Ingegneria industriale e dell'informazione

She graduated in Mathematics in 1992; from 1992 to 1994 she was at the Institute of High Mathematics and in 1998 she obtained the PhD in System Engineering from University of Rome “La Sapienza”. Then she had a post doc position in Identification and Optimal Control. Since 2002 she is assistant professor in Automatic Control at the Department of Computer Control and Management Engineering Antonio Ruberti of Sapienza University of Rome.
From 1992 to 2002 she attended many courses in applied mathematics, automatic controls, biomedical signal analysis, image analysis, physics, and non-linear dynamic.
Since 2003 she is member of the Centro Interdipartimentale di Ricerca per l’Analisi dei Modelli e
dell’Informazione nei Sistemi Biomedici.
Since 2008 she is member of the Centro Ricerche Aerospaziali Sapienza.
She is member of the Editorial Board of the Journals: International Journal of Imaging
International Journal of Computer Interaction & Information Technology; moreover she is member of the Scientific Committee of many Conferences on image analysis.
Since 2012 she is associate editor of the Journal Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization
She is referee of many international scientific journal.
Since 2001 she has participated to financial programs from the Faculty and the University;
- 2005 PRIN project “Studio, progetto e realizzazione di algoritmi efficienti di classificazione mediante reti neurali artificiali di immagini di provini metallografici di ghisa sferoidale; modelli dinamici basati su reti neurali artificiali di fenomeni di frattura” (Partecipazione)
2010-2012 SARFIRE project, ASI financial program (Partecipazione)
2009-2012 PRIN project: “Characterization of fatigue damaging in ductile cast iron by image analysis and artificial neural network” (Responsabile di una Unità di Roma)
Her main research interests from methodological and applicative point of view are in 1-D and 2-D signal processing, estimation theory, system identification, optimal control, neural network, bioengineering. In particular image analysis techniques are considered in material science and in biomedical application. Optimal control theory has been studied and applied to control epidemic disease in SIR and SIRC model, bone remodeling and to determine the optimal allocation of band resources. Neural network were studied both in image processing and in material science to predict crack propagation.
Currently in collaboration with colleagues of University of L’Aquila she has been studying EEG signals and their classification for human machine interface.
TEACHING ACTIVITY
2002-2014: Automatic Control at Sapienza University of Rome, Faculty of Civil Engineering
2014-2015: Optimal Control in Master of Science in Control Engineering and Control in Biological Systems in Biomedical Engineering at Sapienza University of Rome
For more details: www.dis.uniroma1.it/~iacoviel

Progetti di Ricerca Piccoli : 262 progetti finanziati da 4.000 a 5.000 euro

In this project the human machine interface is considered to provide a multi-class classification algorithm based on EEG signals from self- induced emotions.
The aim is to provide a new channel of communication based on emotions like disgust, happiness, fear, hunger, and so on, so to yield an efficient, parametric, general, and completely automatic real time classification method of EEG signals from emotions. Preliminary and promising results were obtained considering the ancestral emotion of disgust/relax (see references below), in collaboration with a team of researchers of University of L’Aquila led by Prof. G. Placidi. The general EEG signal classification problem will be faced; the signals to be considered will be self-induced, and not induced by a video or external stimuli. Therefore there are problems related with EEG low amplitude signals.
Signal processing techniques, like wavelet transform, will be studied and applied, determining the suitable level of transformation; in general in the EEG signals the information content relies in the frequency interval [0 46]Hz and it will be interesting to study the localization of the signals for the different emotions and to determine if a more specific range of frequency could be identified in each case.
The signals are acquired by the team of University of L’Aquila. To provide a classifier, on each transformed signal suitable features must be identified; to reduce redundancy, Principal Component Analysis (PCA) could be performed. PCA features’ selection or features’ extraction should be tested.
For the classification of the features (and therefore of the signals) support vector machine (SVM) appeared to be the most promising tool even if in literature also neural networks are implemented. A comparison among different classification strategy will be implemented. The multi-class classifier will be tested over a significant number of subjects of different categories and applied to a demonstrative device.

