**1. Introduction**

Activity can be defined as the state or quality of being active, which implies that the activity can be emotional, intellectual, physical, etc. Typical activity recognition systems focus on daily life activities such as walking, running, exercising, scrubbing and cooking [1–5], mental tasks [6,7] or emotion recognition [8]. Activity-state recognition systems can be applied to human error prevention tasks in many professional activities such as first responders, crane operators or train drivers. The present work aims at deeply studying several features found in the literature to characterize the signals of electrocardiogram, thoracic bioimpedance and electrodermal activity, whose objective is to recognize four different activities: emotional, mental, physical and neutral activity (resting).

Currently, there are different methods for detecting activity. For instance, Inertial Measurement Units (IMUs) [9,10] in combination with Global Positioning System (GPS) data for outdoor applications [11] or sensor located indoors for smart homes [3,12] for detecting physical activity. On the other hand, speech and gestures can be useful for assessing emotional activity [13–16]. Another alternative is physiological signals captured through sensors located in the body of the subject. Wearable biomedical sensing through smart clothing [17,18] allows the recording of physiological

measurements such as the Electrocardiogram (ECG), the Thoracic Electrical Bioimpedance (TEB) or the Electrodermal Activity (EDA), among others, which contain not only information about specific body functions and physiological states, but also valuable information about the activity and the person's condition regarding emotional state, mental load and physical activity [19].

In the literature, numerous works are found in which these three signals are used to detect stress, emotions, and activity. For instance, ECG is affected by these factors, since the heart rate is directly related to the body and mind condition [20–22]. In this sense, the Heart Rate Variability (HRV) has been widely used to extract information about the status of the autonomous nervous system and emotions [23]. On the other hand, TEB can be used as an indicator of the breathing function, and it has been used in different studies for activity recognition [24] and stress detection [25]. EDA measures the activity of sweating glands on the skin which are directly controlled by the sympathetic nervous system, and thus can also be used for emotion recognition [26–29].

However, few papers provide deep studies including all these three signals with the same purpose, comparing the physiological signals under study and determining which physiological signal provides more relevant information about the individual activity. For instance, the features extracted from TEB signal acquired together with the ECG and the heart sound can be used to study cardiovascular reactivity during emotional activation in men and women [24]. Numerous features have been found for this purpose in the literature, but there is not a clear rule of which ones are more relevant for a given problem. In general, the larger the number of features, the greater the generalization problems, that is, the ability to handle unseen data [30]. Selecting a subset of features results mandatory for many activity recognition application.

Taking all this into account, the present paper aims at assessing the utility of features extracted from ECG, TEB, and EDA in activity recognition systems. These physiological signals have been recorded using sensorized garments combined with wearable instrumentation. We intend to recognize four different activities: emotional activity, mental activity, physical activity, and resting. The paper is structured as follows: Section 1 introduces the problem tackled in this paper; Section 2 is a review of the literature about physiological sensing, window length, features, and possible classifiers; Section 3 summarizes the sensors used to acquire the signals and the mental activity states that are considered; Section 4 presents the experiments carried out; Section 5 includes the obtained results; Section 6 presents the main conclusions. A set of Appendixes A–C are also included with a detailed description of the considered features extracted from the different acquired signals.
