The adoption of Multi-Access Edge Computing (MEC) stems from the need to overcome the limitations of mobile device resources and to optimize the performance of computation-intensive applications such as virtual and augmented reality. Additionally, by leveraging the content reuse property, multicast transmission emerges as an efficient strategy to alleviate network load, enabling the delivery of common content to multiple users simultaneously. This work investigates the challenges and opportunities associated with integrating multicast into MEC networks. We formulate a complex decision-making problem involving task offloading at the MEC level, designing multicast groups, and transmission scheduling. We develop a greedy heuristic solution along with a genetic algorithm-based approach. The simulation results provide valuable insights into the fundamental benefits of multicast in MEC networks, demonstrating its potential to optimize resource usage. They also highlight the superiority of our genetic approach as it efficiently balances multicast and unicast transmissions.

Federated learning (FL) was introduced recently as a new machine learning (ML) paradigm. It is a distributed network of client nodes that train ML and deep learning (DL) models on their local data without sharing them to preserve data privacy (DP). However, these data are heterogeneous by nature as they are collected in different contexts using various sources such as IoT devices. Consequently, data heterogeneity (DH) in FL has brought new performance-related challenges. Few of these challenges have been addressed in the literature; moreover, context heterogeneity and balance rate were not explored at all. In this paper, we introduce an FL approach in which a trade-off between personalization and generalization is achieved to mitigate the impact of DH and obtain better performance. We focus on three DH challenges: context, non-independent and identically distributed (non-IID) data, and balance rate. For the implementation, fall detection (FD) data is used to demonstrate the potential of our approach in improving the FL system’s performance. FD is an important subject and is particularly prevalent for the safety of elderly people. Hence, we collected fall data from two sensors: accelerometer (ACC) and heart rate (HR), then, we used two ML models to evaluate our approach. We utilized XGBoost (XGB) for balanced and unbalanced clients and One-Class Support Vector Machine (OC-SVM) for one-label clients. Our approach achieved an average F1-score of 88%. A comparative study was also conducted with previous works on FD. Our results showed a performance improvement which exceeded 94.30% on average.

In recent years, significant advances in biometrics have essentially been driven by machine learning (ML) and deep learning (DL) progress. Numerous human identification applications are currently available using physical traits such as fingerprints, face, and voice. With the development of Internet of Things (IoT) sensors and the availability of a variety of ML algorithms, there has been increased research interest in subject identification (SI) based on behavioral cues. For example, several research works have been published on SI based on gait analysis. Sensors like accelerometers (ACC), gyroscopes, (GYR), and magnetometers (MAG) were used to collect data during limited activities such as walking. We believe that using data for one activity is not sufficient to adequately capture behavioral cues for the purpose of SI. Considering other cues such as gestures or head shaking and using a variety of sensors located on different parts of the human body are essential to developing a scenario that includes expressive human activities. We designed a specific scenario that included several activities, such as walking, giving a talk, chatting while sitting, and climbing stairs, using five inertial measurement units (IMU) located on various parts of the human body. Several ML algorithms, namely Linear Discriminant Analysis (LDA), K-Nearest Neighbours (KNN), Random Forest (RF), and XGBoost (XGB) were used. Our results show that SI yields beyond 99% accuracy for most activities. Furthermore, we succeeded in implementing a real-time IoT system for SI based on our best offline results. We achieved 98.04% accuracy within 0.06 ms of processing time (PT).

Fall is one of the main causes of injuries for the elderly, and fall detection (FD) for senior monitoring has received considerable attention from both the academic community and healthcare industries. In recent years, there has been an increasing interest in using wearable sensors, such as accelerometers to monitor the subject’s body movement and apply Machine Learning (ML) methods to detect and prevent falls. Since it is extremely difficult to collect accelerometer data of real falls during activities of daily living (ADL), researchers tended to rely on simulating falls in well-protected environments. They collected ADLs separately, applied ML algorithms to classify falls and ADLs, and reported very high FD accuracy rates. However, these studies cannot be applied in a real fall context. In this paper, instead of classifying ADL and fall separately, we propose to incorporate fall data within ADL data to obtain more realistic datasets and apply ML to detect falls. Several ML algorithms including CatBoost (CB), Decision Tree (DT), Random Forest (RF), and XGBoost (XGB) were applied to the datasets. Experimental results show a fall detection accuracy of $88.70 \%$. We also extend our work to cover slow fall which, to the best of our knowledge, was not extensively addressed in previous works.