Brain Computer Interface (BCI) or Brain Machine Interface (BMI), (Alonso 2012), represents a stimulating research area aiming at transforming thought into action, by providing a new channel of output for the brain. BCI analyzes signals generated by voluntary neural activity of the central nervous system, electroencephalographic signals (EEG), i.e. cortical surface recording, or extracellular brain recording. The subject, thinking at an intention, generates voluntary brain signals to be translated into commands for an output device, (Birbaumer 2006). The neural activity useful for BCI can be measured by electroencephalography either by external electrodes or through microelectrodes implanted inside the skull. EEG with externally placed electrodes is safe, not expensive, not invasive, and maintains high temporal resolution, (Srinivasan, 1999). Though BCI is currently applied in a lot of different fields ranging from video games to military to the study of epileptic subjects’signals (Das et al.2013), it is mostly applied as a support for disabled people. In the last ten years BCI has revealed its possibilities especially for severely disabled and locked-in individuals with very limited possibility to interact with the ambient and other subjects, (Birbaumer et al.2007). Generally, the BCI communication for disabled subjects is based on event-related signals, (Escolano et al. 2012); in (Gandhi et al. 2014) an adaptive user interface allowing the movement of a robot device used motion imagery. In (Wodlingeret al. 2015) BMI control of an anthropomorphic prosthetic arm was studied. The different techniques proposed mainly depend on the kind of signal to be analyzed and the information to be retrieved.
To this aim, useful features are extracted from the signal adequately processed, generally decomposed into different frequency bands by the Discrete Wavelet Transform. The most common features extracted from the EEG signals were recalled in (Subasi et al.2005): the entropy, the energy and the standard deviation. Sometimes the set of feature may be redundant and Principal Component Analysis (PCA) is considered for features’ extraction or features’ selection.
Once a set of feature is determined, different types of classifiers may be implemented to face with multiclass EEG signal classification; in (Poyhonen et al.2005) multilayer perceptron neural network, probabilistic neural network and multiclass support vector machines (SVM) were considered to study signals from subjects affected by epilepsy.
The SVM is an efficient method for classification of EEG signals. Generally they are applied as a two-classes classifier but can be extended for multi-class classifiers.
Recently, great attention has been devoted in the classification of emotional states by EEG-based functional connectivity patterns, (Lee et al.2014), (Rached et al. 2013). A new interesting stimulus generated by the disgust produced by remembering an unpleasant odor has been presented in (Placidi et al.2015) to be used as an alternative task to drive an EEG-based communication BCI for severely unpaired people, for which classical stimuli are ineffective.
The proposed paradigm allowed the construction of a binary BCI in the class of affective BCI, based on the measurement of the emotions. The method, based on the short time Fourier transform, was very simple and effective in classify binary data (activation, through “disgust” and not activation, through “relax”); it assumed that the disgust, being a negative emotion, was generated and located in the right hemisphere of the brain and that the emotions were normally measured in the range of gamma frequencies; a limit relied in the binary nature of the classification: the signals to be classified had just binary outcomes (YES or NO). For this reason, the described classification strategy used just signals from the right hemisphere of the brain, analyzed mainly the gamma frequencies and was specifically designed for binary outputs.
The idea of this project was inspired by an ongoing collaboration with the research team led by Prof. Placidi of University of L’Aquila. Some results are under review (see the list below). The results obtained considered EEG signals from 8 channels: P4, C4, T8, P8, P3, C3, T7,P7. The signals were those produced by the disgust auto-induced by remembering an unpleasant odor. With this specific emotion the information content in the EEG signal lies in precise frequency range. The proposed algorithm yielded a binary classifier, stimulus or absence of stimulus. The peculiarity of this signal required ad-hoc signal processing by wavelet decomposition, and the definition of a set of features able to characterize the signal and discriminate different conditions. The proposed method is a two stages algorithm; the first one is off-line and is devoted at the calibration. The second stage is the test on new data. To avoid redundancy in the set of features, PCA theory is adapted and applied; the classification of the selected features, and therefore of the signals, is obtained by the SVM. The flow-chart of the proposed method is :


