International audience

International audience

International audience

International audience

International audience

Early detection of breast cancer is crucial for improving survival rates and patient outcomes. However, accurately identifying the early signs of breast cancer remains a significant challenge for healthcare professionals, often due to the subtlety of early-stage indicators and the complexity of interpreting medical data. In this paper, we propose a comprehensive AI-driven approach to assist healthcare professionals in the early detection of breast cancer. Comparing both Machine Learning (ML) and Deep Learning (DL) techniques, we conduct an extensive evaluation of several models, using the Wisconsin Breast Cancer dataset. Our analysis reveals that ensemble learning models, particularly Random Forest (RF) and Adaptive Boosting (AB), demonstrate superior performance, with both models achieving an accuracy of$\mathbf{9 8. 2 \%}$and an Area Under the Curve of$\mathbf{0. 9 9}$. These results underscore the potential of these models to accurately predict breast cancer, thereby helping in early detection.

The Internet of Medical Things (IoMT) leverages interconnected devices to enhance healthcare operations through data collection, monitoring, and automation. However, IoMT systems are vulnerable to cyberattacks, necessitating robust detection mechanisms. This paper presents a comprehensive comparison of traditional Machine Learning (ML) and Deep Learning (DL) models for detecting attacks in IoMT environments. We address class imbalance using the Synthetic Minority Over-sampling Technique and evaluate models based on various classification performance metrics and computational complexity. Our findings highlight that XGBoost (XGB) achieves the highest performance metrics, including accuracy (0.957), F1-score (0.955), and AUC (0.992), while maintaining competitive prediction times. In comparison, DL models, despite their high accuracy, exhibit significantly higher computational demands. This study demonstrates the effectiveness of XGB for real-time attack detection in IoMT systems, offering practical insights for selecting models that balance performance with computational efficiency. This research contributes to the development of more resilient IoMT systems by enhancing cyberattack detection strategies.

Anomaly detection in the Internet of Medical Things (IoMT) is important for ensuring the timely identification of potential health issues. To address this challenge, this paper presents a novel approach combining clustering and unsupervised learning with online adaptation. Patient data collected from IoMT sensors, including vital signs and health alerts, is first preprocessed to ensure quality and reliability. We then employ clustering algorithms to identify clusters containing normal data, which are then used to train unsupervised anomaly detection models. The models evaluated include Isolation Forest, One-Class Support Vector Machine (OCSVM), Local Outlier Factor (LOF), and Elliptic Envelope (EE). To maintain the model's effectiveness over time, an online learning mechanism updates the model with new data every 5000 samples. Our experimental results demonstrate that the LOF model provides the best performance with high precision, recall, F1-score, and Area Under the Curve (AUC) while maintaining computational efficiency with varying amounts of training data. For instance, with 30,000 samples, the LOF model achieved an accuracy of 0.867, and an AUC score of 0.964. Despite similar performance metrics, Spectral Clustering was found to be impractical due to its excessive training time. These results highlight the effectiveness and practicality of our approach for real-time anomaly detection in IoMT.

This paper presents a comprehensive approach for detecting respiratory diseases using IoMTs and Machine Learning (ML) algorithms, leveraging audio recordings from multiple sensor locations on the body. By capturing, extracting, and analyzing diverse audio features, such as MFCC, STFT, and Mel-spectrogram, we aim to detect and classify six respi-ratory conditions: Bronchiectasis, Bronchiolitis, COPD, Healthy, Pneumonia, and URTI. We used a publicly annotated dataset to conduct the experiment and analyze the performance of our proposed approach. This dataset underwent preprocessing, which included feature extraction, removal of rare diseases, data flattening, and encoding for model training. Our findings demonstrate that Deep Learning (DL), such as the Convolutional Neural Network (CNN) model achieved the highest accuracy of 92.4 % and an AU C of 97 %, highlighting its potential in audio-based diagnostics. Our experimental results prove that DL, particularly CNN, outperforms traditional ML techniques in detection accuracy, which makes them a good choice in developing non-invasive, efficient, and cost-effective solutions for respiratory disease detection.

With the growing importance of privacy in data-driven applications, ensuring the security and confidentiality of personal information has become a significant challenge. Federated Learning (FL) offers a promising solution by enabling collaborative model training across multiple clients while keeping individual data localized and private. In FL, outputs computed by various devices are aggregated at a central server, which uses iterative algorithms to develop a globally shared model. However, the presence of malicious participants can result in the intentional manipulation of training data or the model, compromising the system's accuracy and reliability. In this study, we propose a novel FL technique based on Split Learning (SL) to enhance robustness against data poisoning attacks. Our approach aims to develop a robust FL system based on SL, integrate it with existing aggregation methods, and compare its performance with traditional FL approaches.

