature that differentiates this algorithm from other methodologies. Initially, we are required to understand how a stimulus is processed by a modality such as sound, smell, light, temperature, and more to understand how scent is accounted for. Depending on these three terms, namely stimulus intensity, sensory modality, and power exponent, are the entire processing and sense of the modality concept. The local search capacity of the algorithm is enhanced here by integrating the method termed local search scheme ( Mousa et al., 2012). The butterfly implements a probabilistic action choice rule for deciding the next position to visit. The butterfly at position o selects the next position t to move if r ≤ r ₀,

(18) $$q_{ot}^l=\{\begin{array}{ll}1\hfill & ift=\arg \underset{v\in O_l(o)}{\max i}\left\{{\displaystyle \sum _{m=1}^4{x}_m\cdot {\sigma }_{mv}^m(1+{\beta }_{ov}^l)\cdot {\gamma }_{ov}^{\alpha }}\right\}\hfill \\ 0\hfill & \text{otherwise}\hfill \end{array}$$

else

(19) $$q_{ot}^l=\frac{{\displaystyle {\sum }_{l=1}^4{x}_l}{\left|{\sigma }_{ot}^m(1+{\beta }_{ot}^m)\right|}^{\mu }{\left|{\lambda }_{ot}\right|}^{\psi }}{{\displaystyle {\sum }_{\xi \in O_O^L}\left({{\displaystyle \sum _1^4{X}_m{\left|{\sigma }_{o\xi }^m(1+{\beta }_{o\mu }^m)\right|}^{\mu }\left|{\lambda }_{o\xi }\right|}}^{\psi }\right)}},\hspace{0.17em}if\hspace{0.17em}t\hspace{0.17em}\in \hspace{0.17em}{O}_l(o)$$

where ꞵ _ot , ꞵ ϵ (0,1)· ꞵ _ot represent the new pheromone developed into the model to attract butterflies. In the search space, the butterfly migrates from one place to another particular place, so the fitness of the butterfly and this scent vary; the fitness will change depending on their position. The mathematical formulation of the fragrance function is

(20) $$f=t{K}^{\alpha }$$

Here, K is the stimulus intensity, the fragrance is represented as f, the power exponent is represented as α, and t describes the sensory modality. Initiation is the first phase in which a few butterflies are generated into a population with initial and individual parameters. The global search phase is the iteration process, and whenever it finds a scent from that butterfly in the search area, it moves to the better butterfly. The global search phase equation is expressed in the below equation:

(21) $$z_i(y+1)=z_i(y)+(ran{d}^2\times b^{\ast }-z_i(y))\times f_i$$

where y denotes the number of iterations, rand ϵ [1,0], and z _i means the vector position of the ith butterfly. f _i indicates a fragrance of the ith butterfly, and the global optima are described as b ^*. Within the search space, a butterfly cannot identify the scent of another butterfly; it moves irregularly; this is referred to as the local search stage. The local search phase is expressed in the below equation.

(22) $$z_i(y+1)=z_i(y)+(ran{d}^2\times z_l(y)-z_j(y))\times f_i$$

where jth and lth butterflies create the population of z _j and z _l .

The BOA offers distinct advantages in optimizing the parameters of the SMGR-BS method within the context of AAL. Specifically, the BOA excels in handling high-dimensional and non-linear optimization problems, which are common in AAL applications due to sensor data’s diverse and complex nature. Additionally, the BOA’s local search strategy complements the deep learning framework employed in SMGR-BS, allowing for fine-tuning of model parameters to enhance its effectiveness in capturing intricate temporal dependencies in sensor data. By leveraging the BOA, the SMGR-BS method can achieve optimal performance and robustness, making it well-suited for developing AAL solutions that meet the unique needs of aging and disabled individuals.

The proposed SMGR-BS algorithm

Figure 2 explains the workflow of the proposed SMGR-BS algorithm for the development of AAL. The data were collected from the MHEALTH dataset. Here, the stacking multiple GRU with the deep RNN is employed to develop AAL.

Figure 2:

Workflow of the SMGR-BS algorithm. Abbreviations: BOA, butterfly optimization algorithm; SMGR-BS, stacking multiple gated recurrent-based butterfly search.

