logo i3s
Laboratoire I3S,

2000, route des Lucioles
Les Algorithmes - bât. Euclide B
06900 Sophia Antipolis - France
+ 33 (0)4 92 94 27 92
        Asya Metelkina my photo

Grâce à ses outils, l'homme possède un pouvoir sur la nature extérieure, dont il dépend cependant quant aux buts qu'il poursuit. (citation d'Hegel traduite de l'allemand: "An seinen Werkzeugen besitzt der Mensch die Macht über die äußerliche Natur, wenn er auch nach seinen Zwecken ihr vielmehr unterworfen ist")

Programming is easy. It’s knowing what to program that’s hard. (citation d'un professeur dans le livre de Chapman sur la programmation en Matlab)

Research fields: Statistics, optimization and machine learning

 Keywords: statistical modeling, random search methods, design of sequential experiments, point processes, survival analysis, stochastic interpolation,  regression and classification, unsupervised and adaptive learning, model selection, bayesian methods, biostatistics

Qualifications: CNU 26'2016, CNU 25'2013

Membership: SFdS, SMF, Femmes et Mathématiques, La Société Mathématique Européenne

Present: I am a postdoc researcher in I3S laboratory of UNS-CNRS in Sophia Antipolis France. My current research is supported by the University of Nice Sophia Antipolis. I work on the project "Neuromodulation: optimization of preference" together with Luc Pronzato and Joao Rendas. My work concerns an adaptive sequetial optimization of the electrical Spinal Cord Stimulation based on the patient's preferencies concerning his parasthesia sensations and on the information about the cobverage of the painful body zone. My project is a part of the big collaborative project on Neuromodulation between four members of ComUE University of Côte d'Azur: CHU of Nice (Neurosurgery team of Dr. D. Fontaine), I3S Laboratory (Designs team of L. Pronzato), LJAD (F. Grammont from Interactions team and Probability and P. Reynaud-Bouret, R. Diel, M. Lerasle from Probability and Statistics team), and IBV (Genetics of mouse brain development team of M. Studer). The ultimate goal of the main project is to attain a new level of understanding of pain mecanisms and of principles of neuromodulation thanks to the unified effort of reseachers from different domains.

2015: I worked as a CNRS post-doctoral researcher in I3S laboratory (UNS-CNRS) with Luc Pronzato on the construction of the Covariate-Adjusted Response-Adaptive clinical trials. This work was focused on the construction of sequential allocation procedures targeting the optimal allocation measure. The optimal allocation measure is defined as the optimum of a convex compromise criterion for balancing the information about the model parameters (measured by an information criterion) and the regret of not allocating the best treatment. The proposed criterion allows to choose the relative importance of the regret with respect to the information by selecting an appropriate tuning parameter. We illustrated on numerical examples that the proposed optimal allocation outperforms the ad-hoc procedures for balancing information and regret at each allocation step, proposed in literature. The key problem of the ad-hoc allocation procedures is that the target allocation and its properties are unknown, which distinguishes our approach. See preprints.
2014-2015: I worked as a CNRS post-doctoral researcher in I3S laboratory (UNS-CNRS) as a member of the project SAFEDIVE (funded by DGA). In the framework of this project, I worked on the modelling of decompression sickness accidents (bends) in human underwater scuba diving by a Cox process. My research was focused on the risk assessment understood as the prediction of the distribution of individuals risks in population of divers with confidence bounds. To perform this task, I've developed a new approach to the risk model estimation, SimKrige, based on the construction of surrogate models of individual diver's body response to dive conditions. I proposed a model selection and predictive quality assessment suitable for the specificity of the bends risk: (i) the rarity of the events and (ii) the double variability in exposure (one controlled from external conditions and one random from individual's properties).
2014: I've worked with Maureen Clerc at Athena team at INRIA during the internship (of my second Master studies). I've made an offline statistical study of the EEG signals recorded from patients during the clinical experience of use of P300 Speller (collaborative project between INRIA and CHU of Nice). P300 Speller is a brain-computer interface consisting in EEG commanded virtual keyboard dedicated to the communication in disabled patients (with Amyotrophic Lateral Sclerosis). The scoring method of letters classification used in P300 Speller is based on the fuzzy classification of sub-samples of EEG signal (corresponding to the event-related potentials) into "target" and "non-target" classes. I used the machine learning methods for unsupervised classification of those sub-samples and of adaptive learning incorporating different mixture of experts voting schemes.

Collaborations: I've collaborate with CHU of Nice, most particularly with Dr. Christian Pradier (director of DSP) and K. Tifratene, on the survival and regression modelling  for Alzheimer’s disease on the data from National Alzheimer’s Bank (BNA). This collaboration started with a short internship under the supervision of Dr. Christian Pradier in 2013 and continued after. In particular, we have modelled the conversion from amnesic and non-amnesic MCI state to Alzheimer’s disease by the Cox model in order to investigate the influence of relevant covariates. See publications.


[1] K. Tifratene, Ph. Robert, A. Metelkina, Ch. Pradier, J.F. Dartigues Progression of mild cognitive impairment to dementia due to Alzheimer’s in clinical settings (published in Neurology 2015(85))

[2] Ch. Pradier, Ch. Sakarovitch, F. Le Duff, R. Layese, A. Metelkina, S. Antony, K. Tifratene, Ph. Robert The mini Mental Score at the time of Alzheimer’s Disease diagnosis, according to age, education, gender and lifestyle: a cross-sectional study among the French population using the National Alzheimer Database (published in PLoS ONE 9(8): e103630)

[3] A. Metelkina Etude de perturbations à variations lentes des opérateurs de Schrödinger sur la droite réelle  (PhD thesis, University Paris XIII, september 2011)

[4] A. Metelkina Periodic and quasiperiodic approximations for slowly oscillating potential, (published in OWR 50/2011, p.29-31)


[1] A. Metelkina, L. Pronzato Covariate-adjusted response-adaptive compromise optimal designs (on HAL)