From
Timetable to
Place Campus Saint-Germain-des-Prés, U. Paris Cité
Ioannis Bargiotas and Charles Truong from the Centre Borelli will present their work at the Borelli@Saints-Pères seminar.
Speaker 1: Ioannis Bargiotas
Titre: Prediction of falls in individuals without a history of falls and longitudinal follow-up of postural control in seniors
Abstract: Background: Postural control is an individual's capacity to maintain a controlled, upright position. The loss of postural control and, eventually falling, is one of the significant causes of injury among the elderly that promotes decreased mobility, loss of autonomy in daily activities (bathing, cooking, etc.), institutionalization, or even death. However, the limited ability of the functional clinical tests to provide quantified/objective values about the future risk of falling (fRoF), makes any medium/long-term longitudinal analysis practically impossible and, therefore, limits the capacity to predict personalized fRoF. A common way to evaluate posture is by recording the center-of-pressure displacement (stabilograms) with force platforms. The study's objectives were to create universal and diseased-specific models of prospective prediction of fall and fRoF evaluation considering his/her past assessments and constantly monitor the individual's state compared to high fRoF individuals.
Methods: We included 1537 individuals, with up to 11 posturographic follow-up assessments (irregular time intervals), without excluding any neurological or other impairment. The resulting stabilograms, were analyzed, and 184 unique fRoF-related features were calculated. We proposed a machine learning framework for 1) managing the follow-up dataset, 2)prospectively predicting the individual's fRoF, 3)monitoring the evolution of the individual's balance control compared to fallers and non-fallers from his/her particular group (if any disease) and the general population.
Results/Conclusions: The results were very promising and indicate that such a framework might provide supplementary information about the fRoF of an individual and be of significant usefulness in screening the pre-fragile population about balance deficits.
Speaker 2: Charles Truong
Title: Supervised calibration of change-point detection methods
Abstract: In numerous applications involving time series, change-point detection is a common step of the data processing pipeline. A significant difficulty of change detection methods is the calibration. Finding an appropriate set of calibration parameters often requires a laborious manual trial-and-error procedure and expert knowledge in both the application domain (e.g. biostatistics) and change-point methods. This work proposes an automatic method to calibrate change-point detection algorithms for high-dimensional time series that contain mean-shifts. Our procedure builds on the ability of an expert (e.g. a bio-medical researcher) to produce target segmentations on a few training signals. Inspired by large-margin structured learning procedures, the problem is formulated as a convex optimization problem, and recent algorithmic advances in change-detection lead to efficient resolution. Our approach learns a weight vector that can be applied component-wise on out-of-sample signals. Thanks to sparsity-inducing regularization, dimensions that are relevant (i.e. contain a change-point) are selected while noisy components are discarded. Experiments for several real-world scenarios show that supervision markedly ameliorate detection accuracy.