First, initialize the maximum number of iterations, calculate the update gate with the time steps, and estimate the rest gate, the weight of the initial layer, and the hidden layer. Choose the loss function and the activation function of the deep RNN with the matrix weight. If the condition is satisfied, then the outcome is obtained; otherwise, it can go through the BOA. The BOA is employed to enhance the hyperparameter of the deep RNN by stacking multiple GRUs. Here, the parameters are to be initialized first, and then the fitness values are to be calculated. The fitness values of the optimization problem are evaluated, and the individual’s positions are also updated in the optimization. However, this has a low convergence speed to enhance the local search strategy employed in the BOA. Then, the conditions are checked and the predicted output is obtained; otherwise, the process can be repeated. The final layer is the visualization layer; the major contribution of this layer is analyzing the data given by IoT environments such as smartphones, sensors, and more. It collects information about aging and disabled people and conveys it to the medical practitioner.

Robustness

Ensuring the robustness of the SMGR-BS method against adversarial attacks or noisy sensor data is paramount for its reliability and resilience in practical settings. To address this concern, several strategies and techniques have been employed. First, data preprocessing techniques, such as noise reduction algorithms and outlier detection methods, are implemented to enhance sensor data quality and mitigate the impact of noisy measurements. Additionally, model regularization techniques, including dropout and weight decay, are applied to prevent overfitting and improve the generalization capability of the SMGR-BS model. Furthermore, ensemble learning approaches, such as model averaging or boosting, are utilized to aggregate multiple models’ predictions and enhance robustness against adversarial attacks or model biases. Moreover, continuous monitoring and model retraining mechanisms are established to adapt to evolving environmental conditions and maintain the performance of the AAL solution over time. By integrating these strategies and techniques, the SMGR-BS method demonstrates enhanced robustness and resilience against adversarial attacks or noisy sensor data, ensuring its reliability in practical AAL settings.

EXPERIMENTAL RESULTS AND DISCUSSION

The effectiveness of the SMGR-BS method for the development of AAL for aging and disabled people and the results achieved from the study are demonstrated in this section. The SMGR-BS method is evaluated with various evaluation measures, namely specificity, recall, F1-score, accuracy, and precision, and the results are compared with existing methods such as AAL using supervised learning ( Patro et al., 2021), DTPAAL ( Cascone et al., 2022), edge AAL ( Ardito et al., 2022), and IoT-based AAL ( Gulati and Kaur, 2022).

Experimental setup

In this work, the SMGR-BS method is implemented in a Home Sensor Simulator, and the system uses a NODE-RED simulation platform in a real-time environment. Furthermore, the computing platform meets minimum software and hardware requirements, including sufficient storage capacity, computational power, and compatibility with stacking multiple GRU-based deep RNN frameworks. Considering these system requirements and platform specifications, this study ensures reliable and efficient implementation of the SMGR-BS method for the development of an AAL for aging and disabled people.

Parameter setting

The performance of the SMGR-BS method is enhanced by implementing the parameter setting, and Table 1 lists the parameter setting of the study. In this process, optimal parameter values are created to improve the performance of the SMGR-BS method.

Table 1:

Parameter setting.

Technique	Parameter	Value
GRU	Optimizer	Adam
	Learning factor	0.001
	Delay sequence	5
	Batch size	32
	Epochs	500
	Activation function	Hyperbolic tangent
	Loss function	Mean squared error (MSE)
RNN	Optimizer	Adam
	Batch size	64
	Learning rate	0.005
	Loss function	Mean squared error (MSE)
	Epochs	500
	Activation function	ReLU

Abbreviations: GRU, gated recurrent unit; ReLU, rectified linear unit; RNN, recurrent neural network.

In this work, the GRU parameter setup consists of the following: the delay sequence is 5, the activation function is the hyperbolic tangent, the number of epochs is 500, the loss function is the mean squared error (MSE), the learning factor is 0.001, the batch size is 32, and the Adam is utilized as an optimizer. Also, in the RNN parameter setting, the rectified linear unit is used as an activation function, Adam is utilized as an optimizer, the learning rate is 0.005, the loss function is the MSE, the batch size is 64, and the number of epochs is 500. In this study, the BOA with a local search strategy is utilized for hyperparameter tuning to enhance the effectiveness of the SMGR-BS method. This work ensures efficient and reliable implementation of the SMGR-BS method, considering these parameter details for the development of an AAL for aging and disabled people. The SMGR-BS method exhibits promising computational efficiency, making it well-suited for deployment in resource-constrained AAL environments. Leveraging stacked multiple GRUs and the BOA, the method efficiently captures intricate temporal dependencies in sensor data while optimizing model parameters. During experimental evaluation, the computational requirements of SMGR-BS were found to be manageable, with modest processing times for both training and inference phases. Notably, the model’s architecture is designed to strike a balance between computational complexity and predictive performance, ensuring practical feasibility in real-world AAL deployments.