References
L.Alonso, Review Brain Computer Interfaces, Sensors, 1211-1279, 2012.
N. Birbaumer, Breaking the silence: Brain–computer interfaces (BCI) for communication and
motor control, Psychophysiology, 43, 6, 517–532, 2006.
N. Birbaumer et al., “Brain-computer interfaces: Communication and restoration of movement in paralysis”, J. of Physiology, 579, 3, 621–636, 2007
P.Das et al., Epilepsy disorder detection from EEG signal, J. Intelligent Computing and Applied Sciences, 1, 1, 41-49, 2013.
C. Escolano et al., A telepresence mobile robot controlled with non invasive brain-computer- interface, IEEE Trans. on Systems, Man, and Cybernetics,-Part B, 42, 3, 793-804, 2012.
V. Gandhi et al., EEG-based mobile robot control through an adaptive brain robot interface, IEEE Trans. on Systems, Man, and Cybernetics,-Part B, 44, 9, 1278-1285, 2014
Y.Lee et al., Classifying different emotional states by means of EEG-based functional connectivity patterns, PLOS one, 9, 4, 1-13, 2014
G. Placidi et al., “Basis for the Implementation of an EEG-based Single-Trial Binary Brain Computer Interface through the Disgust Produced by Remembering Unpleasant Odors”, Neurocomputing, 2015.
S. Poyhonen, et.al., Coupling pairwise support vector machines for fault classification, Control Engineering Practice, 13, 759-769, 2005.
T.Rached et al., Emotion recognition based on brain computer interface systems, in "Brain-Computer Interface Systems - Recent Progress and Future Prospects", 2013
R.Srinivasan, Methods to improve the spatial resolution of EEG, Int. J. of Bioelectromagnetism, 1, 102–111, 1999
A. Subasi et al, “Classification of EEG signals using neural network and logistic regression”, Computer Methods and Programs in Biomedicine, 78, 87-99, 2005
B. Wodlinger, et.al, Ten-dimensional anthropomorphic arm control in a human brain−machine interface: difficulties, solutions, and limitations J. Neural Eng. 12, 1-17, 2015
Papers submitted in collaboration with the Research team of L’Aquila
D.Iacoviello, A.Petracca, M.Spezialetti, G.Placidi, A classification algorithm for BCI driven by EEG signals from self-induced emotions, submitted to Computer Methods and Programs in Biomedicine.
D.Iacoviello, A.Petracca, M.Spezialetti, G.Placidi , A classification algorithm with optimized parameters for noisy EEG signals, submitted to IEEE Trans. on Cybernetics
F. Pistoia, A.Carolei, D.Iacoviello, A.Petracca, S.Sacco, M. Sarà, M. Spezialetti, G. Placidi, EEG-detected olfactory imagery to reveal covert consciousness in minimally conscious state, submitted to J.Neurology, Neurosurgery & Psychiatry
G.Placidi, A.Petracca, M.Spezialetti, D.Iacoviello, A Modular Framework for EEG Web Based binary Brain Computer Interfaces to Recover Communication Abilities in Impaired People, submitted to Int. J. of Human-Computer Studies

The project will be taken forward in collaboration with the Research Team led by Prof. Giuseppe Placidi of University of L’Aquila. Therefore regular meetings with the colleagues of L’Aquila will be planned.
The collaboration has already started, as referred in section 2.4 leading to preliminary results about the use of one emotion, the disgust, for a binary output.
For increasing the quality and the efficiency of a BCI based on emotions classification, the cardinality of the alphabet should be increased in order to reduce the time necessary for communication. This is possible by increasing the number of emotions to be recognized (for example, by considering, besides “disgust”, also “happiness”, “fear”, “hunger”, and so on). The increment of the number of emotions implies that also signals coming from the left side of the brain have to be considered, that the whole frequency spectrum has to be analysed, and that “features” different from the ones already studied have to be considered to help in discriminating between very similar, but different, emotions. This approach could contribute to avoid confusion between different emotions and to reduce the number of classification errors.
Hence, the aim of this project is to provide a new channel of communication based on emotions like disgust, happiness, fear, hunger, and so on, to yield an efficient, parametric, general, and completely automatic real time classification method of EEG signals from emotions.
The particular choice of the auto-induced stimuli and the efficiency of the classifier proposed could yield to a simplification of the nowadays used Human Machine Interfaces and a significant cost reduction for possible implementation.
The classification procedure will be a two stages one: the off-line one, that is the calibration stage, and the on-line stage, that is the application of the classifier to a new unknown signal.
The output is a multi-class classifier to translate auto-induced emotions into actions and the realization of a demonstrator device for testing and proving the applicability and the feasibility of a simple non-invasive EEG based physical interface. The effectiveness of the approach will be validated through final experimental tests on different categories of subjects.

