Sylvie Charbonnier

EMG feature evaluation for improving myoelectric pattern recognition robustness

In pattern recognition-based myoelectric control, high accuracy for multiple discriminated motions is presented in most of related literature. However, there is a gap between the classification accuracy and the usability of practical applications of myoelectric control, especially the effect of long-term usage. This paper proposes and investigates the behavior of fifty time-domain and frequency-domain features to classify ten upper limb motions using electromyographic data recorded during 21days. The most stable single feature and multiple feature sets are presented with the optimum configuration of myoelectric control, i.e. data segmentation and classifier. The result shows that sample entropy (SampEn) outperforms other features when compared using linear discriminant analysis (LDA), a robust classifier. The averaged test classification accuracy is 93.37%, when trained in only initial first day. It brings only 2.45% decrease compared with retraining schemes. Increasing number of features to four, which consists of SampEn, the fourth order cepstrum coefficients, root mean square and waveform length, increase the classification accuracy to 98.87%. The proposed techniques achieve to maintain the high accuracy without the retraining scheme. Additionally, this continuous classification allows the real-time operation.

Trends extraction and analysis for complex system monitoring and decision support

This paper presents an effective trend extraction procedure, based on a simple, yet powerful, representation. Its usefulness for complex system monitoring and decision support is illustrated by three examples. The method extracts semi-qualitative temporal episodes on-line, from any univariate time series. Three primitives are used to describe the episodes: {Increasing, Decreasing, Steady}. The method uses a segmentation algorithm, a classification of the segments into seven temporal shapes and a temporal aggregation of episodes. It acts on noisy data, without prefiltering. The first illustration is devoted to decision support in intensive care units. The signals contain information and noise at very different frequencies, and smoothing must not mask some interesting high-frequency data features. The second illustration is dedicated to a food industry process. On-line trends of key variables represent a very useful monitoring tool to control the end product quality despite high variations of raw materials at the input and a long delay. The last example concerns operator support and predictive maintenance. The results issued from a diagnostic module are complemented by the extrapolation of the key variable trends, which gives an idea of the time left to repair or reconfigure the process.

Feature selection for sleep/wake stages classification using data driven methods

This paper focuses on the problem of selecting relevant features extracted from human polysomnographic (PSG) signals to perform accurate sleep/wake stages classification. Extraction of various features from the electroencephalogram (EEG), the electro-oculogram (EOG) and the electromyogram (EMG) processed in the frequency and time domains was achieved using a database of 47 night sleep recordings obtained from healthy adults in laboratory settings. Multiple iterative feature selection and supervised classification methods were applied together with a systematic statistical assessment of the classification performances. Our results show that using a simple set of features such as relative EEG powers in five frequency bands yields an agreement of 71% with the whole database classification of two human experts. These performances are within the range of existing classification systems. The addition of features extracted from the EOG and EMG signals makes it possible to reach about 80% of agreement with the expert classification. The most significant improvement on classification accuracy is obtained on NREM sleep stage I, a stage of transition between sleep and wakefulness.

On-Line Detection of Drowsiness Using Brain and Visual Information

A drowsiness detection system using both brain and visual activity is presented in this paper. The brain activity is monitored using a single electroencephalographic (EEG) channel. An EEG-based drowsiness detector using diagnostic techniques and fuzzy logic is proposed. Visual activity is monitored through blinking detection and characterization. Blinking features are extracted from an electrooculographic (EOG) channel. Features are merged using fuzzy logic to create an EOG-based drowsiness detector. The features used by the EOG-based detector are voluntary restricted to the features that can be automatically extracted from a video analysis of the same accuracy. Both detection systems are then merged using cascading decision rules according to a medical scale of drowsiness evaluation. Merging brain and visual information makes it possible to detect three levels of drowsiness: “awake,” “drowsy,” and “very drowsy.” One major advantage of the system is that it does not have to be tuned for each driver. The system was tested on driving data from 20 different drivers and reached 80.6% correct classifications on three drowsiness levels. The results show that EEG and EOG detectors are redundant: EEG-based detections are used to confirm EOG-based detection and thus enable the false alarm rate to be reduced to 5% while the true positive rate is not decreased, compared with a single EOG-based detector.

Self-evaluated automatic classifier as a decision-support tool for sleep/wake staging

An automatic sleep/wake stages classifier that deals with the presence of artifacts and that provides a confidence index with each decision is proposed. The decision system is composed of two stages: the first stage checks the 20s epoch of polysomnographic signals (EEG, EOG and EMG) for the presence of artifacts and selects the artifact-free signals. The second stage classifies the epoch using one classifier selected out of four, using feature inputs extracted from the artifact-free signals only. A confidence index is associated with each decision made, depending on the classifier used and on the class assigned, so that the users confidence in the automatic decision is increased. The two-stage system was tested on a large database of 46 night recordings. It reached 85.5% of overall accuracy with improved ability to discern NREM I stage from REM sleep. It was shown that only 7% of the database was classified with a low confidence index, and thus should be re-evaluated by a physiologist expert, which makes the system an efficient decision-support tool.