Dataset description

In this work, the MHEALTH dataset ( Jain, 2020) is utilized to implement the SMGR-BS method for the development of an AAL for aging and disabled people. It encompasses information from sensors, user inputs, and other sources such as electronic health records. In this study, 10,000 observations are collected from the MHEALTH dataset, and these observations include 12 types of physical activities, namely walking, cycling, running, jogging, climbing stairs, knee bending, sitting and relaxing, jumping front and back, frontal elevation of arms, lying down, standing still, and bending waist forward performed by 10 individuals who wore three sensors. The observations are split into training and testing in the ratio of 80:20 for the development of an AAL for aging and disabled people.

Evaluation measures

The effectiveness of the SMGR-BS method for the development of an AAL for aging and disabled people is evaluated through various evaluation measures, namely specificity, accuracy, recall, precision, and F1-score. The performance evaluation of these metrics is conducted based on the mathematical expressions mentioned below.

Accuracy

Accuracy ( a _y ) is the measurement of correctly classified instances to the total number of instances. Accuracy can be calculated as

(23) $$a_y=\frac{t+t^{\prime }}{t+t^{\prime }+f+f^{\prime }}$$

Precision

Precision ( p _n ) is the ratio of correctly predicted positive instances to all instances predicted as positive. The precision can be formulated as

(24) $$p_n=\frac{t}{t+f}$$

Recall

Recall ( r _e ) is the ratio of correctly predicted positive instances out of all actual positive instances. The recall can be represented as

(25) $$r_e=\frac{t}{t+f^{\prime }}$$

F1-score

F1-score ( f1 _score ) is the harmonic mean of recall and precision, and during the precision–recall tradeoff, if the precision increases the recall decreases. The F1-score can be calculated as

(26) $$f{1}_{score}=2\times \frac{(p_n\times r_e)}{(p_n+r_e)}$$

Specificity

Specificity ( s _p ) is the ratio of correctly predicted negative instances out of all actual negative instances. The specificity can be expressed as

(27) $$s_p=\frac{t^{\prime }}{t^{\prime }+f}$$

In Equations 23– 27, t, t′, f, and f′ denote true positive, true negative, false positive, and false negative, respectively.

Performance analysis

The performance analysis of the SMGR-BS method to develop an AAL for aging and disabled people using the specified performance metrics, namely specificity, F1-score, recall, precision, and accuracy, provides a comprehensive evaluation of its effectiveness. The performance is evaluated by comparing the SMGR-BS method with the existing methods, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL. Figures 3– 7 illustrate the comparative graphical representation of the SMGR-BS method and the existing methods for various evaluation metrics based on the development of an AAL for aging and disabled people.

Figure 3:

Performance validation based on accuracy. Abbreviations: AAL, ambient assisted living; DTPAAL, digital twinning pepper and ambient assisted living; IoT, Internet of Things; SMGR-BS, stacking multiple gated recurrent-based butterfly search.

Figure 4:

Graphical representation of precision analysis. Abbreviations: AAL, ambient assisted living; DTPAAL, digital twinning pepper and ambient assisted living; IoT, Internet of Things; SMGR-BS, stacking multiple gated recurrent-based butterfly search.

Figure 5:

Recall analysis for performance evaluation. Abbreviations: AAL, ambient assisted living; DTPAAL, digital twinning pepper and ambient assisted living; IoT, Internet of Things; SMGR-BS, stacking multiple gated recurrent-based butterfly search.

Figure 6:

Performance validation based on the F1-score. Abbreviations: AAL, ambient assisted living; DTPAAL, digital twinning pepper and ambient assisted living; IoT, Internet of Things; SMGR-BS, stacking multiple gated recurrent-based butterfly search.

Figure 7:

Graphical representation of specificity analysis. Abbreviations: AAL, ambient assisted living; DTPAAL, digital twinning pepper and ambient assisted living; IoT, Internet of Things; SMGR-BS, stacking multiple gated recurrent-based butterfly search.

