Pierre Jallon

Separation of Convolutive Mixtures of Cyclostationary Sources: A Contrast Function Based Approach

Fourth-order cumulants are quite popular in the field of blind separation of convolutive mixtures of stationary sources. Their use in the context of cyclo-stationary sources cannot be taken for granted because consistent estimation of the temporal mean of the fourth-order cumulants needs the knowledge of the cyclic frequencies of the received signal. In this paper, we introduce a cost function whose estimation does not need the knowledge of the cyclic frequencies. We show that under some reasonable sufficient conditions, its maximization allows to separate the sources.

Separation of instantaneous mixtures of cyclo-stationary sources

It has been observed that, generally, the Comon algorithm successfully achieves to separate instantaneous mixtures of non-stationary signals, despite it has been designed for stationary environments. We address the theoretical justification of this fact. We provide a tractable condition on the statistics of the sources ensuring that the Comon function is a contrast. This condition is worked out for various digital communication signals. We finally explain why the Comon and JADE algorithms remain strongly connected in a cyclo-stationary context.

Second-Order Based Cyclic Frequency Estimates: The Case of Digital Communication Signals

Both the standard estimate of the second-order cyclic frequency of a digital communication signal and its improved versions (based on a weighted criterion or a denoising of the cyclic periodogram) do not take into account a key property of the signal, indeed, the support of the cyclospectrum at the true cyclic frequency is a narrow interval centered around half this cyclic frequency. This fact is taken into account in this contribution. A theoretical fact explains the improvement over the standard method. Simulations confirm the expectation

A graph based method for timed up & go test qualification using inertial sensors

A graph based classifier is proposed to recognize the different time phases of the up & go test based on signals collected by an inertial sensor set on a person chest. This test being a sequential set of actions, a graph is used to model it and enforce the classification algorithm to estimate a solution with this constraint. The graph is described by a Markov chain A(m). Based on the hidden Markov model theoretical framework which by construction fits with this kind of modelling, the proposed method extends this framework to other classifiers: Bayes, LDA and SVM are discussed in this paper. These classifiers and their graph enforced versions are applied and their results compared to the analysis of the timed up & go test to recognize its different phases.

Estimation of physical activity monitored during the day-to-day life by an autonomous wearable device (SVELTE project)

Physical activity (PA) and the energy expenditure it generates (PAEE) are increasingly shown to have impacts on everybodys health (e.g. development of chronic diseases) and to be key factors in maintaining the physical autonomy of elderlies. The SVELTE project objective was to develop an autonomous actimeter, easily wearable and with several days of autonomy, which could record a subjects physical activity during his/her daily life and estimate the associated energy expenditure. A few prototypes and dedicated algorithms were developed based on laboratory experiments. The identification of physical activity patterns algorithm shows good performances (79% of correct identification), based on a trial in semi-free-living conditions. The assessment of the PAEE computation algorithm is under validation based on a clinical trial.

Improving activity recognition using temporal coherence

Assessment of daily physical activity using data from wearable sensors has recently become a prominent research area in the biomedical engineering field and a substantial application for pattern recognition. In this paper, we present an accelerometer-based activity recognition scheme on the basis of a hierarchical structured classifier. A first step consists of distinguishing static activities from dynamic ones in order to extract relevant features for each activity type. Next, a separate classifier is applied to detect more specific activities of the same type. On top of our activity recognition system, we introduce a novel approach to take into account the temporal coherence of activities. Inter-activity transition information is modeled by a directed graph Markov chain. Confidence measures in activity classes are then evaluated from conventional classifiers outputs and coupled with the graph to reinforce activity estimation. Accurate results and significant improvement of activity detection are obtained when applying our system for the recognition of 9 activities for 48 subjects.

Estimation of the symbol rate of linearly modulated sequences of symbols

The blind estimation of the symbol rate is one of the main issues for blind characterization of linearly modulated sequences. The estimator proposed in this article is based on the autocorrelation function of this kind of signal and exploits its limited frequency support. We prove theoretically that taking into account this property to modify some classical algorithms allows improving significantly their performances. We illustrate this result by numerical simulations, and we show that our estimator correctly detect the symbol rate in nearly 100% of the cases in many contexts as long as the signal to noise ratio is higher than 5dB.

