A Self-Organized Learning Model for Anomalies Detection: Application to Elderly People

In a context of a rapidly growing population of elderly people, this paper introduces a novel method for behavioural anomaly detection relying on a self-organized learning process. This method first models the Circadian Activity Rhythm of a set of sensors and compares it to a nominal profile to determine variations in patients' activities. The anomalies are detected by a multi-agent system as a linear relation of those variations, weighted by influence parameters. The problem of adaptation to a particular patient then becomes the problem of learning the adequate influence parameters. Those influence parameters are self-adjusted, using feedback provided at any time by the medical staff. This approach is evaluated on a synthetic environment and results show both the capacity to effectively learn influence parameters and the resilience of this system to parameter size. Details on the ongoing real-world experimentation are provided.

I. Introduction

Clinical observations have proven that human's biological functions (such as temperature, weight or arterial pressure) follow periodical variations regulated by the internal biological rhythm [1]. Human's daily activities also have periodical rhythms due to biological imperatives (sleeping, eating, drinking …), environmental conditions (days and nights cycles, season cycles), and social components (agenda, education, culture, sports …) [2]. Human's daily activities and biological rhythms are thus intrinsically correlated, and monitoring the Circadian Activity Rhythm¹

A Circadian Activity Rhythm is a daily rhythmic activity cycle, based on 24-hour intervals, that is exhibited by many biological organisms such as humans

of an individual provides useful information that may be used in a medical follow-up. This biological tendency for humans to have some regularities in their everyday life enables to model everyday life behavioural rhythms in order to study deviation and anomalies [3] [2].