The accuracy of the SMGR-BS method and the existing methods is demonstrated by the graphical analysis as shown in Figure 3. The SMGR-BS method achieved a high accuracy of 98.67%, while the existing methods, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL, obtained a low accuracy of 97.54%, 96.41%, 95.36%, and 94.23%, respectively. Figure 4 illustrates the graphical analysis to depict the precision of the SMGR-BS method and the existing methods. The SMGR-BS method attained a high precision of 97.93%, while the existing methods, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL, obtained a low precision of 96.87%, 96.12%, 95.36%, and 94.79%, respectively.

In Figure 5, the graphical analysis illustrates the recall of the SMGR-BS and existing methods. The SMGR-BS method achieved a high recall of 97.86%, while the existing methods, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL, obtained a low recall of 96.57%, 96.15%, 95.42%, and 94.78%, respectively. Figure 6 represents the graphical analysis to determine the F1-score of the SMGR-BS method and the existing methods. The SMGR-BS method attained a high F1-score of 97.82%, while the existing methods, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL, obtained a low F1-score of 96.45%, 96.03%, 95.37%, and 94.71%, respectively.

The specificity of the SMGR-BS method and the existing methods is depicted by the graphical analysis represented in Figure 7. The SMGR-BS method achieved a high specificity of 97.84%, while the existing methods, such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL, obtained a low specificity of 96.38%, 95.63%, 94.51%, and 94.07%, respectively. The performance analyses evaluate the effectiveness of the SMGR-BS method for the development of an AAL for aging and disabled people. The results illustrate that the SMGR-BS method achieved high precision, accuracy, specificity, recall, and F1-score compared to state-of-the-art methods.

CONCLUSION

The SMGR-BS method holds significant advantages for the development of AAL for aging and disabled people. In this study, stacking multiple GRU-based deep RNNs enables the extraction of vast and heterogeneous data generated in AAL environments, helping to monitor and predict the behaviors of residents. Moreover, the integration of the BOA with the local search strategy is utilized to optimize the hyperparameters. The utilization of the MHEALTH dataset, with its diverse and extensive health-related data, empowers AAL systems to make informed decisions and provide timely assistance to individuals with specific health needs. The effectiveness of the SMGR-BS method is evaluated using different evaluation measures, namely specificity, F1-score, precision, accuracy, and recall, and these results are compared with existing methods such as AAL using supervised learning, DTPAAL, edge AAL, and IoT-based AAL. The SMGR-BS method achieved high accuracy of 98.67%, precision of 97.93%, recall of 97.86%, F1-score of 97.82%, and specificity of 97.84%, respectively. The results illustrate that the proposed method achieved better results in developing AAL for aging and disabled people.

Incorporating interpretability into the SMGR-BS model is crucial for enhancing personalized care and support for aging and disabled individuals. By elucidating how the model’s predictions are generated and which features contribute most significantly to these predictions, caregivers can gain valuable insights into the specific needs and behaviors of individuals under their care. For instance, understanding patterns detected by the model, such as changes in activity levels or deviations from routine behaviors, can inform timely interventions and tailored support measures. Moreover, transparent explanations of the model’s decisions can foster trust between caregivers, individuals, and the technology itself, ultimately promoting more effective and empathetic care practices.

Despite the effectiveness of the SMGR-BS method, several challenges could be encountered during the implementation. For instance, the availability and quality of sensor data may vary across different AAL environments, leading to potential biases or limitations in model generalization. Furthermore, the computational complexity associated with training deep learning models, such as SMGR-BS, may pose challenges in resource-constrained environments or real-time applications. Additionally, the choice of evaluation metrics and validation procedures may influence the interpretation of results and comparisons with existing methodologies.

The deployment of AAL systems raises concerns regarding data privacy, consent, and potential biases inherent in algorithmic decision-making. In the context of the proposed SMGR-BS method, we acknowledge these concerns and have implemented several measures to address or mitigate them. First, we prioritize data privacy by adhering to strict data protection protocols and anonymizing sensitive information to prevent unauthorized access or misuse. Additionally, we ensure transparency and accountability by providing clear explanations of the model’s decision-making processes and allowing users to understand and challenge algorithmic outputs. Furthermore, we actively engage with stakeholders, including individuals, caregivers, and healthcare professionals, to incorporate their perspectives and preferences into the design and deployment of AAL systems. By fostering open dialogue and collaboration, we aim to promote the ethical and responsible use of deep learning models in AAL applications, ultimately enhancing trust and acceptance among end-users and stakeholders.