The project is organized in specific tasks which compose the different phases according to a development plan that considers the evolution of the research state:

Task1: Bibliographical research
Task2: Device analysis
Task3: Data acquisition
Task4: Data analysis
-Task 4.1: Signal processing
-Task 4.2: Features Computation
-Task 4.3: Principal Component Analysis (PCA)
-Task 4.4: Signal Classification
Task5: Validation of the classification algorithm
Task6: Study of signal-device interface



Fig.1.Phases and Tasks

Phases and tasks description:

PHASE 1:
Phase 1 is divided into two tasks, required for letting all the research group acquire a common and up to date background. Hence, all the researchers will be involved in this phase and in each tasks.

Task 1. Bibliographical research:
The state-of-art dealing with induced and auto-induced emotion should be thoroughly studied with respect to the different aspects of the project.
More precisely, the following problems will be deeply addressed: signal acquisition, signal processing and data analysis.
-Signal acquisition: the problem of the identification of the self –induced emotion calls for the analysis of different aspects, among which the most important refers to the localization of the information both in the spatial and in the frequency domain.
--a.Problem of spatial localization: a good knowledge of which brain area is activated in presence of self-induced emotions is greatly helpful for better understanding and proving which channels are more sensitive to such stimuli. Up to now, considering just the auto-induced ancestral emotion of “disgust” the right hemisphere and in particular the channel T8 appeared the most sensitive. Therefore, regarding to a binary decision output one could think at a simplified device, considering a limited number of electrodes. In case of a multi-class classifier the problem is to determine the spatial localization of each signal. This knowledge could help in a more efficient signal acquisition procedure and in the simplification, and a possible reduction of the costs, for any device realization for real use.
--b.Problem of frequency localization: in general the information in EEG signals relies in the range [0, 46]Hz; when dealing with the “disgust-emotion” the information is in the interval 8-12 Hz, the so called alpha band and, more significantly, in the interval 30-42 Hz into the gamma band. This knowledge could help in a more efficient signal processing. In fact the acquired signal needs a processing and the knowledge of the involved frequency interval could help when a suitable transform must be implemented. For example, when dealing with the disgust emotion the level 3 of the Meyer wavelet transform has been effectively used.
-Signal processing: once the signal is available, the problem is to determine the most suitable transformation to be applied. Generally Discrete Wavelet transform (DWT) or Continuous Wavelet Transform (CWT) are widely applied and the determination of the specific wavelet is an important issue for the effectiveness of the result. For EEG data different wavelet functions may be considered; usually, as can be found for example in the cited paper of Subasi 2007, the Meyer wavelet and the Daubechies 4 are considered. In fact, orthogonal wavelet transforms are particularly useful in analyzing brain signals for their capability in separating different frequency bands; moreover they have a significant contribution for reducing the noise effects. For these reasons, in the application on the auto-induced disgust emotion the Meyer wavelet was preferred to others; it is an orthogonal symmetric wavelet, infinitely differentiable, with infinite support.
-Data analysis: from the transformed signals useful features must be identified to be classified. The features classification procedure used in literature in similar problems will be studied in deep as the starting point for the development of a performing approach to multi class classification.

Task 2. Device analysis:
Since the project starts on the bases of the EEG signals currently acquired at the University of L’Aquila, and in view of subsequent new acquisitions according to the development of the project, the first phase will be devoted in studying the technical aspects of the device to be used and the characteristics of these kind of signals, when induced by different emotions.
Up to now the system used to record the EEG by the L’Aquila research team has been EnobioNE®ttp://www.neuroelectrics.com/enobio), an 8 channels (two more channels are used one as reference and another for ground) precise and robust wireless EEG equipment that uses a neoprene cap to fix the channels in the desired brain locations. The channels can be dynamically associated to variable positions in the international 10-20 system. The channels of the 10-20 international system we used were the following: P4, C4, T8, P8, P3, C3, T7 and P7.



Figure 2. International 10-20 positioning system. In red the channels used up to now


A dry copper electrodes (coated by a silver layer) fixed to the cap that ensured the contact with the subject’s scalp is used. A study on the new commercial acquisition system will be also performed in view of the development of a working testing device.

After the first phase of the project a deep knowledge of the state of the art about the current signal processing techniques applied to auto-induced emotions and the instrumental device will be obtained by the members of the group.

PHASE 2
All the researchers will be involved in this phase.

Task 3. Data acquisition:
The set of signals will be acquired by the Research Team of L’Aquila. In this phase the Sapienza team will be strictly in contact with the colleagues of L’Aquila, supporting them in the acquisition procedure and getting confidence with the acquisition device. During the execution of this task, a certain number of travels to L’Aquila for supporting the measurement sessions are to be considered.
Signals from different channels will be acquired considering, as a first step, the results from literature. This phase will be strictly connected with the successive one in which the signals will be processed.