Ethernet Motion-Sensor Based Alarm System for Epilepsy Monitoring

In several biomedical domains, it would be interesting to monitor subjects over night time using wearable motion sensors and trigger an alarm if a specific movement has been detected by processing the accelerometer readings. In this paper, the authors describe an innovative architecture for such an alarm system in the context of epilepsy monitoring. The main ingredients of the proposed system are wireless motion sensors, a radio-frequency transceiver linked to an Ethernet gateway and an acquisition server that incorporates real-time detection method. This motion analysis system is further integrated in the dataflow of an existing medicalized alarm system and an event is sent to healthcare professionals every time a seizure is detected by the expert system. The EPIMOUV system has been evaluated, during a 6-month period, in a specialized institution with epilepsy pharmaco-resistant residents.

Wearable impedance monitoring system for dialysis patients

This paper describes the development and the validation of a prototype wearable miniaturized impedance monitoring system for remote monitoring in home-based dialysis patients. This device is intended to assess the hydration status of dialysis patients using calf impedance measurements. The system is based on the low-power AD8302 component. The impedance calibration procedure is described together with the Cole parameter estimation and the hydric volume estimation. Results are given on a test cell to validate the design and on preliminary calf measurements showing Cole parameter variations during hemodialysis.This paper describes the development and the validation of a prototype wearable miniaturized impedance monitoring system for remote monitoring in home-based dialysis patients. This device is intended to assess the hydration status of dialysis patients using calf impedance measurements. The system is based on the low-power AD8302 component. The impedance calibration procedure is described together with the Cole parameter estimation and the hydric volume estimation. Results are given on a test cell to validate the design and on preliminary calf measurements showing Cole parameter variations during hemodialysis.

CA-213: Comparaison in silico de modèles physiologiques et de leur implémentation pour un pancréas artificiel pour le traitement du diabète de type 1 via la création d'un simulateur

Objectif Montrer que la personnalisation du modele dHovorka permet dinterpreter les trajectoires de glycemies pour un historique patient de donnees reelles. Materiels et Methodes Les algorithmes de controle selon le modele predictif (MPC) sont des modeles physiologiques compartimentaux du couple insuline/ glucose, utilises dans le developpement de la plupart des projets de boucle fermee pour le traitement du diabete de type 1 (egalement appele Pancreas Artificiel). Les 2 principaux modeles utilises sont ceux de Cobelli et al., et de Hovorka et al. Une implementation du modele dHovorka a ete faite dans un simulateur Diabeloop/CEA, et comparee avec le simulateur de lequipe de Cambridge. A partir de ce simulateur et des donnees reelles de patients (basal, bolus, repas, mesure de glucose interstitiel continue (CGM)) obtenues lors des campagnes de mises au point du pancreas artificiel Diabeloop, un algorithme de personnalisation du modele a ete developpe et qualifie sur sa capacite a interpreter et reconstruire les trajectoires de glycemies des patients reels. Resultats Les performances de la methode ont ete evaluees en calculant lerreur quadratique moyenne entre lhistorique de glycemies calculees par lalgorithme et celles mesurees par le CGM. Ces erreurs ont ete calculees sur 10 patients, sur des plages de 8 heures et des donnees durant 48 heures. Les resultats montrent que ce critere derreur quadratique moyenne est souvent acceptable (i.e. les courbes de glycemies reconstruites et obtenues par le CGM se superposent). Conclusions La personnalisation du modele dHovorka est possible a partir de donnees reelles et les experimentations menees ont permis de montrer les bonnes performances de ce modele pour decrire les evolutions de glycemies des patients. Cette approche semble donc interessante pour la mise au point dun algorithme de type MPC de regulation automatisee de la glycemie pour le sujet diabetique de type 1.