Neurostimulation is an emerging treatment for drug-resistant epilepsies when surgery is contraindicated. Recent clinical results demonstrate significant seizure frequency reduction in epileptic patients, however the mechanisms underlying this therapeutic effect are largely unknown. This study aimed at gaining insights into local direct current stimulation (LDCS) effects on hyperexcitable tissue, by i) analyzing the impact of electrical currents locally applied on epileptogenic brain regions, and ii) characterizing currents achieving an "anti-epileptic" effect (excitability reduction). First, a neural mass model of hippocampal circuits was extended to accurately reproduce the features of hippocampal paroxysmal discharges (HPD) observed in a mouse model of epilepsy. Second, model predictions regarding current intensity and stimulation polarity were confronted to in vivo mice recordings during LDCS (n = 8). The neural mass model was able to generate realistic hippocampal discharges. Simulation of LDCS in the model pointed at a significant decrease of simulated HPD (in duration and occurrence rate, not in amplitude) for cathodal stimulation, which was successfully verified experimentally in epileptic mice. Despite the simplicity of our stimulation protocol, these results contribute to a better understanding of clinical benefits observed in epileptic patients with implanted neurostimulators. Our results also provide further support for model-guided design of neuromodulation therapy.

Stroke affects the mobility, hence the quality of life of people victim of this cerebrovascular disease. Part of research has been focusing on the development of exoskeletons bringing support to help regaining independence in daily life. One example is Xosoft, a soft modular exoskeleton currently being developed in the framework of the European project of the same name. On top of its assistive properties, the soft exoskeleton will provide therapeutic feedback via the analysis of kinematic data stemming from inertial sensors mounted on the exoskeleton. Prior to these analyses however, the activities performed by the user must be known in order to have sufficient behavioral context to interpret the data. Four activity recognition chains, based on machine learning algorithm, were implemented to automatically identify the nature of the activities performed by the user. To be consistent with the application they are being used for (i.e. wearable exoskeleton), focus was made on reducing energy consumption by configuration minimization and bringing robustness to these algorithms. In this study, movement sensor data was collected from eleven stroke survivors while performing daily-life activities. From this data, we evaluated the influence of sensor reduction and position on the performances of the four algorithms. Moreover, we evaluated their resistance to sensor failures. Results show that in all four activity recognition chains, and for each patient, reduction of sensors is possible until a certain limit beyond which the position on the body has to be carefully chosen in order to maintain the same performance results. In particular, the study shows the benefits of avoiding lower legs and foot locations as well as the sensors positioned on the affected side of the stroke patient. It also shows that robustness can be brought to the activity recognition chain when the data stemming from the different sensors are fused at the very end of the classification process.