An auditory oddball brain–computer interface for binary choices

Abstract

Objective

Brain–computer interfaces (BCIs) provide non-muscular communication for individuals diagnosed with late-stage motoneuron disease (e.g., amyotrophic lateral sclerosis (ALS)). In the final stages of the disease, a BCI cannot rely on the visual modality. This study examined a method to achieve high accuracies using auditory stimuli only.

Methods

We propose an auditory BCI based on a three-stimulus paradigm. This paradigm is similar to the standard oddball but includes an additional target (i.e. two target stimuli, one frequent stimulus). Three versions of the task were evaluated in which the target stimuli differed in loudness, pitch or direction.

Results

Twenty healthy participants achieved an average information transfer rate (ITR) of up to 2.46 bits/min and accuracies of 78.5%. Most subjects (14 of 20) achieved their best performance with targets differing in pitch.

Conclusions

With this study, the viability of the paradigm was shown for healthy participants and will next be evaluated with individuals diagnosed with ALS or locked-in syndrome (LIS) after stroke.

Significance

The here presented BCI offers communication with binary choices (yes/no) independent of vision. As it requires only little time per selection, it may constitute a reliable means of communication for patients who lost all motor function and have a short attention span.

Introduction

Brain–computer interface (BCI) systems provide a means of communication for locked-in state (LIS) patients or serve as an aid in restoration of motor function for stroke patients (Kübler and Neumann, 2005). Patients in the LIS have only residual control over few muscles (e.g., vertical eye movement) which may be unreliable. A common principle of all BCI paradigms is that they must be controllable without the use of muscles to suit the needs of the aforementioned patient groups. Control is provided to the user by recording a signal from the brain and translating this signal into a defined action, the result of which is fed back to the user. Currently most BCI paradigms are based on control signals extracted from the electroencephalogram (EEG). Alternative BCIs not using EEG but functional magnetic resonance imaging (Sitaram et al., 2007a), implanted sensors (Leuthardt et al., 2004, Hochberg et al., 2006), magnetoencephalography (MEG) (Mellinger et al., 2007) and near infrared red spectroscopy (NIRS) (Sitaram et al., 2007b, Wriessnegger et al., 2008) have also been demonstrated but have severe limitations such as the need for surgery or limited practicability for daily use (non-portable devices: MEG, fMRI). Therefore, only EEG controlled BCIs are currently applicable to a broad range of patients. EEG BCIs have been implemented using different components of the EEG such as slow cortical potentials (SCPs), sensorimotor rhythms (SMR) and event-related potentials (ERPs) and have been used extensively with patients in the LIS (Kübler et al., 2007). One particular EEG based BCI controlled by the P300 ERP (Farwell and Donchin, 1988) has been successful with patients (Nijboer et al., 2008, Hoffmann et al., 2008, Sellers and Donchin, 2006, Neshige et al., 2006, Silvoni et al., 2009).

The P300 is a positive deflection in the EEG with variations in latency between 250 and 700 ms on central to parietal locations (Polich et al., 1997). The visual P300 BCI paradigm consists of a matrix, usually 6 × 6, with letters and numbers. For visual stimulation rows and columns flash randomly. For selection the user is required to focus attention on the cell with the desired item. Any cell flashes only twice in a sequence of 12 flashes (6 rows, 6 columns) and thus, becomes an oddball which elicits a P300. Apart from spelling, these systems can also be used to control various applications, such as an internet browser (Mugler et al., 2008), a painting application (Kübler et al., 2008) and even wheelchair control (Iturrate et al., 2009, Rebsamen et al., 2006).