On-line automatic detection of driver drowsiness using a single electroencephalographic channel

In this paper, an on-line drowsiness detection algorithm using a single electroencephalographic (EEG) channel is presented. This algorithm is based on a means comparison test to detect changes of the alpha relative power ([8–12]Hz band). The main advantage of the method proposed is that the detection threshold is completely independent of drivers and does not need to be tuned for each person. This algorithm, which works on-line, has been tested on a huge dataset representing 60 hours of driving and give good results with nearly 85% of good detections and 20% of false alarms.

Selection of the Most Relevant Physiological Features for Classifying Emotion

With the development of wearable physiological sensors, emotion estimation becomes a hot topic in the literature. Databases of physiological signals recorded during emotional stimulation are acquired and machine learning algorithms are used. Yet, which are the most relevant signals to detect emotions is still a question to be answered. In order to better understand the contribution of each signal, and thus sensor, to the emotion estimation problem, several feature selection algorithms were implemented on two databases freely available to the research community (DEAP and MANHOB-HCI). Both databases manipulate emotions by showing participants short videos (video clips or part of movies respectively). Features extracted from Galvanic Skin response were found to be relevant for arousal estimation in both databases. Other relevant features were eye closing rate for arousal, variance of zygomatic EMG for valence (those features being only available for DEAP). The hearth rate variability power in three frequency bands also appeared to be very relevant, but only for MANHOB-HCI database where heat rate was measured using ECG (whereas DEAP used PPG). This suggests that PPG is not accurate enough to estimate HRV precisely. Finally we showed on DEAP database that emotion classifiers need just a few well selected features to obtain similar performances to literature classifiers using more features.

Mental fatigue and working memory load estimation: Interaction and implications for EEG-based passive BCI

Current mental state monitoring systems, a.k.a. passive brain-computer interfaces (pBCI), allow one to perform a real-time assessment of an operators cognitive state. In EEG-based systems, typical measurements for workload level assessment are band power estimates in several frequency bands. Mental fatigue, arising from growing time-on-task (TOT), can significantly affect the distribution of these band power features. However, the impact of mental fatigue on workload (WKL) assessment has not yet been evaluated. With this paper we intend to help fill in this lack of knowledge by analyzing the influence of WKL and TOT on EEG band power features, as well as their interaction and its impact on classification performance. Twenty participants underwent an experiment that modulated both their WKL (low/high) and time spent on the task (short/long). Statistical analyses were performed on the EEG signals, behavioral and subjective data. They revealed opposite changes in alpha power distribution between WKL and TOT conditions, as well as a decrease in WKL level discriminability with increasing TOT in both number of statistical differences in band power and classification performance. Implications for pBCI systems and experimental protocol design are discussed.

Statistical and fuzzy models of ambulatory systolic blood pressure for hypertension diagnosis

A solution to reduce Ambulatory Systolic Blood Pressure (ASBP) variability during the analysis of a 24-h profile is studied and presented in this article. It consists of equipping the portable ASBP recorder device with other sensors, a three-axes accelerometer and a heart rate recorder, so as to enable an analysis of the tensional profile in the light of these concomitant data. A database has been collected, and two models linking ASBP variations with body acceleration and heart rate measurements are developed: a regression one, based on a priori knowledge of ASBP variations, and a fuzzy one, automatically built from experimental data. Their performances are tested in prediction and compared. Then, the results are compared to those obtained with one of the solutions currently used by the physicians to deal with ASBP variability. The results obtained on 16 young subjects from the database are significantly improved and thus, encouraging.

Comparison between EOG and high frame rate camera for drowsiness detection

Drowsiness is responsible for a large number car crashes. Blinks analysis from electrooculogram (EOG) signal brings reliable information on drowsiness but EOG recording condition can be really disturbing for the driver. On the other hand, video approaches seem a lot more practical but the standard acquisition rate does not give the same accuracy than EOG for blinks analysis. So, a high frame rate camera seems a good compromise. The purpose of this paper is to study to what extent a high speed camera could replace the EOG for the extraction of blinks features in order to design a system to detect drowsiness. An original method to detect and characterize blinks from the video is presented. This method uses two energy signals extracted from the video analysis: one related to the contours of the eyes and the other one to the moving contours. A comparison between the different features extracted from the EOG and from the video is then performed. This study shows that duration, frequency, PERCLOS 80 and dynamic features extracted from the EOG and from the video signals are highly correlated. The frame rate influence on the accuracy of the different features extracted is also studied.