Heart disease is a major health concern, and this study investigates how machine learning, particularly Random Forest (RF), can be used to identify it early and assess risk. We tested various algorithms using a large dataset (BRFSS 2022) and found that RF performed best, achieving an accuracy of 93% and an AUC of 98%. To make the model run faster, we identified the four most important features and focused on those. This reduced accuracy slightly (86% for RF and Extra Trees) but with AUCs of 0.92 and 0.91 respectively. Importantly, training and testing times improved significantly. This trade-off between accuracy and speed makes these models more suitable for real-world use, potentially aiding healthcare professionals in the early detection and prevention of heart disease.

Our study challenges the conventional understanding that deep learning models consistently outperform traditional machine learning approaches in classification tasks. By evaluating the performance of Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) and Random Forest (RF) models on QUIC traffic classification, we demonstrate that RF achieves superior accuracy, precision, recall, F1-score, and computational efficiency compared to CNN and LSTM. This finding underscores the importance of considering both performance and computational efficiency when selecting an appropriate model. Additionally, we emphasize the practical applicability of RF, especially in resource-constrained environments, where its efficiency makes it a compelling alternative to deep learning methods. These insights offer valuable guidance for enhancing network security, optimizing resource utilization, and deploying effective traffic classification systems in real-world scenarios.

Early detection of anomalies in network traffic data is critical for robust cybersecurity. This study investigates the effectiveness of various Machine Learning and Deep Learning models for identifying anomalous patterns in NetFlow v9 traffic. We address data preprocessing challenges and explore feature engineering techniques to optimize anomaly detection system performance. Our study evaluates the performance of several models based on key metrics like accuracy, Area Under the Curve (AUC), and computational efficiency. The results highlight the strengths and limitations of each model, emphasizing the importance of balancing performance with real-world deployment feasibility. Random Forest emerged as the most effective model, achieving an accuracy of 93.8% and an AUC of 0.99. Additionally, it demonstrated superior training and testing times, requiring only 0.19 seconds for training and 0.23 microseconds per prediction. Conversely, the Recurrent Neural Network model exhibited limitations in training efficiency and overall performance. Through a nuanced analysis of model performance and computational considerations, this study contributes to advancing anomaly detection techniques for network security applications.

In this paper, we propose an approach for anomaly detection in Internet of Medical Things based on the soft voting between the most accurate machine learning algorithms. We compare 12 machine learning and 5 deep learning algorithms for anomaly detection to identify the top 3. We apply these algorithms over public annotated dataset with network and physiological parameters. The soft voting predicts the anomaly based on the predicted probability by each individual model from the selected 3. We also compare the performance of the 17 algorithms before and after dimensionality reduction using two different techniques, and we found that soft voting using CatBoost, XGBoost and LightGBM outperforms hard voting and other algorithms, achieving a detection accuracy of 97.45 % and a false alarm rate of 2%.

Cyberbullying is a growing problem in today's digital society, with serious consequences for the mental health, professional life, and general well-being of victims. To combat this phenomenon, it is essential to effectively detect cases of cyberbullying online. In this article, we propose a cyberbullying detection model based on machine learning. We collect a dataset containing messages classified as cyberbullying, with multiclass labels. We then use machine learning techniques to extract relevant features from these messages and identify cases of cyberbullying. The results of our experiments show that our detection model achieves high accuracy, making it effective for detecting cyberbullying online. This research thus contributes to the fight against cyberbullying by proposing an innovative method for detecting this phenomenon online.

The purpose of this paper is to develop an intrusion detection system for IoT systems using deep learning and the genetic algorithm to optimize (or, more precisely, compress) the model. To meet the security requirements of IoT systems, we aim to design a powerful IDS system capable of detecting novel attacks, with a lightweight architecture and reduced classification time. We proposed two neural network architecture-based systems: the multi-layer perceptron and the auto-encoder model. The NSL-KDD dataset is utilized for the evaluation of these two systems.

Recent advancements in the realm of Graph Neural Networks (GNN) have introduced interesting opportunities for classifying and detecting anomalies in medical data. This paper explores the effectiveness of GNNs by utilizing various graphical representations, such as tree and bipartite graphs, to capture and exploit the intricate relationships inherent in medical data. Our investigation demonstrates that through this graph conversion process, GNNs acquire the capability to train and classify on a limited range of configurations, thereby improving both efficiency and classification performance. Furthermore, by employing complete graphs and multi-parametric temporal graphs, we can identify anomalies in health values by considering correlation factors. This sequential approach, where GNNs are repeatedly applied to graphs free from erroneous values, facilitates more precise classification, leading to exceptional accuracy rates. These findings highlight the crucial role of graphs in medical data analysis and demonstrate the effectiveness and robustness of GNNs in this specific domain.

