Determining the stress level of a subject in real time could be of special interest in certain professional activities to allow the monitoring of soldiers, pilots, emergency personnel and other professionals responsible for human lives. Assessment of current mental fitness for executing a task at hand might avoid unnecessary risks. To obtain this knowledge, two physiological measurements were recorded in this work using customized non-invasive wearable instrumentation that measures electrocardiogram (ECG) and thoracic electrical bioimpedance (TEB) signals. The relevant information from each measurement is extracted via evaluation of a reduced set of selected features. These features are primarily obtained from filtered and processed versions of the raw time measurements with calculations of certain statistical and descriptive parameters. Selection of the reduced set of features was performed using genetic algorithms, thus constraining the computational cost of the real-time implementation. Different classification approaches have been studied, but neural networks were chosen for this investigation because they represent a good tradeoff between the intelligence of the solution and computational complexity. Three different application scenarios were considered. In the first scenario, the proposed system is capable of distinguishing among different types of activity with a 21.2% probability error, for activities coded as neutral, emotional, mental and physical. In the second scenario, the proposed solution distinguishes among the three different emotional states of neutral, sadness and disgust, with a probability error of 4.8%. In the third scenario, the system is able to distinguish between low mental load and mental overload with a probability error of 32.3%. The computational cost was calculated, and the solution was implemented in commercially available Android-based smartphones. The results indicate that execution of such a monitoring solution is negligible compared to the nominal computational load of current smartphones.

Activity and emotion recognition based on physiological signal processing in health care applications is a relevant research field, with promising future and relevant applications, such as health at work or preventive care. This paper carries out a deep analysis of features proposed to extract information from the electrocardiogram, thoracic electrical bioimpedance, and electrodermal activity signals. The activities analyzed are: neutral, emotional, mental and physical. A total number of 533 features are tested for activity recognition, performing a comprehensive study taking into consideration the prediction accuracy, feature calculation, window length, and type of classifier. Feature selection to know the most relevant features from the complete set is implemented using a genetic algorithm, with a different number of features. This study has allowed us to determine the best number of features to obtain a good error probability avoiding over-fitting, and the best subset of features among those proposed in the literature. The lowest error probability that is obtained is 22.2%, with 40 features, a least squares error classifier, and 40 s window length.

Assessing emotional state is an emerging application field boosting research activities on the topic of analysis of non-invasive biosignals to find effective markers to accurately determine the emotional state in real-time. Nowadays using wearable sensors, electrocardiogram and thoracic impedance measurements can be recorded, facilitating analyzing cardiac and respiratory functions directly and autonomic nervous system function indirectly. Such analysis allows distinguishing between different emotional states: neutral, sadness, and disgust. This work was specifically focused on the proposal of a k-fold approach for selecting features while training the classifier that reduces the loss of generalization. The performance of the proposed algorithm used as the selection criterion was compared to the commonly used standard error function. The proposed k-fold approach outperforms the conventional method with 4% hit success rate improvement, reaching an accuracy near to 78%. Moreover, the proposed selection criterion method allows the classifier to produce the best performance using a lower number of features at lower computational cost. A reduced number of features reduces the risk of overfitting while a lower computational cost contributes to implementing real-time systems using wearable electronics.

This paper presents the Real-Life Indoor Sound Event Dataset (ReaLISED), a new database which has been developed to contribute to the scientific advance by providing a large amount of real labeled indoor audio event recordings. They offer the scientific community the possibility of testing Sound Event Classification (SEC) algorithms. The full set is made up of 2479 sound clips of 18 different events, which were recorded following a precise recording process described along the proposal. This, together with a described way of testing the similarity of new audio, makes the dataset scalable and opens up the door to its future growth, if desired by the researchers. The full set presents a good balance in terms of the number of recordings of each type of event, which is a desirable characteristic of any dataset. Conversely, the main limitation of the provided data is that all the audio is recorded in indoor environments, which was the aim behind this development. To test the quality of the dataset, both the intraclass and the interclass similarities were evaluated. The first has been studied through the calculation of the intraclass Pearson correlation coefficient and further discard of redundant audio, while the second one has been evaluated with the creation, training and testing of different classifiers: linear and quadratic discriminants, k-Nearest Neighbors (kNN), Support Vector Machines (SVM), Multilayer Perceptron (MLP), and Deep Neural Networks (DNN). Firstly, experiments were carried out over the entire dataset, and later over three different groups (impulsive sounds, non-impulsive sounds, and appliances) composed of six classes according to the results from the entire dataset. This clustering shows the usefulness of following a two-step classification process.

Emotional state recognition through speech is being a very interesting research topic nowadays.\ud Using subliminal information of speech, it is possible to recognize the emotional state of the\ud person. One of the main problems in the design of automatic emotion recognition systems is the\ud small number of available patterns. This fact makes the learning process more difficult, due to\ud the generalization problems that arise under these conditions.\ud In this work we propose a solution to this problem consisting in enlarging the training set\ud through the creation the new virtual patterns. In the case of emotional speech, most of the\ud emotional information is included in speed and pitch variations. So, a change in the average\ud pitch that does not modify neither the speed nor the pitch variations does not affect the\ud expressed emotion. Thus, we use this prior information in order to create new patterns applying\ud a pitch shift modification in the feature extraction process of the classification system. For this\ud purpose, we propose a frequency scaling modification of the Mel Frequency Cepstral\ud Coefficients, used to classify the emotion. This proposed process allows us to synthetically\ud increase the number of available patterns in thetraining set, thus increasing the generalization\ud capability of the system and reducing the test error