Integration of this technology into existing AAL systems or healthcare infrastructures holds promise for enhancing the quality of care and support for aging and disabled individuals. For example, SMGR-BS could be utilized in smart home environments equipped with sensor networks to monitor and analyze daily activities, detect anomalies or changes in behavior, and provide timely assistance or alerts to caregivers or healthcare professionals. Additionally, integration with wearable devices or mobile applications could enable individuals to receive personalized feedback and recommendations based on their health and activity data, promoting self-management and independence. Furthermore, collaboration with healthcare providers and stakeholders could facilitate the implementation of SMGR-BS in clinical settings for early detection of health issues, remote patient monitoring, and optimizing care delivery pathways. Overall, exploring these real-world applications and deployment scenarios underscores the potential impact of SMGR-BS in transforming AAL systems and healthcare delivery.

Future research can focus on further improving the predictive capabilities of AAL systems. Developing more sophisticated deep learning frameworks to better capture complex temporal relationships within the dataset is necessary. The generalizability of the proposed SMGR-BS approach beyond the MHEALTH dataset is indeed an important aspect to consider. While the initial experiments demonstrate promising results on the MHEALTH dataset, it has been recognized the need to validate the SMGR-BS method on other datasets with varying characteristics to assess its robustness and applicability across diverse AAL environments.

ACKNOWLEDGMENTS

The authors extend their appreciation to the King Salman Center for Disability Research for funding this work through Research Group no KSRG-2022-043.

REFERENCES

1. Abidi MH, Alkhalefah H, Mohammed MK, Umer U, Qudeiri JEA. 2020a. Optimal scheduling of flexible manufacturing system using improved lion-based hybrid machine learning approach. IEEE Access. Vol. 8:96088–96114. [Cross Ref]

2. Abidi MH, Umer U, Mohammed MK, Aboudaif MK, Alkhalefah H. 2020b. Automated maintenance data classification using recurrent neural network: enhancement by spotted hyena-based whale optimization. Mathematics. Vol. 8(11):2008. [Cross Ref]

3. Abidi MH, Alkhalefah H, Umer U. 2022a. Fuzzy harmony search based optimal control strategy for wireless cyber physical system with industry 4.0. J. Intell. Manuf. Vol. 33(6):1795–1812. [Cross Ref]

4. Abidi MH, Mohammed MK, Alkhalefah H. 2022b. Predictive maintenance planning for industry 4.0 using machine learning for sustainable manufacturing. Sustainability. Vol. 14(6):3387. [Cross Ref]

5. Alluhaibi R, Alharbe N, Aljohani A, Al Mamlook RE. 2023. Selection of an efficient classification algorithm for ambient assisted living: supportive care for elderly people. Healthcare. Vol. 11(2):256. [Cross Ref]

6. Ardito C, Di Noia T, Fasciano C, Lofù D, Macchiarulo N, Mallardi G, et al.. 2022. An edge ambient assisted living process for clinical pathwayAmbient Assisted Living. Springer International Publishing. Cham: p. 363–374

7. Arora S, Singh S. 2019. Butterfly optimization algorithm: a novel approach for global optimization. Soft Comput. Vol. 23(3):715–734. [Cross Ref]

8. Cascone L, Nappi M, Narducci F, Passero I. 2022. DTPAAL: digital twinning pepper and ambient assisted living. IEEE Trans. Industr. Inform. Vol. 18(2):1397–1404. [Cross Ref]

9. Cicirelli G, Marani R, Petitti A, Milella A, D’Orazio T. 2021. Ambient assisted living: a review of technologies, methodologies and future perspectives for healthy aging of population. Sensors (Basel). Vol. 21(10):3549. [Cross Ref]

10. Correia L, Fuentes D, Ribeiro J, Costa N, Reis A, Rabadão C, et al.. 2021. Usability of smartbands by the elderly population in the context of ambient assisted living applications. Electronics. Vol. 10(14):1617. [Cross Ref]

11. Dasios A, Gavalas D, Pantziou G, Konstantopoulos C. 2015. Hands-on experiences in deploying cost-effective ambient-assisted living systems. Sensors (Basel). Vol. 15(6):14487–14512. [Cross Ref]