This paper presents a comparative analysis of various machine learning models applied to a diabetes dataset. The study evaluates a range of ML algorithms, employing exploratory data analysis and feature engineering techniques. The primary objective is to forecast the onset of diabetes mellitus in a high-risk population of Pima Indians. Each model is trained, evaluated using performance metrics such as accuracy and F1-score, and assessed using ROC curves and AUC values. The results highlight the performance and suitability of each model for the prediction task. Notably, an Support Vector Machines (SVM) model achieved an accuracy of 77.27%. This study contributes valuable insights for selecting appropriate models in similar scenarios.

Bluetooth Low Energy (BLE) is extensively employed in various fields such as medication, home automation, transportation, and agriculture, enabling the integration of smart applications into resource-constrained devices that can operate for prolonged periods, often powered by a coin cell battery. While many BLE devices rely on the Just Works pairing option to establish connections with peer devices, the responsibility for security implementation falls on application developers and device manufacturers due to its lightweight nature. Unfortunately, the market's reluctance to invest in security measures has resulted in a proliferation of vulnerable smart devices. This paper aims to address the vulnerability of the Just Works pairing method in BLE. Through a case study involving a smart lightbulb, we intercept BLE exchanges between two entities and propose an algorithm to enhance the security of the Just Works pairing mechanism.

The deployment of the Internet of Medical Things (IoMTs) for remote monitoring has grown exponentially as an alternative for in-hospital diagnosis. The COVID-19 lockdown and the shortage of human and medical apparatus pushed the detection and diagnostics toward the Wireless Body Area Network to prevent the spread of infection. In this paper, we present a novel approach to detect cardiovascular diseases using the IoMTs. The proposed approach is based on the use of autoencoders to derive a compressed representation of the input record, followed by a Random Forest to classify the latent data. The objective is to provide a lightweight and accurate model compatible with the constrained resources of sensors. Our experiment results on a public annotated dataset show that our approach can enhance the performance where the obtained area under the curve is 89%. We also compare the performance of widely used classification algorithms to prove the efficiency of our proposed model.

The objective of this paper is to propose a new system to assist drivers with hearing disability, deaf or unfocused persons by recognizing and transforming audible signals, such as emergency vehicle sirens or honks into alerts displayed in the dashboard. Such conversion from audio to alert messages attracts the attention of unfocused drivers to hear the honking of other cars, and enhances the safety and the quality of life for deaf or hard-of-hearing drivers. We develop an IoT based system to identify, denoise and translate any significant voice signal around the driver's car into alert messages. The signal acquired by sensors is processed to identify the source and display the associated message through the use of several machine learning models with majority voting. Our experiments results show that the proposed solution is able to achieve a 95% accuracy when trained and validated against a real dataset of 600 files.

Cyberbullying is a prevalent social issue that can cause significant harm to individuals, particularly young people. Traditional approaches to detecting cyberbullying rely on text-based analysis of online messages, but the use of multiple modes of communication in cyberbullying presents challenges for detection. In this study, we compare the effectiveness of audio-only, visual-only, and text-only approaches for cyberbullying detection. We collected different datasets for the study and used machine learning algorithms, speech emotion recognition, text conversion, middle finger, and landmarks techniques to analyze the data. Our results show that the text-only approach achieved the highest accuracy of 95%. Our study demonstrates that text-based features remain the most informative for cyberbullying detection, while audio and visual features alone are less effective. These findings suggest that the development of more comprehensive cyberbullying detection methods should focus on improving text-based analysis while considering the potential benefits of incorporating audio and visual features.

The exponential development and widespread emergence of the Internet of Medical Things (IoMT) have led to a growing need for effective anomaly detection techniques to ensure the reliability and security of healthcare systems. This article provides a review of existing machine learning and deep learning algorithms for anomaly detection in IoMT, followed by the presentation of a novel approach combining ARIMA for predicting health parameter values and a decision tree for anomaly detection. This hybrid approach aims to improve the accuracy and efficiency of anomaly detection in IoMT by leveraging both time series models and the discriminative features of decision trees. The preliminary results of this approach are presented and discussed, highlighting its potential to enhance early detection of anomalies in IoMT and contribute to safer and more reliable healthcare.

This paper proposes a dual autoencoder approach for anomaly detection in industrial network systems. Our system is based on two separate autoencoders: one trained on raw sensor data and the other on derived features, both aimed at capturing normal system behavior. Anomalies are identified by comparing reconstruction errors from both autoencoders and employing classification algorithms such as ANN. The proposed approach achieves a high accuracy on benchmark datasets, surpassing traditional anomaly detection methods.