Hand gesture recognition (HGR) is a vital area of research with widespread applications and employing disruptive technologies such as Artificial Intelligence (AI) and the Internet of Things (IoT). HGR is very important in human-computer interaction (HCI), especially for people with disabilities. Several approaches have been described in the literature, including image-based, radar-based, and haptic-based ones. However, these approaches are difficult to implement in real time as they require high processing time (PT). Inertial sensors (IS) offer a worthy alternative for HGR and require much less PT. Several research papers have been published in this domain. However, models capturing complex temporal patterns using traditional machine learning (ML) algorithms have limitations. In this paper, we propose a transformer-based approach to accurately recognize dynamic gestures. A dataset we collected from accelerometer (ACC) and gyroscope (GYR) sensors was used to evaluate the performance of our approach. Our experimental results achieved an accuracy of 96.77% and outperformed traditional ML algorithms in terms of accuracy using the same dataset: Linear Discriminant Analysis (LDA) with 58.06%, KNearest Neighbor (KNN) with 70.97%, XGBoost (XGB) with 74.19%, Convolutional neural network (CNN) with 74.19%, Random Forest (RF) with 77.42%, and Support vector machine (SVM) with 93.55%. Furthermore, we implemented a real-time HGR system that can achieve a very short response time of less than 300 ms.

Detecting Freezing of Gait (FOG) is crucial for Parkinson's disease (PD). Several research works have been published on FOG detection and prediction using machine learning (ML) and limited accelerometer data. FOG data collection is challenging and generally conducted on a limited number of persons as freezing occurs randomly, unlike walking which is a recurrent activity. This makes directly applying ML inefficient. In this paper, to improve previously published results, we apply three different ML algorithms: Linear Discriminant Analysis (LDA), Extreme Gradient Boosting (XGB), and Extra Trees (ET), in addition to a Transformer model to detect and predict FOG. This is achieved by modelling the publicly available DAPHNet dataset, which contains accelerometer readings from three wearable sensors. The data quality is first improved by using data augmentation, data balancing, and feature extraction. Our results show that accuracy (ACC), along with other metrics like sensitivity (SEN), specificity (SPE), and F1-score (F1) is important to quantify the performance of our results. From FOG detection results, we can see that XGB and ET algorithms provide 100% ACC, and slightly lower performance, 99.96% with LDA. For FOG prediction, our results prove we can achieve 99.21%, 96.40%, 96.21%, and 94.70% ACC with LDA, XGB, ET, and Transformer, respectively. These results outperform previously published results on the same dataset.

Falls are one of the most common causes of injury among the elderly. As a result, fall detection has received in the last decade considerable attention from both academia and the healthcare industry. Accelerometer data, collected from simulated falls, were widely used with classical machine learning (ML) algorithms as well as with threshold-based methods to identify fall situations that can be used to launch an alert for help. As collecting real fall data is challenging, most of the research papers on fall detection have used limited data which do not reflect the complexity of real fall situations. Fortunately, a comprehensive fall dataset called “Simulated Falls and Daily Living Activities Dataset” has recently become available. This dataset includes 1827 simulated falls of 20 different types. In this paper, we use this dataset to evaluate the possibility of fall detection, more precisely, impact and pre-impact which correspond to fall and pre-fall, respectively. Unlike the classical ML algorithms and threshold-based methods commonly used in previous research works, in this paper, we implement a bidirectional long short-term memory (Bi-LSTM) algorithm which we believe better reflects the impact and pre-impact context as it takes into consideration both backward and forward sequence information at every time step. Our experimental results showed that Bi-LSTM achieves an accuracy of 99.97% and 99.95%, with 99.80% and 99.30% sensitivity, and 100% and 99.99% specificity for fall and pre-fall detections, respectively. These results largely exceed previously published results.

