Ent?te

Logo du LIFL

Depuis le 1er janvier 2015 le LIFL et le LAGIS forment le laboratoire CRIStAL

  1. Doctoral studies

Thesis of

Julie Jacques

Monday 2 December 2013
Amphithéâtre de l'IRCICA

Classification of medical data using optimization methods applied to patient screening in clinical trials.

Directeur de Thèse : Clarisse DHAENENS, Professeur des Universités, Université Lille I Laetitia JOURDAN, Professeur des Universités, Université Lille I Rapporteurs : Jean-Charles BILLAUT, Professeur des Universités, Université de Tours Nadia BRAUNER, Professeur des Universités, Université Grenoble I Membres : Stéphane BONNEVAY, Maître de conférences, HDR, Université Lyon I Denis BOUYSSOU, Directeur de recherche, Université Paris Dauphine Sophie TISON, Professeur des Universités, Université Lille I David DELERUE (Invité), Gérant, Société Alicante

Medical data suffer from uncertainty and a lack of uniformisation, making them hard to use in medical software, especially for patient screening in clinical trials. In this PhD work, we propose to deal with these problems using supervised classification methods. We will focus on 3 properties of these data : imbalance, uncertainty and volumetry. We propose the MOCA-I algorithm to cope with this partial classification combinatorial problem, that uses a multi-objective local search algorithm. After having confirmed the benefits of multiobjectivization in this context, we calibrate MOCA-I and compare it to the best algorithms of the literature, on both real data sets and imbalanced data sets from literature. MOCA-I generates rule sets that are statistically better than models obtained by the best algorithmes of the literature. Moreover, the models generated by MOCA-I are between 2 to 6 times shorter. Regarding balanced data, we propose the MOCA algorithm, statistically equivalent to best algorithms of literature. Then, we analyze both theoretically and experimentally the behaviors of MOCA and MOCA-I depending on imbalance. In order to help the decision maker to choose a solution and reduce over-fitting, we propose and evaluate different methods to handle all the Pareto solutions generated by MOCA-I. Finally, we show how this work can be integrated into a software application.

Ours

UMR 8022 - Laboratoire d'Informatique Fondamentale de Lille - Copyright © 2012 Sophie TISON - Crédits & Mentions légales

Page respectant XHTML et CSS.

Pour tout commentaire / Comments and remarks : webmaster