Bluetooth Low Energy (BLE) is widely applied in various gadgets and Internet of Things (IoT) packages through-out diverse fields which include healthcare, home automation, transportation, and agriculture. It evolves classic Bluetooth generation, permitting integration into aid-constrained gadgets that can operate for months, or even years, on an unmarried coin cell battery. To establish connections with peer gadgets, the bulk of BLE devices in the marketplace hire the “Just Works” pairing approach. This lightweight mode locations the onus of protection implementation on application builders and device producers. Unfortunately, marketplace pressures frequently sideline safety issues, leading to a plethora of inclined clever devices. In this paper, we observe the vulnerabilities related to the “Just Works” pairing in BLE. We behavior a case examination related to a smart mild bulb to intercept BLE communications between entities. Subsequently, we advocate an algorithm aimed at improving the safety of the “Just Works” pairing method.

In this article, we present a method for detecting cyberbullying that goes beyond understanding the text's content and also considers the underlying intent. Our approach involves analyzing both the acoustic and linguistic features of audio to gain insights into the emotional intent conveyed in the text. Using the librosa library, we extract acoustic features from audio files to identify whether the emotions expressed are aggressive or non-aggressive. We then combine this information with the results of linguistic analysis to develop a comprehensive understanding of the text's content. By incorporating the emotional aspect, our method enables more precise and nuanced identification of cyberbullying instances.

Cyberbullying is a growing concern, with serious consequences especially for children. In this paper, we propose a machine learning approach to detect cyberbullying messages accurately and distinguish them from regular ones. We used a public dataset of social media messages to fit a binary classification of either cyberbullying or non-bullying messages. We compared 15 classifiers using two methods: Term Frequency-Inverse Document Frequency (TF-IDF) for traditional algorithms and word embedding for deep learning algorithms. The voting classifier, a combination of the best algorithms from the first method, achieved the highest accuracy of 96.5% during tests. This approach can be used in social media or chat applications to detect and prevent cyberbullying.

In this paper, we propose a secure framework for healthcare monitoring using the Internet of Medical Things (IoMT). In spite of their deployment, these devices still vulnerable to several cyber-attacks, ranging from unauthorized access to private medical data to data modification and injection. These attacks can compromise the privacy of the monitored patient, reduce the reliability of the monitoring system and may harm the life of monitored patient. In this paper, we propose a new framework to detect attacks and secure the communications in IoMT. To prevent eavesdropping and modification attacks, we propose the Ephemeral Elliptic Curves Diffie-Hellman (EECDH) to derive a session key used to provide confidentiality and authenticity. To detect injected measurements, flooding triggered by compromised devices and medical changes in physiological data, we applied the sequential change point detection algorithm Pruned Exact Linear Time (PELT) followed by the boxplot. Our experimental results show that our approach is able to increase the reliability and the accuracy of remote monitoring system, while reducing the false alarms triggered by injected measurements.

To improve the communications between the deaf and the hearers using hand-held devices, we propose a lightweight approach to quickly identify the word in American Sign Language (ASL). We acquire inertial data and muscular activity during hands movements. Then we aggregate the received data to reduce the required processing complexity and memory usage in a portable device. Afterward, we feed extracted features from aggregated data into the Support Vector Machine (SVM) to identify the associated word. Our experimental results showed that our data aggregation approach was able to enhance the recognition accuracy of the associated word when comparing the performance of SVM and Decision Tree (DT) classifiers with and without data aggregation. We conduct a performance analysis and we showed that our proposed approach is faster and able to achieve better recognition accuracy (92%) when compared with existing work.

The Internet of Medical Things are susceptible to Man-in-the-Middle (MitM) attack, which can identify healthcare emergency of monitored patients and replay normal physiological data to prevent the system from raising an alarm. In this article, we propose a framework to prevent a MitM from disrupting the operations and prohibiting the raise of alarms by the remote healthcare monitoring system. To reduce energy consumption for normal data transmission, and preserve the privacy of health data, our framework transmits a smaller size signature derived from acquired data with message authentication code, where the key is derived from received signal strength indication. Our experimental results for emergency detection show that our approach can achieve a high detection accuracy with a low false alarm rate of 3%.

The use of Wireless Body Area Networks (WBANs) in healthcare for pervasive monitoring enhances the lives of patients and allows them to fulfill their daily life activities while being monitored. Various non-invasive sensors are placed on the skin to monitor several physiological attributes, and the measured data are transmitted wirelessly to a centralized processing unit to detect changes in the health of the monitored patient. However, the transferred data are vulnerable to various sources of interference, sensor faults, measurement faults, injection and alteration by malicious attackers, etc. In this article, we propose a change point detection model based on a Markov chain for centralized anomaly detection in WBANs. The model is derived from the Root Mean Square Error (RMSE) between the forecasted and measured values for whole attributes. The RMSE transforms the monitored attributes into a univariate times series which is divided into overlapping sliding window. The joint probability of the sequence of RMSE values in each sliding window is calculated to decide whether a change has occurred or not. When an effective change is detected over k consecutive windows, the number of deviated attributes is used to distinguish faulty measurements from a health emergency. We apply our proposed approach on real physiological data from the Physionet database and compare it with existing approaches. Our experimental results prove the effectiveness of our proposed approach, as it achieves high detection accuracy with a low false alarm rate (5.2%).

