A SCORE BASED METHOD FOR P300 COLLABORATIVE BCI L. Bianchi 1 , F. Gambardella 1 , C. Liti 1 , V. Piccialli 1 1 Department of Civil Engineering and Computer Science, University of Rome Tor Vergata, Rome, Italy E-mail: chiara.liti@uniroma2.it ABSTRACT: Group decision-making is the process where two or more people are engaged in generating a so- lution for a given problem. In the last decade, researchers started exploiting collaborative Brain-Computer Inter- faces to enhance group performance. Various meth- ods have been proposed to integrate EEG data of mul- tiple users showing the improvement in group decision- making over single-user BCIs and non-BCI systems. In this study, we investigate four EEG integration strategies: EEG averaging across participants, the standard majority voting rule and two weighted voting system. For each ap- proach, we evaluate three different scenarios varying the number of iterations necessary to perform a single selec- tion. In all cases, it is possible to exceed 90% of accuracy with at least one collaborative BCI. INTRODUCTION Group (or Collective, or Collaborative) decision-making is the process where two or more people are engaged in generating a solution for a given problem [1]. Combina- tion of sensing and cognition capabilities allow a group to make better decisions than single individuals [2]. Nev- ertheless, group decisions can be negatively affected by several factors, such as lacking of time, sharing of infor- mation, group and leadership style and communication biases [2, 3]. In the last decade, researchers started to use Brain-Computer Interfaces (BCIs) to enhance group decision-making. BCIs allow people to interact with the environment without requiring any peripheral muscle ac- tivity to complete the interaction [4]. Brain signals are acquired and processed by a computer to identify a partic- ular type of neural process called event-related potential (ERP), which is the brain response resulting after spe- cific sensory or cognitive events. ERP-based BCIs use an oddball paradigm to elicit the ERP components: the user has to focus on ’target’ (rare) stimuli which are inserted in a stream of ’non-target’ (frequent) stimuli. As target and non-target stimuli elicit different responses, they can be distinguished and exploited by the BCI. Single-user BCIs are widely exploited for clinical purposes, most of them aiming at restoring communication capabilities to severely disable people [5]. Instead, a collaborative Brain-Computer Interface (cBCI) is a system designed for integrating brain signals from a group of users for im- proving a decision-making process. In the last decade, various approaches to integrate EEG signals have been proposed. For example, single-trial ERPs can be averaged across group members and then processed as a single-user BCI. Alternatively, neural fea- tures can be inferred from the EEG data of each user and concatenated afterwards to build a feature vector for the group, which is then passed to a single classifier. Finally, the output of several single-user BCI can be combined by means of a voting system to compute the group deci- sion [6]. In [7], these approaches have been applied to the EEG data collected from 20 subjects in a movement- planning experiment. In the voting method, a SVM clas- sifier was trained for each subject. The classification out- put was then weighted according to each user’s training accuracy. All the three cBCIs outperformed the single- user BCIs. Moreover, the voting strategy turned out to be the optimal method for collaborative EEG-based classifi- cation. In [8] and [9] the same strategy has been applied to a visual target detection task and a visual Go/NoGo task, respectively. The output of the single-user SVMs has been used as the input for a second-layer SVM. In [10], the authors integrated the EEG of 20 individuals en- gaged in the discrimination among pictures of cars and faces, using various voting decision rules for combining information across user. The advantages of a cBCI has been evaluated also in [11], where data recorded using a P300 speller paradigm have been analyzed showing that combining data from users led to an improved accuracy with respect to fusing data from the same participant over time. In [3], a completely different weighted majority rule has been introduced. The authors developed a hybrid cBCI that does not predict the user decision but combines neural signals and response times to determine the deci- sion confidence of each user and then weights their be- havioral responses accordingly to produce the group de- cision. This hybrid cBCI was evaluated on several tasks, such as visual matching [3], visual search with simple shapes [12], visual search with realistic stimuli [2]. In literature, cBCIs have been tested on several visual tasks showing their reliability. Moreover, various stud- ies suggest that voting methods are often optimal for collaborative EEG-based classification, especially when the scores of the single classifier (instead of the pre- dicted class) are used for the integration [6]. In this work we propose a voting method for a cBCI that exploits several information behind a standard ERPs stimulation Proceedings of the 8th Graz Brain-Computer Interface Conference 2019 DOI: 10.3217/978-3-85125-682-6-57