Nowadays, hypertension diagnosis often uses data measured by ambulatory blood-pressure monitoring (ABPM). Together with blood pressure automatic measured by ABPM, patients have to fill their preceding behavior and current status into a card, though such behavior recording is not efficient due to the strain caused by repeated handwriting, and recorded information lack of objectivity because it is self-reported data. We have been studying a method of behavior classification from sensor data, aiming at automatic behavior recording for ABPM. We selected a set of actions useful for ABPM based on physicians' interviews results, and for each action, we used wearable sensors to measure the acceleration and the angular velocity of both the trunk and the right wrist on volunteer participants. We propose an original behavior classification method in two steps. The first step consists in classifying the data into three groups of actions regarding their influence on blood pressure variation, that are the motion-related actions group, the work-related actions group, and the stillness group. The second step consists in action detection after features selection using the correlation ratio. As a result, the proposed method achieved an average action groups correct recognition rate of more than 95%, and proved to be an effective way to identify objectively the main causes of blood pressure variations.