In this paper, we propose a novel model for real-time detection of nocturnal epileptic seizures from physiological and inertial data, which are collected by a wireless wristband with integrated muscular activity sensors, 3D accelerometer and 3D gyroscope. The wristband transmits the data to a portable unit for processing and seizures detection. It must be able to distinguish normal nocturnal movements from seizures, and to raise an alarm upon detection of seizures for patient relatives and for predefined contacts. Our real-time detection model starts by reducing the dimensionality of collected data through the use of root mean square to derive one signal from 3D accelerometer and one signal from 3D gyroscope. With the derived 3 signals (accelerometer, gyroscope and electromyogram), we apply the vector triple product to derive one signal used as input for anomaly detection mechanism. The robust version of z-score is applied on the resulting product signal to detect deviations associated with seizures before raising an alarm for patient relative for assistance to prevent further injuries when the patient loses control with the excessive discharge of neurons.

Accurate and fast recognition of sign language would greatly improve communications between the deaf and the hearers using hand-held devices. We used Myo armband as our wireless data measurement device, which is wearable technology equipped with on-board 3D Accelerometer, 3D Gyroscope and 8 channel Electromyogram acquisition system. The main objective of this paper is to provide a lightweight approach for American Sign Language recognition by reducing the dimensionality of inputs using a novel method. Data from each sensors are aggregated into one dimension to reduce the required time for data processing, as well as the amount of required memory for storage. Afterward, we extract features from aggregated data and we proceed to classification using Support Vector Machine (SVM). We compare the performance of SVM with and without our aggregation approach. Our experimental results prove that our proposed approach improves the speed of model derivation (four times faster than existing methods) and reduces the size of input data with the same accuracy.

The accurate individuals' activities description is an important assignments in human computer interface. Many applications can be developed based on correct and fast activity recognition, such as healthcare monitoring and fall detection applications. Human Activity Recognition (HAR) is an active research field, where some key factors still challenging and need to be enhanced for faster and accurate recognition. This paper presents a novel approach aims to improve the activity recognition time by aggregating the raw data from inertial sensors into one time series used as input, then various features are extracted from the resulted signal by dimensionality reduction procedure, Those features are used as input data for the classification algorithm without affecting the information associated with activity in the raw data. The results of the novel method showed an improvement in recognition time while retaining the same level of accuracy.

This paper presents a lightweight approach for the early detection of nocturnal epileptic seizures through analysis of inertial data and muscle contractions. Our approach uses an overlapping sliding window to derive the variance of data acquired by the MPU 9,250 motion tracking device and single channel surface ElectroMyoGram (sEMG). The Exponentially Weighted Moving Average (EWMA) is used to forecast the current value of the data variance. When the Kullback-Leibler divergence between the forecasted and measured variances deviates from past values, a signal is transmitted to the base station to set the current counter in an alarm window. If the filling ratio of the alarm window is greater than a predefined threshold, an alarm is triggered by the base station. The proposed approach is intended to improve the performance of existing detection systems based on data analysis from Accelerometer. The MPU 9,250 is 9-axis motion tracking and used to detect motor seizures, and it contains a 3-axis Accelerometer, Gyroscope, and Magnetometer. The sEMG is used to detect silent seizures without jerky movements. Our experimental results on a real dataset from an epileptic patient show that our proposed approach is able to increase detection accuracy and reduce the low false alarm rate. Comparison with a Probability Density Function (PDF) further demonstrates the detection efficiency of our approach.

The collected data by biomedical sensors must be analyzed for automatic detection of physiological changes. The early identification of an event in collected data is required to trigger an alarm upon detection of patient health degradation. Such alarms inform healthcare professionals and allow them to quickly react by taking appropriate actions. However, events result from physiological change or faulty measurements, and lead to false alarms and unnecessary medical intervention. In this paper, we propose a framework for automatic detection of events from collected data by biomedical sensors. The proposed approach is based on the Kalman filter to forecast the current measurement and to derive the baseline of the time series. The power divergence is used to measure the distance between the forecasted and measured values. When a change occurs, this metric significantly deviates from past values. To distinguish emergency events from faulty measurements, we exploit the spatial correlation between the monitored attributes. We conduct experiments on real physiological data set and our results show that our proposed framework achieves a good detection accuracy with a low false alarm rate. Its simplicity and processing speed make our proposed framework efficient and effective for real-world deployment.