The Internet of Things (IoT) generates massive amounts of data everywhere through sensors of every kind which are disseminated in a variety of objects. This data contains incredibly valuable information useful for multiple applications. Knowing the context in which it was generated is extremely important and constitutes one of the first steps in extracting the knowledge it contains. Thereby, Context-Aware Learning (CAL) has become an important area of research as machine learning (ML) is a fast and ever-evolving technology. Wearable devices, ranging from accelerometers (ACC), frequently used, to magnetic field sensors, are used to monitor and recognize human activities (HA). Beyond ML Algorithms (MLA), accurate Human Activities Recognition (HAR) or context identification, depends not only on the kinds of sensors used but also on their location. In this paper, we study the impact of three types of sensors: ACC, gyroscope (GYR), and magnetometer (MAG); and their locations on the performance of MLA for HAR. Our results show that magnetic field sensors, less frequently used in the literature, placed at a specific location, provide the best performance in terms of HAR. Using a publicly available dataset, PAMAP2, we implement and evaluate the performance of HAR using five MLA: Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QLA), K-Nearest Neighbors (KNN), Decision Tree (DT), and Random Forest (RF). Our results show that the success rate of these algorithms is 98.3%, 90.4%, 97.6%, 99.9%, and 100% respectively, which exceeds the results obtained in a previous work based on the same dataset.

Wearable sensors are frequently used for monitoring physical activities and medical conditions. A variety of sensors are used, such as the accelerometer (ACC), gyroscope (GYR), and magnetometer (MAG), which are often embedded in Inertial Measurement Units (IMU). Data collected from these sensors can be used to identify context or physical activities through context-aware learning methods that apply a variety of learning algorithms. Implementing a real-time system (RTS) that serves a specific application like fall detection or heart condition, for example, is challenging as response time must be within a certain interval. This response time depends on the speed of data collection and transmission as well as the prediction time. Even though a lot of research was done in this area, to the best of our knowledge there is no comparative study based on the criteria we are using here. In this paper, we propose an Edge-based RTS for health-related applications and conduct a comparative study of several Machine Learning Algorithms (MLA) according to four criteria: source of data, sampling period, prediction time (PT), and success rates (SR). Even though MLA demonstrate different behavior toward these criteria, our simulation results showed that it is possible to implement a RTS that can identify or predict accurately physical activities within acceptable time constraints which are application dependant. Simulations were performed on wearable sensors’ data that we collected from 24 participants practicing five different physical activities. Our simulation results showed that the Decision Tree algorithm, with SR of 97.93% and PT of 0.17 seconds, outperformed all other algorithms.

IoT sensors are extensively used in a variety of medical applications such as patient monitoring. In recent years, multiple papers have been published on the measurement of blood alcohol level (BAL) using more specialized biosensors. Transdermal alcohol content (TAC), a practical and non-invasive tool for measuring BAL, was employed in several recent research works. As BAL affects the person's way of walking, accelerometers (ACC) combined with TAC data fed to machine learning algorithms (MLA) were applied to predict the state, drunk or sober, of a person. However, the accuracy of prediction was not high enough to be considered reliable. In this paper, using an archived “BAR CRAWL” dataset, we implemented five MLA: Linear Discriminant Analysis (LDA), Decision Tree (DT), Random Forest (RF), Extra Trees (ET), and Ada Boost (AB), with a variety of features to accurately predict the state of a person based on his alcohol levels. Furthermore, we defined various alcohol thresholds to predict the level of intoxication. Also, we were able to identify the intoxicated person. Our experimental results showed that we can achieve up to 28.9 % higher prediction accuracy with less processing time (PT) when compared to previous published works using the same dataset.