PHASE 3
All the researchers will be involved in this phase, with a major participation of Iacoviello to the algorithms development and of Di Giamberardino to the data handling and the computation.

Task4. Data Analysis:
In this phase the acquired signals will be processed and the hypothesis under investigation will be verified. To better validate the results, during this phase the group could need more data related to different channels or in different conditions. Therefore in any of the steps in this phase it could be needed to acquire new data and hence the necessity of travels to L’Aquila should be taken into account.
Dealing with different emotions and aiming at a multi-class classifier, starting from the results presented in literature it will be important to use or introduce a correct and effective technique suitable for the classes of signals involved.

The ordered steps requiring continuously updates from one step to the others are:
-Task 4.1. Signal processing:
Starting from the CWT and/or DWT the different kind of wavelet transforms will be tested; the first approach will involve the Meyer wavelet and the Daubechies’. The level of the decomposition needs to be established once a frequency localization is identified, otherwise all the relevant interval of frequencies should be used. The choice of the decomposition level depends on the frequency of interest; more precisely, assuming a sampling frequency of 500 Hz (the sampling frequency of the EEG system used for the experiments), and therefore the presence of frequencies in the signal up to 250 Hz, at the l-level of a diadic decomposition the approximation CA_l contains information in the interval [0, 250/2l] Hz, whereas the detail CD_l yields information in the interval (250/2l, 250/2l-1] Hz. The diadic wavelet decomposition is used to select and retain just specific frequency bands of the original signal before to skip to the following step of the classification method.
-Task 4.2. Features computation:
From the processed signals useful characteristics able to discriminate the components corresponding to different auto-induced emotions will be determined. Some features, the mean value, the median, the standard deviation, for example, have been very commonly used but some new variables related with the specificity of the signal could be investigated.
-Task 4.3. Principal Component Analysis:
Since at this very point probably a large number of features (from literature and/or new) will be considered, without really knowing which one is important, the Principal Component Analysis (PCA) could be considered; there are two main strategy: features’ extraction and features’ selection. In the former case from the set of features a new one is obtained by suitably combining the others; in the latter the most significant ones are selected among the features of the set. During this task the most efficient strategy will be determined for the specific signals.
-Task 4.4. Signal classification:
Support vector machines (SVM) constituted a promising tool for signals classification when the emotion to be identified was the disgust and when only a binary decision set up needed to be addressed. In the contest of this project the effectiveness of SVM needs to be tested and different classification procedure will be considered and compared. For example, neural network have been applied with success to study EEG signals from epileptic signals.
The classifier we are going to implement will be constituted by two stages, an off line one, for the calibration, and an on-line one. An important aspect that will be taken into account is the computational time; the aim to provide an on line classifier, therefore the time requested for the on-line phase will be optimized providing an efficient procedure in the next phase.

PHASE 4
All the researchers will be involved in this phase, with a major contribution by Iacoviello for Task 5 and by Di Giamberardino for task 6.
This phase will be mainly devoted to the experimental validation of the technique and of the algorithm. The testing activities will be divided into two tasks:
-Task 5. Validation of the classification algorithm:
Testing the procedure over the signals acquired from different categories: women and men, different classes of ages, left and right handed, for example. These tests will be meaningful in terms of spatial localization, helping in determining the most significant channels, and in frequency localization allowing to determining the frequencies most involved by the different emotions.
-Task 6. Study of a signal device interface:
Since it is intention of the proposers to validate all the results that can be obtained at the end of the project, the development of a first simple but effective experimental device will be taken into consideration. This task is of importance since the validation of the on line applicability of the entire EEG signal classification procedure is essential for all the possible application that can come from the results.
Being the on line working device depending on both the hardware solutions and the software implementation, in this phase also the efficiency of the software will be studied and optimized.
The idea is to create an interface between a commercial low cost helmet, that can be bought with less than 1000 euro with the open source software included, with an actuator device which can perform some demonstrative tasks, like turning on or turning off one or more lights, as well as activating an electrical motor or making a bell ring.
It will be important to evidence the powerfulness of the approach proving that more than one device can be controlled at once thanks to the extension from one bit (one emotion recognition) information to a multibit (multi emotion recognition) one.