The goal of this paper is to provide a lightweight approach for the early detection of epileptic seizures using data from inertial measurement unit and muscular activity. The detection procedure runs in a portable data collection device and raises an alarm for family member or other persons in the vicinity for fast assistance and eventually for saving the life of the monitored patient. The instantaneous power of sliding window is derived for inertial measurements from 3D Accelerometer (ACM), 3D Gyroscope (Gyro) and for the muscular activity from the ElectroMyoGram (EMG). The residual between forecasted and measured power is used as input for the detection algorithm based on Shewhart control chart. When the error between forecasted and derived power exceeds statistical chart limits [lower, upper] for several consecutive slots, an alarm is raised. The proposed approach is intended to improve the performance of existing detection systems by increasing the detection accuracy and reducing the false alarms through correlation analysis of collected data from 3D ACM, 3D Gyro and EMG. Our experimental results on real data set collected in Necker hospital from epileptic patients show that our proposed approach is robust against nocturnal movements and achieves a high level of detection accuracy with low false alarm rate.

The main goal of this paper is to propose an effective approach for the early detection of nocturnal seizures from multi-signals gathered by an Armband. An inertial measurement unit and muscular activities acquisition sensors are integrated in the armband, and used to collect acceleration, angular velocity and electromyogram from the arm of the monitored patient. These signals are transmitted from the arm-band to a portable unit (e.g., SmartPhone) for preprocessing, detection and identification of seizures. Our approach starts by deriving the root mean square for acceleration and gyroscope, followed by the normalization of whole signals in the same range, and aggregation into one signal. The chart's control with its upper and lower limits are derived in the training phase (when the patient at rest) and used to detect abnormal seizures and to raise an alarm for patient relative for assistance to prevent further injuries when he loses consciousness.

The electromyogram signals generated by muscles are used in numerous fields such as augmented reality, biomedical, kinematics, gaming, 3D animations and Human-Machine Interfaces. The latter is specifically used to help persons with reduced mobility or amputee with specific constraints to remote control machines. In this paper, we present a novel EMG-based system that aims to control multimedia player in simple, efficient and flexible manner. The implementation of our proposed approach was realized in order to achieve experiments and to conduct performance analysis. Our approach uses pattern recognition and contraction duration to derive four predefined actions. Our experimental results show the capacity of our system to achieve good detection accuracy of user EMG-based commands and to translate these commands into actions in media player system.

The goal of this paper is to propose a new approach for controlling smart home devices using physiological and movements signals (electromyogram, accelerometer and gyroscope). Our proposed approach exploits the ElectroMyoGram (EMG) signal to detect user muscles contraction and to trigger the associated actions on the controlled device. The triggered action is based on the position of the hand or its movement pattern. The Support Vector Machine (SVM) is used to classify the gyroscope data and to detect hand movements. The position of the hand is determined using the K-Nearest Neighbors (KNN) algorithm. Our proposed approach can be used to command any electronic device or connected objects and it is intended to work with data from wearable arm band containing triaxial accelerometer, gyroscope and able to measure EMG signal. Our experimental results are very encouraging where we achieve fast processing and reliable evaluation of hand movement with very low rate of actions miss interpretation.

Cardiovascular diseases (CVD) are known to be the most widespread causes to death. Therefore, detecting earlier signs of cardiac anomalies is of prominent importance to ease the treatment of any cardiac complication or take appropriate actions. Electrocardiogram (ECG) is used by doctors as an important diagnosis tool and in most cases, it's recorded and analyzed at hospital after the appearance of first symptoms or recorded by patients using a device named holter ECG and analyzed afterward by doctors. In fact, there is a lack of systems able to capture ECG and analyze it remotely before the onset of severe symptoms. With the development of wearable sensor devices having wireless transmission capabilities, there is a need to develop real time systems able to accurately analyze ECG and detect cardiac abnormalities. In this paper, we propose a new CVD detection system using Wireless Body Area Networks (WBAN) technology. This system processes the captured ECG using filtering and Undecimated Wavelet Transform (UWT) techniques to remove noises and extract nine main ECG diagnosis parameters, then the system uses a Bayesian Network Classifier model to classify ECG based on its parameters into four different classes: Normal, Premature Atrial Contraction (PAC), Premature Ventricular Contraction (PVC) and Myocardial Infarction (MI). The experimental results on ECGs from real patients databases show that the average detection rate (TPR) is 96.1% for an average false alarm rate (FPR) of 1.3%.

