For several healthcare applications, it is important to monitor the attention level of people, especially in the fields of rehabilitation and psychology. The recent availability of cheap and portable EEG readers has enabled continuous and unobtrusive acquisition of EEG signals. Those signals may be preprocessed and analysed with machine learning algorithms to estimate the attention level of people without interfering with their current activities. In this paper, we report our experience with attention level estimation using two kinds of devices: an of-the-shelf portable EEG headset, and a more sophisticated Pervasive healthcare, human attention monitoring, EEG sensor data, supervised machine learning CEUR
The increasing availability of portable and wearable
sensors, more and more integrated in everyday objects, is
paving the way to a new generation of applications to
support personal health and well-being. Consequently,
impressive research eforts have been devoted to devise
efective techniques for recognizing human activities and
complex behaviors based on those sensor data [
Interestingly, while a vast amount of healthcare
applications use sensor-based artificial intelligence for
addressing the physical dimension of health, the mental
dimension is less investigated [
In this paper, we evaluate the use of a cheap and unobtrusive portable electroencephalography (EEG) sensor for monitoring the human attention level. Indeed, the ability to monitor human attention is fundamental for treating several conditions, including the diagnosis and nEvelop-O (D. Riboni) 2.
Material and methods
ing brainwave data on which we have applied the same
feature extraction and classification techniques. The first
dataset, named ‘Image-labeling dataset’, was acquired
using an of-the-shelf portable EEG headset with 4
channels. The second dataset, named ‘Epoc’, was acquired
using a more sophisticated EEG device with data acquired
from 7 channels. We experimented the performance of
machine learning algorithms in distinguishing attentive,
distracted, and drowsed states of the individual based
on EEG signal processing. In our experiments, only the
preprocessing phase of EEG data diverges. Indeed, the
data of the Epoc dataset are raw, so it was necessary
to employ Fast Fourier transform algorithms to extract
Delta, Theta, Alpha and Beta brainwaves. Delta waves
are related to deep sleep, unconsciousness, anesthesia,
and lack of oxygen; Theta waves activity occurs when a
person experiences emotional pressure, unconsciousness,
mental disabilities, which require ad-hoc and personal- In our work, we have considered two datasets
containor deep physical relaxation; Alpha waves are instead vis- Since brainwaves data in the dataset are raw, we
preible when an individual is in a state of consciousness, processed the data by applying the fast Fourier
transstillness, or rest, whereas when one is thinking, blinking form [
The first dataset, collected by our research group,
considers the attention level of annotators labeling a series
of images. A detailed description of the dataset can be
found in [
The brainwave values are absolute band powers, based
on the logarithm of the spectral power density (PSD) of Feature vectors are used to train and test a Random Forest
the EEG data for each channel. These values are calcu- (RF) classifier [
The participant’s task was to label indoor images that sifiers. RF and random trees were also successfully used
appeared randomly in a data annotation interface by se- for run-time brain-computer interface applications [
The second dataset was taken from the work of [
1https://choosemuse.com/ 2https://mind-monitor.com/ 3https://www.emotiv.com/epoc/ Tester 1 Tester 2 Tester 3 Tester 4 Tester 5 Tester 6 Overall 68% 62% 86% 84% 82% 100% 80%
Attent.
41 19 Attent.
47 32 Attent.
49 5 Attent.
50 9 Attent.
48 10 Attent.
60 0 Attent.
295 75
Attent.
Distracted
Table 3 and Table 4 report the results of the subjectspecific approach applied to the Epoc dataset to solve the attentive/distracted and attentive/drowsed classification problems, respectively. Finally, Table 5 and Table 6 show the results obtained by applying the leave-one-personout approach to the Epoc dataset to solve the same problems. We can make similar considerations to those made previously comparing the results of Tables 1 and Table 2, although in this case, the gap between the results obtained with the application of the two approaches is less evident. In particular, in the subject-specific approach, we have an overall Accuracy of 72% (Table 3) and 86% (Table 4), compared to an accuracy of 68% (Table 5) and 80%
Attentive Drowsed
Considering the Image-labeling dataset, we can observe that the average accuracy of distinguishing attentive and distracted states is 80% when we use a subject-specific cross validation approach; i.e., when the classifier is trained on the data of the same individual used for testing. Unfortunately, when we use a leave-one-person-out cross validation approach, the accuracy drops to 61%, which is a rather weak result for a binary classification problem. With the latter approach, we use more extensive training data, but those data are acquired from diferent people than the individual used for testing.
With the Epoc dataset, we achieved similar results. Indeed, the average accuracy of distinguishing attentive and distracted states is 72% when we use a subjectspecific cross validation approach. With the same approach, the average accuracy of distinguishing attentive and drowsed states is 86%. The recognition achieved with the latter problem is higher, probably because drowsiness is easier to distinguish from attentiveness with respect to distraction. Also with this dataset, using a leaveone-person-out cross validation approach determines a considerable drop of accuracy; i.e., 68% accuracy in distinguishing attentive from distracted states, and 80% accuracy in distinguishing attentive from drowsed states.
These results indicate that our method achieves
relatively high accuracy only when the system is trained with
data acquired from the final user of the system. Training
the system with data acquired from diferent persons
determines a relevant drop in accuracy. This fact
undermines the practical utility of this technique for some
applications, since the system would require an initial
training phase by the user which may be time-expensive
and uncomfortable. This problem may be addressed by
using transfer learning methods explicitly proposed for
EEG data [
Another worth noting finding of our experiment is that the more sophisticated device used for the Epoc dataset achieves essentially the same accuracy of the simpler device used for the Image-labeling dataset. This result indicates that even an of-the-shelf device may be efective to support some attention-aware applications. Future work includes investigating diferent machine learning algorithms for the classification task, including deep learning methods, to improve the accuracy of the system, and feature selection techniques to reduce overfitting.