Machine Learning techniques have been recently applied in the healthcare eld and particularly for electroencephalographic signal classi cation, opening new possibilities for brain activities and diseases analysis through peculiar applications like the Brain Computer Interfaces. The project proposal for the Ph.D. thesis work brie y described in the following wants to address the problems arising from these biomedical heterogeneous data, starting from the preliminary signal processing for noise removal, moving to possible data normalisation for subject and population based analysis and exploiting the outputted manipulated data to create classi ers for peculiar brain activities labelling, diseases identi cation, Brain Computer Interface development. These steps will require an evaluation of the state-of-the-art, which present mostly semi-automatic or manual signal processing techniques, that will be used to create fully automated denoising modules for every type of data and integrated for scenario-dependent signal reconstruction procedures. Also, there is a narrow number of studies addressing the normalisation problem, which is to be considered for population-based analysis. Finally, the recent works on electrophysiological signal classi cation will be used to evaluate commonly used Machine Learning algorithms and to create best-practices for feature extraction, a benchmark for deep learning techniques application and the study of Brain Computer Interface mainly for rehabilitation purposes.
In the last decades the constant technological improvement and the availability
of a greater amount of data have led to an increasing interest over the
application of Machine Learning (ML) techniques in the biomedical eld, posing new
challenges for the development of faster and more accurate classi cation
algorithms.
The project proposal for the Ph.D. thesis work described in the following, wants
to address the problems arising from the heterogeneity of a peculiar kind of
biomedical data, i.e. the Electroencephalographic (EEG) signals.
Recently, EEG has begun to be extensively used in the medical and research elds
due to its characteristics: it is non-invasive, records the cerebral bio-electric
potentials through multiple sensors (electrodes) placed on the scalp [
The EEG has been used, for example, in face recognition experiments [
However, the development of these kinds of applications encounters some di
culties due to the fact that the EEG signal is weak, time varying and easily a ected
by biological (ocular, muscular, cardiac movements) and non-physiological
(direct current, electrical leakage) noises [
Also, to make assumptions over a population, the speci city of each recording must be considered and so the di culty of normalising the data, i.e. trying to t a speci c recording into a canonical space, arises. This necessity, to which an unique solution is yet to be found, may allow researchers to move from a classical subject-based analysis to a population-based one. 2
Starting from an experiment of face recognition conducted in collaboration with
professor Daini1 [
Afterwards, the obtained features will be used to train a classi er to develop a tool for semi- to completely-automated noise removal, which will be expected to run with slight variations in any kind of case scenario.
Finally, the Ph.D. work will consider the problem of normalisation to allow analysis inter-subject or between di erent populations (e.g. access the di erences in the brain activations between a normal recogniser population and an impaired one). This last step will be introduced as an addition to the de ned pipeline, lacking of state-of-the-art references and having the heterogeneity and subjectspeci city of the EEG data as di cult issues to address. 1 Department of Psychology, University of Milano - Bicocca, Milan, Italy
As well depicted by Uriguen et al. [
As stated in the introduction, the di culties of developing an automated procedure arise from the speci c nature of the EEG signal: it varies from patient to patient, depends on the recording hardware, may present a mixture of the actual electrophysiological data and contaminated components.
Denoising techniques are mostly scenarios-dependent, but despite that, some
recent studies con rm the success of methods used in the identi cation of noisy
components and suggests some useful tips for EEG automatic signal
manipulation. For example, Al-Qazzaz et al. [
Therefore, starting from the study of semi-automatic procedures, the noise
removal methodology suggested for the Ph.D. work wants to obtain a modular set
of procedures, that could be algorithms applicable to any kind of experiment
and to scenario-dependent ones and move to the identi cation of noisy patterns
through some metrics, which could be used for the training of peculiar classi ers.
In fact, the state-of-the-art presents a good amount of ML techniques for EEG
signal classi cation. The most used are k-Nearest Neighbour (k-NN), which
assigns to a tested sample the label of the k-nearest training sample, and Support
Vector Machine (SVM), that segregates the data through an hyper plan with
maximal margins, as supervised classi ers and Nave Bayes (NB), which is based
on Bayes' theorem and determines the class of earlier probabilities through a
maximum probability algorithm and uses a feature probability distribution from
a training set [
To allow a classi cation that could be executed with both un- and supervised
approaches, recently the research community involved in healthcare began to
explore and develop applications based on deep learning techniques.
In this regard, the review edited by Miotto et al. [
Therefore, the Ph.D. project wants to (1) start from the supervised classi ers
and create best practices for feature extraction in scenario-based experiment
and for noise patterns classi cation, where the most used features would be
extracted computing the Power Spectral Density (PSD) over the frequency bands
that characterise the EEG signal (e.g. average spectral power, spectral power for
each frequency and approximate entropy) and (2) create a benchmark through
which evaluate the possibility of developing a deep learning classi er.
The last cited item opens new issues, like the fact that a deep learning model
requires a great amount of data, which should be clean and well-structured. This
could be achieved by manipulating the raw state of the EEG signal to be more
clean and interpretable through the signal processing procedure previously cited.
Also, there are no robust and well-maintained deep learning procedures on EEG
applications, even though there have been recent studies for the use of
Convolutional Neural Networks (ConvNets). Schirrmeister et al. [
Finally, from the necessity of evaluating brain activities and functions in nonand pathological conditions or between di erent states, comes the issue regarding a proper way to normalise the EEG signal, which is not only characterised by a high dimensionality, but also - as for other biomedical data - presents heterogeneity, temporal dependency, sparsity and irregularity. This problem has been discussed between numerous researchers (mainly on ResearchGate2), but an univocal solution has yet to be found.
The advised approaches emerging from the discussions and, only minimally, in
the state-of-the-art are mainly (1) the normalisation of the power spectra for
each frequency band and sensor [
The Ph.D. project wants to evaluate, as an additional step, the suggested solutions and nd better methods for data normalisation, moving from a subjectbased approach to a population-based one. 4
The project for the Ph.D. work is divided in three main steps: signal processing, classi cation of heterogeneous EEG data and normalisation.
Each of them could be divided in and expanded with di erent sub-modules: novel
algorithms for denoising based on peculiar signal characteristics, normalisation
procedures less commonly cited as the min-max normalisation [
The cleared signal and noisy patterns classi cation will be validated by experts
of the Department of Psychology at University of Milano - Bicocca, evaluating
2 https://www.researchgate.net/post/Normalization_of_resting_EEG_data_
for_comparisons_between_different_subjects
the accuracy, sensitivity and speci city of the obtained results [