The sleep apnea is a sleep disorder characterized by cessation of respiratory flow (apnea) or a reduction in the flow (hypopnea). This disorder is often invalidating and may in some cases lead to death. During the night, symptoms can include nocturnal choking, heavy snoring, sweating, restless sleep, impotence, and witnessed apnea. As the sleep centers for apnea detection are usually overloaded and inaccessible, an automatic apnea detection algorithm for portable devices is required for in-home detection. In this paper, we propose a lightweight approach for pervasive detection of sleep apnea using Wireless Sensor Networks. The experimental results show that our proposed approach achieves good detection accuracy with low delay and low false alarm rate.

The aim of this paper is to provide a lightweight approach for early detection of nocturnal epileptic seizures using data from wireless 3-D accelerometer sensors. We use the exponentially weighted moving average algorithm to forecast the current value of the accelerometer measurement, and when the difference between measured and forecasted values is greater than the dynamic threshold on any axis, a notification is transmitted to the base station, which maintains a sliding window of received notifications. When the filling ratio is greater than a predefined threshold, an alarm is triggered by the base station. The proposed approach is intended to improve the performance of existing mobile health detection systems based on the analysis of electroencephalogram (EEG). To reduce their false alarm rate, we seek to correlate detection results from 3-D accelerometer with other physiological parameters through a majority voting. Our experimental results on real dataset collected from the epileptic patient show that our proposed approach is robust against temporal fluctuations and achieves a high level of detection accuracy, which in turn proves the effectiveness of this approach in enhancing the reliability of existing detection approaches based on EEG signal analysis.

In this paper, we propose an Anomaly Detection (AD) approach for medical Wireless Sensor Networks (WSNs). This approach is able to detect abnormal changes and to cope with unreliable or maliciously injected measurements in the network, without prior knowledge of anomalous events or normal data pattern. The main objective is to reduce the false alarms triggered by abnormal measurements. In our proposed framework, each sensor applies the Exponentially Weighted Moving Average (EWMA) for one-step forecasting. To reduce the energy consumed by periodic data transmission to the Local Processing Unit (LPU), the sensor transmits only when the data point (measured, expected) falls outside the dynamically updated ellipsoidal region enclosing the normal data. The LPU exploits correlation and uses chi-square distance for spatial analysis before raising a medical alarm. We evaluate our approach on real medical data set. Experimental results through computer simulation demonstrate that our proposed approach can achieve a good detection accuracy with low false alarm rate (less than 4%).

Electroencephalogram (EEG) is the electrical signal of brain which contains valuable information about its activities. In this paper, we propose a new approach for the early detection of epileptic seizure in EEG. The proposed approach is based on Discrete Wavelet Transform (DWT) and Ant Colony (AC) Classifier. We started by applying DWT to decompose the EEG signal into its sub-bands to extract the energy ratio from wavelet coefficients. Beside we extract some statistical features from the original signal, and we use the extracted features as the input for the AC algorithm to derive classification rules, which are used to detect epileptic seizures in the EEG of the monitored patient. Our experimental results on real dataset show that our proposed approach achieves a high level of detection accuracy.

In this paper, we propose a lightweight approach for online detection of faulty measurements by analyzing the data collected from medical wireless body area networks. The proposed framework performs sequential data analysis using a smart phone as a base station, and takes into account the constrained resources of the smart phone, such as processing power and storage capacity. The main objective is to raise alarms only when patients enter in an emergency situation, and to discard false alarms triggered by faulty measurements or ill-behaved sensors. The proposed approach is based on the Haar wavelet decomposition, nonseasonal Holt-Winters forecasting, and the Hampel filter for spatial analysis, and on for temporal analysis. Our objective is to reduce false alarms resulting from unreliable measurements and to reduce unnecessary healthcare intervention. We apply our proposed approach on real physiological dataset. Our experimental results prove the effectiveness of our approach in achieving good detection accuracy with a low false alarm rate. The simplicity and the processing speed of our proposed framework make it useful and efficient for real time diagnosis.

Cardiovascular diseases are the leading cause of death in the world, and Myocardial Infarction (MI) is the most serious one among those diseases. Patient monitoring for an early detection of MI is important to alert medical assistance and increase the vital prognostic of patients. With the development of wearable sensor devices having wireless transmission capabilities, there is a need to develop real-time applications that are able to accurately detect MI non-invasively. In this paper, we propose a new approach for early detection of MI using wireless body area networks. The proposed approach analyzes the patient electrocardiogram (ECG) in real time and extracts from each ECG cycle the ST elevation which is a significant indicator of an upcoming MI. We use the sequential change point detection algorithm CUmulative SUM (CUSUM) to early detect any deviation in ST elevation time series, and to raise an alarm for healthcare professionals. The experimental results on the ECG of real patients show that our proposed approach can detect MI with low delay and high accuracy.

