En 2015, la chute de la personne âgée reste toujours un événement majeur, quel que soit l’angle de vue considéré. Elle est toujours associée à une forte morbi-mortalité, nombreuses incapacités, altération la qualité de vie du chuteur, mais aussi, en raison du vieillissement de la population, avec le nombre croissant de chuteurs requérant une prise en charge médicale. Cette situation repose en bonne partie sur notre incapacité à identifier la personne âgée qui est le plus à risque de chute, cette étape étant la première de toute stratégie d’intervention efficace et efficiente. Il est donc nécessaire voir obligatoire aujourd’hui de redoubler nos efforts sur l’amélioration de la prédiction de la chute. En contrepartie de nouvelles opportunités s’ouvrent à nous en raison de l’implantation et de l’informatisation des données médicales. La chute doit être considérée comme un événement chaotique et sa prédiction doit se faire via de nouveaux modèles mathématiques intégrant la particularité de ce comportement. C’est pour cette raison que des méthodes d’analyse basée sur l'intelligence artificielle semblent être une solution appropriée. C’est à partir de ce constat que nous avons émis l’hypothèse que les modèles mathématiques issus de l’intelligence artificielle devaient permettre d’atteindre une qualité de la prédiction meilleure. L’objectif principal de cette thèse est d’étudier la qualité de la prédiction de la chute, récurrente ou non, chez des personnes âgées de 65 ans et plus, en utilisant les réseaux neuronaux et un modèle de logique floue, en les comparant avec des modèles mathématiques linéaires utilisés classiquement dans la littérature. L’ensemble de nos résultats confirme notre hypothèse de départ en montrant que le choix du modèle mathématique influence la qualité de la prédiction de la chute, les modèles non linéaires, et notamment les réseaux neuronaux et les systèmes de logique flous, étant plus performants que les modèles linéaires pour la prédiction des chutes surtout lorsqu’elles sont récurrentes. / Falls in the elderly are still a major issue in 2015 because they are associated with high rate of morbidity, mortality and disability, which affect the quality of life. From the patient’s perspective, it is still associated with high morbidity, mortality and disability, which affect the quality of life. The number of fallers requiring medical and/or social care is growing up due to aging population. This fact seems paradoxical since during the recent years the knowledge about the mechanisms of falls and the quality of interventions to support fallers significantly increased. This is largely based on our inability to predict correctly the risk of falling among the elderly person, knowing that this is the first step of any efficient and effective intervention strategies. Therefore it is necessary today to double our efforts in improving the prediction of falls. Nonetheless, new opportunities and advanced technologies provide to us the possibility of computerizing of medical data and research, and also to improve prediction of falls using new approaches. A fall should be considered as a chaotic event, and its prediction should be done via new mathematical models incorporating the feature of this behaviour. Thus, the methods ofartificial intelligence-based analysis seem to be an appropriate solution to analyse complex medical data. These artificial intelligence techniques have been already used in many medical areas, but rarely in the field of fall prediction. Artificial neural networks are the most commonly used methods while other promising techniques based on fuzzy logic are less often applied.Based on this observation we have formulated the hypothesis that non-linear mathematical models using artificial intelligence are the models, which are the most likely to achieve the bestquality of the prediction. The main objective of this thesis is to study the quality of theprediction of falls, recurrent or not, among the adults aged 65 years and more,applying neuralnetworks and fuzzy logic models, and comparing them either among themselves or with the linear mathematical models conventionally employed in the literature for fall prediction. The first cross-sectional study was conducted by using a decision tree to explore the risk of recurrent falls in various combinations of fall risk factors compared to a logistic regression model. The second study was designed to examine the efficiency of artificial neural networks (Multilayer Perceptron and Neuroevolution of Augmenting Topologies) to classify recurrent and nonrecurrent fallers by using a set of clinical characteristics corresponding to risk factors measured among seniors living in the community. Finally, in the third study we compared the results of different statistical methods (linear and nonlinear) in order to identify the risk of falls using 7 clinical variables, separating the collection mode (retrospective and prospective) of the fall and its recurrence. The results confirm our hypothesis showing that the choice of the mathematical model affects the quality of fall prediction. Nonlinear models, such as neural networks and fuzzy logic systems, are more efficient than linear models for the prediction of falls especially for recurrent falls. However, the results show that the balance between different criteria used to judge the quality of the forecast (sensitivity, specificity, positive and negative predictive value, area under the curve, positive and negative likelihood ratio, and accuracy) has not been always correct, emphasizing the need to continue the development of the models whose intelligence should specifically predict the fall.