In this paper, we study health monitoring, using ambulatory sensors, where the data available are limited, and can be both unreliable and ambiguous. Hence, the need to consider a person's context: surrounding environment and previous situations. We propose studying multiple situational hypotheses, and the relations between hypotheses present and past. Such hypotheses are managed with a multi-agent system: the agents embody hypotheses on several levels of abstraction, from a general, rough scenario, down to precise states of both physiology and activity. These agents' hypotheses are evaluated and compared so that plausible hypotheses emerge. We discuss both the representation of situations, and multi-agent adaptive control mechanisms. This is a mainly theoretical approach, although these proposals are illustrated by an application on real data from a daily o ce life scenario.