12. Guerra BMV, Schmid M, Beltrami G, Ramat S. 2022. Neural networks for automatic posture recognition in ambient-assisted living. Sensors (Basel). Vol. 22(7):2609. [Cross Ref]

13. Gulati N, Kaur PD. 2022. FriendCare-AAL: a robust social IoT based alert generation system for ambient assisted living. J. Ambient Intell. Humaniz. Comput. Vol. 13(4):1735–1762. [Cross Ref]

14. Jain G. 2020. Mobile health human behavior analysis. Kaggle. https://www.kaggle.com/datasets/gaurav2022/mobile-health

15. Joshi AP, Patel BV. 2020. Data preprocessing: the techniques for preparing clean and quality data for data analytics process. Orient. J. Comput. Sci. Technol. Vol. 13(2, 3):78–81. [Cross Ref]

16. Liyakathunisa, Alsaeedi A, Jabeen S, Kolivand H. 2022. Ambient assisted living framework for elderly care using Internet of Medical Things, smart sensors, and GRU deep learning techniques. J. Ambient Intell. Smart Environ. Vol. 14:5–23. [Cross Ref]

17. Madhusanka BGDA, Ramadass S. 2021. Implicit intention communication for activities of daily living of elder/disabled people to improve well-beingIoT in Healthcare and Ambient Assisted Living. Marques G, Bhoi AK, de Albuquerque VHC, Hareesha KS. p. 325–342. Springer. Singapore:

18. Maskeliūnas R, Damaševičius R, Segal S. 2019. A review of Internet of Things technologies for ambient assisted living environments. Future Internet. Vol. 11(12):259. [Cross Ref]

19. Mishra P, Biancolillo A, Roger JM, Marini F, Rutledge DN. 2020. New data preprocessing trends based on ensemble of multiple preprocessing techniques. TrAC Trends Anal. Chem. Vol. 132:116045. [Cross Ref]

20. Mousa AA, El-Shorbagy MA, Abd-El-Wahed WF. 2012. Local search based hybrid particle swarm optimization algorithm for multiobjective optimization. Swarm Evol. Comput. Vol. 3:1–14. [Cross Ref]

21. Oguntala GA, Hu YF, Alabdullah AAS, Abd-Alhameed RA, Ali M, Luong DK. 2021. Passive RFID module with LSTM recurrent neural network activity classification algorithm for ambient-assisted living. IEEE Internet Things J. Vol. 8(13):10953–10962. [Cross Ref]

22. Patro SP, Padhy N, Chiranjevi D. 2021. Ambient assisted living predictive model for cardiovascular disease prediction using supervised learning. Evol. Intell. Vol. 14(2):941–969. [Cross Ref]

23. Pavithra M, Saruladha K, Sathyabama K. 2019. GRU based deep learning model for prognosis prediction of disease progressionProceedings of the 2019 3rd International Conference on Computing Methodologies and Communication (ICCMC); p. 840–844. Erode, India

24. Queirós A, Dias A, Silva AG, Pacheco Rocha N. 2017. Ambient assisted living and health-related outcomes—a systematic literature review. Informatics. Vol. 4(3):19. [Cross Ref]

25. Shang K, Chen Z, Liu Z, Song L, Zheng W, Yang B, et al.. 2021. Haze prediction model using deep recurrent neural network. Atmosphere. Vol. 12(12):1625. [Cross Ref]

26. Srinivasu PN, JayaLakshmi G, Jhaveri RH, Phani Praveen S. 2022. Ambient assistive living for monitoring the physical activity of diabetic adults through body area networks. Mobile Inform. Syst. Vol. 2022:3169927. [Cross Ref]

27. Taramasco C, Rimassa C, Martinez F. 2022. Improvement in quality of life with use of ambient-assisted living: clinical trial with older persons in the Chilean population. Sensors (Basel). Vol. 23(1):268. [Cross Ref]

28. Thakur N, Han CY. 2021. Indoor localization for personalized ambient assisted living of multiple users in multi-floor smart environments. Big Data Cogn. Comput. Vol. 5(3):42. [Cross Ref]

29. Thakur N, Han CY. 2022. A simplistic and cost-effective design for real-world development of an ambient assisted living system for fall detection and indoor localization: proof-of-concept. Information. Vol. 13:363. [Cross Ref]