PhD Alumni


Dal Seno Bernardo

Present position: Temporary Researcher
http://home.dei.polimi.it/dalseno/
 

Thesis title:  Toward An Integrated P300- And ErrP-Based Brain-Computer Interface
Advisor:  Matteo Matteucci
Research area:  Artificial intelligence, robotics, and computer vision
Thesis abstract:  
This thesis presents a work in the brain-computer interface (BCI) field that makes use of machine learning and statistics techniques. A BCI is a device that bypasses any muscle or nerve mediation, and it interfaces directly with the brain by measuring signals generated by its activity. Potentially, it could be helpful for people that cannot use “standard” interfaces, such as people affected by a disease that impairs limb movements, or people engaged in physical activities. Machine learning is a broad field of artificial intelligence that deals with techniques used to endow a machine (a computer) with the ability to adapt its behavior to different conditions.
The work focuses mainly on BCIs based on potentials discernible in EEG (electroencephalography) recordings, such as P300 and error potentials. A P300 occurs when a subject detects an occasional target (oddball) stimulus in a regular train of standard stimuli. An error potential can be seen when a subject makes a mistake, and, more relevant to BCI applications, when the machine the subject is interacting with does not behave as the user expects. In a P300-based BCI the user is presented with some possible choices, and they are highlighted one at a time. A user directs his attention to one choice, and thus every highlighting of that choice elicits a P300 potential.
After an overview of some methods and protocols already used for BCIs, a couple of novel algorithms to recognize relevant potentials in the brain are presented. In particular, a genetic algorithm — an optimization method that mimics the way natural evolution works — is developed to extract features from EEG data. A mathematical interpretation of the features found by the genetic algorithm is derived, which can be used to identify the most important time intervals in EEG recordings.
The algorithms are tested on real EEG data. The results are compared with a few methods from other researchers which have been replicated. Data from persons affected by amyotrophic lateral sclerosis are also tested, and for some subjects classification accuracy is well above 80% in single-sweep mode.
The main testbed for the new algorithms consists in two classical applications: a P300 speller and a P300-driven robotic wheelchair. A P300 speller is a program to write text by selecting one letter at a time, where each selection is made with the recognition of a P300 potential. The robotic wheelchair has been developed for another project and is autonomous: It can receive high-level commands from the user, like “go to the room X”. An interface driven by a P300-based BCI is set up to give such commands to the wheelchair. In both applications, error potentials are used to identify mistakes made by the P300 BCI.
Results on offline experiments show that the genetic algorithm performs well and can be effective for a real BCI, and that error potentials are present and detectable in a P300 BCI. Online experiments are done both on the P300 speller and the wheelchair, and they confirm the offline results.
An analysis of the performance gain achievable by using error potentials is also presented, based on the novel concept of utility. After a critique of the methods used in the literature, where in particular some limits of the “information transfer rate” are showed, utility is presented as a formalization of the benefit that the user receives from the behavior of the BCI. A formula for the performance of a speller is derived, and a discussion of when error-potential detection can be helpful is given. Under some simplifying assumptions, a precise characterization of the performance gain (or loss) is computed as a function of the speller accuracy. Utility can be also extended to different kinds of BCIs and used to choose many project parameters. As a proof of the versatility of utility, a more complex interface is analyzed. Utility predictions are validated through Monte Carlo simulations.