Generally, successful use of a BCI is defined by the accuracy (percent of targets selected correctly) achieved, the information content of each selection (i.e. the number of different targets) and the time needed for one selection. All three factors determine the speed of the communication possible with the BCI (usually defined as information transfer rate (ITR) in bits/min). In contrast to SCP and SMR based BCIs, the visual P300 BCI permits fast communication without user training for healthy participants (e.g., 94.6% and 99% in Furdea et al. (2009) and Kleih et al. (in preparation)). Although machine learning approaches for BCIs with the SMR as input signal have been introduced and proven to allow for high accuracies with less than an hour of training, it has not yet been shown whether these approaches can be used with LIS patients (Blankertz et al., 2008). Performance of patients with motor disorders is usually lower than that of healthy controls (e.g., 78.8% in Nijboer et al. (2008), 68.6% in Piccione et al. (2006) and 37.7% in Kübler et al. (2009)). In later stages of amyotrophic lateral sclerosis (ALS) patients lose all voluntary muscular control (Averbuch-Heller et al., 1998, Jacobs et al., 1981), including voluntary control of gaze, which makes the use of a visual BCI difficult if not impossible. Since these are the users that would benefit most from a BCI, this is a problem that must be addressed.

Therefore, P300 BCI systems based on auditory stimulation have been proposed. One of the first studies (Sellers and Donchin, 2006) used the words “Yes”, “No”, “Stop”, “Pass” as possible targets with three healthy subjects and three ALS patients. On average the subjects achieved accuracies of 65% which results in a communication speed of 0.43–1.80 bits/min. A recent study transferred the P300 spelling matrix to the auditory modality. Numbers were assigned to identify rows and columns. For selection, the numbers of the rows and columns were presented to the user (Furdea et al., 2009). The participants achieved 65% accuracy. Due to the high number of bits per selection communication speed was higher (up to 2.85 bits/min) than in the study by Sellers and Donchin (2006). Most recently, a design similar to that of Furdea and colleagues using sounds instead of words was described by Furdea et al., 2009, Klobassa et al., 2009). Online accuracies of 60% and offline accuracies of 70% were achieved with auditory stimulation. Despite these promising results with auditory P300 BCIs all three methods (Sellers and Donchin, 2006, Furdea et al., 2009, Klobassa et al., 2009) have a low average accuracy. Even though 60% of the participants achieved a very reliable performance of more than 70% accuracy in the study of Furdea and colleagues a different method is needed for those that are incapable of using a multi-class auditory P300 BCI. Instead of words, tones can be applied as stimuli. For example, Hill and colleagues presented their participants with two streams of rare and standard tones, each constituting an oddball paradigm from two different directions (Hill et al., 2004). To make a binary selection, users were required to focus their spatial attention on only one of the two auditory oddball streams. Promising accuracies between 63% and 97% were achieved in healthy people.

In addition to the loss of motor control, locked-in patients may also suffer from reduced attention spans (Birbaumer, 2006, Smith and Delargy, 2005, Vieregge et al., 1999). Therefore, a paradigm optimized for LIS patients should employ stimuli distinguishable with minimum effort. Furthermore, we expect accuracies achieved with an auditory P300 BCI to be reduced in patients as compared to healthy controls (Sellers and Donchin, 2006, Furdea et al., 2009, Klobassa et al., 2009). Thus, an auditory BCI for patients has to take into account the trade off between speed and accuracy. In an extreme case, low accuracies close to 50% achieved with only a short time needed per selection lead to high ITRs in bits/min. Nonetheless, this is not feasible in practice as the amount of times the “delete” function has to be selected significantly increases the time to spell the intended letter to unbearable levels (Sellers et al., 2006). Therefore, a BCI for CLIS patients should be designed with a probable reduction of accuracy in mind.

The goal of this study was to evaluate a BCI design based on a simple paradigm that would permit the user a binary selection with high accuracies while still offering a competitive communication speed. A three-stimulus oddball paradigm (Katayama and Polich, 1996, Polich, 2007) was developed for this purpose. In this paradigm the subject is presented with two targets and a series of frequent standard tones randomized in sequences of e.g., 7 stimuli. Target one and target two differed in one physical property (loudness, pitch or location). The experiment aimed at determining which type of stimulus characteristics chosen for discriminating between targets 1 and 2 would lead to the best discrimination and thus, to the highest information transfer rate (ITR). The standard tones were identical in all three tasks (pink or $1/f$ noise). Selections were made by focusing on either one of the two targets. We hypothesized that such a design would boost accuracy to higher levels as seen in current auditory systems, at the cost of bits per selection. We further investigated whether reliable classification would be possible when using only one or two averaged trials . Finally, we analyzed which type of difference (pitch, loudness or spatial location) would lead to the best discrimination results between targets and non-targets.