The aim of this paper is to propose a new approach for the detection and isolation of faulty measurements in medical wireless sensors networks. The proposed approach is based on the combination of statistical model and machine learning algorithm. We begin by collecting physiological data and then we cluster the data collected during the first few minutes using the Gaussian mixture decomposition. We use the resulted labeled data as the input for the Ant Colony algorithm to derive classification rules, which are used to detect abnormal values. Finally, we exploit the spatial correlation between monitored attributes to differentiate between faulty sensor readings and emergency situations. Our experimental results on real patient dataset show that our proposed approach achieves a high level of detection accuracy, which in turn proves the effectiveness of this approach in enhancing the reliability of medical wireless sensors networks.

Wireless Sensor Networks are vulnerable to a plethora of different fault types and external attacks after their deployment. We focus on sensor networks used in healthcare applications for vital sign collection from remotely monitored patients. These types of personal area networks must be robust and resilient to sensor failures as their capabilities encompass highly critical systems. Our objective is to propose an anomaly detection algorithm for medical wireless sensor networks. Our proposed approach firstly classifies instances of sensed patient attributes as normal and abnormal. Once we detect an abnormal instance, we use regression prediction to discern between a faulty sensor reading and a patient entering into a critical state. Our experimental results on real patient datasets show that our proposed approach is able to quickly detect patient anomalies and sensor faults with high detection accuracy while maintaining a low false alarm ratio.

In this paper, we focus on online detection and isolation of erroneous values reported by medical wireless sensors. We propose a lightweight approach for online anomaly detection in collected data, able to raise alarms only when patients enter in emergency situation and to discard faulty measurements. The proposed approach is based on Haar wavelet decomposition and Hampel filter for spatial analysis, and on boxplot for temporal analysis. Our objective is to reduce false alarms resulted from unreliable measurements. We apply our proposed approach on real physiological data set. Our experimental results prove the effectiveness of our approach to achieve good detection accuracy with low false alarm rate.

The increasing use of wireless networks and the constant miniaturization of electrical devices has empowered the development of Wireless Body Sensor Networks (WBSNs). The wireless nature of the network and the wide variety of sensors offer numerous new, practical and innovative applications to improve health care and the Quality of Life. WBSNs like any other sensor networks suffer limited energy resources and hence preserving the energy of the nodes is of great importance. Unlike typical sensor networks WBSNs have few and dissimilar sensors. In addition, an extremely low transmit power per node is needed to minimize interference to cope with health concerns and to avoid tissue heating which means that the existing solution for preserving energy in wireless sensor networks might not be efficient in WBSNs. Most of the attention has been given to the energy routing where energy awareness is an essential consideration. In this paper, we propose a Min-Max multi-commodity flow model for WBSNs which allows to prevent sensor node saturation, by imposing an equilibrium use of sensors during the routing process taking into account the specific characteristics of the wireless environment on the human body. The Min-Max objective is transformed to a Min objective by adding a set of constraints to the model. Based on the energy consumption for sending and receiving data and the available residual energy of nodes, the max-min based mathematical programming model is designed to find optimal routing. Simulation results show that the algorithm balances the energy consumption of nodes effectively and maximize the network lifetime.

Anomaly detection in high speed networks is well known to be a challenging problem. It requires generally the analysis of a huge amount of data with high accuracy and low complexity. In this paper, we propose an anomaly detection mechanism against flooding attacks in high speed networks. The proposed mechanism is based on Jensen-Shannon divergence metric over sketch data structure. This sketch is used to reduce the required memory, while monitoring the traffic, by maintaining them into a predefined fixed size of hash tables. This sketch is also used to develop a probabilistic model. The Jensen-Shannon divergence is used for detecting deviations between previously established and current distributions of network traffic. We have implemented our approach and evaluated it using real Internet traffic traces, obtained from MAWI trans-Pacific wide transit link between USA and Japan. Our results show that the proposed approach is scalable and efficient in detecting anomalies without maintaining per-flow state information.

Internet services are vulnerable to flooding attacks that lead to denial of service. This paper proposes a new framework to detect anomalies and to provide early alerts for flooding attacks in backbone networks. Thus allow to quickly react in order to prevent the flooding attacks from strangling the victim server and its access network. The proposed detection scheme is based on the application of Least Mean Square (LMS) filter and Pearson Chi-square divergence on randomly aggregated flows in Sketch data structure. Instead of analyzing one time series for overall traffic, random aggregation of flows is used to investigate a fixed number of time series for grained analysis. Least mean square filter is used to predict the next value of the time series based on previous values, and Pearson Chi- square divergence is used to measure the deviations between the current and estimated probability distributions. We evaluate our approach using publicly available real IP traces (MAWI) collected from the WIDE backbone network, on trans-Pacific transit link between Japan and USA. Our experimental results show that the proposed approach outperforms existing techniques in terms of detection accuracy and false alarm rate. It is able to detect low intensity attacks covered by the large number of traffic in high speed network.