Abstract— Cognitive workload, which is the level of mental effort required for a cognitive task, can be assessed by monitoring the changes in neurophysiological measures such as electroencephalogram (EEG). This study investigates the performance of an EEG-based Brain-Computer Interface (BCI) to discriminate different difficulty levels in performing a mental arithmetic task. EEG data from 10 subjects were collected while performing mental addition with 3 difficulty levels (easy, medium and hard). EEG features were then extracted using band power and Common Spatial Pattern features and subsequently features were selected using Fisher Ratio to train a Linear Discriminant Classifier. The results from 10-fold cross-validation yielded averaged accuracy of 90% for 2 classes (easy versus hard tasks) and 66% for 3 classes (easy versus medium versus hard tasks). Hence the results showed the feasibility of using EEG-based BCI to measure cognitive workload in performing mental arithmetic. I. INTRODUCTION Cognitive workload or mental workload refers to the level of mental effort required to perform a task such as problem- solving, reading and writing, and performing mental arithmetic. Heavy cognitive workload for a prolonged period of time could lead to cognitive overload. This results in issues such as a drop in learning gain [1], affecting task performance [2]or even critical errors in situations where high levels of sustained attention are required. In contrast, low cognitive workload could lead to cognitive underload, as the task comes too easy, resulting in boredom and reducing the learning gain as well. Hence identification of mental workload during cognitive tasks is an important area in cognitive neuroscience and other real-world applications. The development of a real-time assessment of cognitive workload could lead to building smart systems to adapt training content and keep subjects in an optimal range of cognitive workload to ensure learning progress[1] or to even proactively warn subjects about their stress during these workload states [3]. The level of cognitive workload depends on factors such as prior experience, physical condition, and varies from This work was supported by the Science and Engineering Research Council of A*STAR (Agency for Science, Technology and Research), Singapore. Z. Y. Chin, C. Wang, K. K. Ang are with Institute for Infocomm Research, A*STAR, 1 Fusionopolis Way #21-01 Connexis (South Tower) Singapore 138632 (email: {zychin, ccwang, kkang }@i2r.a-star.edu.sg). X.Zhang was a research attachment student at the Institute for Infocomm Research and is currently studying at Temasek Junior College, Singapore (email: amanda4zhangxin@gmail.com) person to person. Hence, there are conventional methods to measure cognitive workload, which include verbal or written feedback from the subject’s point of view. For example the NASA-Task Load Index (TLX) subjective ratings could be used to evaluate cognitive workload of system operators. However, the assumption is that the subject could accurately identify and report changes in cognitive workload, which may be subjected to confirmation biases [4]. Furthermore, to evaluate the cognitive workload of many subjects concurrently based on subjective feedback in real time e.g. 30-40 students in a classroom for a teacher would be challenging [2]. To complement conventional methods, physiological measures such as Galvanic Skin Response (GSR), heart rate, as well as brain signals, have been investigated for their effectiveness to provide real-time assessment or classification of cognitive workload. Existing studies have investigated the changes in cognitive workload during mental arithmetic. In previous work [5],[6], the authors measured Near-Infrared Spectroscopy (NIRS) brain signal changes when subjects were performing a mental arithmetic task involving digit additions. The task was categorized into 2 difficulty levels which varied in the number of digits involved and the presence / absence of a carry-over operation. They proposed classification algorithms which yielded an overall intra- subject accuracy of 73% in classifying hard versus easy mental arithmetic tasks and 92% in classifying hard task versus a neutral state. Besides NIRS, changes in scalp-recorded brain signals or electroencephalogram (EEG) is also employed to detect changes in mental workload [7]. Due to its relatively higher temporal resolution, EEG can capture fast and dynamically changing brainwave patterns in complex cognitive tasks. Examples of studies that employ EEG-based classification of mental arithmetic are as follows: In [8], the authors trained their classifier to discriminate between 3 conditions: relaxed, low workload and high workload using mental arithmetic, which obtained an average accuracy of 62%. In [9], the authors proposed a generalized Higuchi fractal dimension spectrum method for mental arithmetic task recognition, which yielded a subject-dependent accuracy of 97.9% for mental arithmetic versus the relaxed state. In [1], the authors proposed an inter-subject classifier to classify EEG during 3 levels of mental arithmetic task difficulty. Their classifier yielded around 56% accuracy averaged over 10 subjects, which could classify easy and medium tasks at 68% and 84% but below chance level at 16% for hard tasks. However, one limitation that the authors identified was the tasks were presented in order of increasing difficulty level and hence classification accuracy could have been affected by EEG-based discrimination of different cognitive workload levels from mental arithmetic Zheng Yang Chin, Xin Zhang, Chuanchu Wang, Kai Keng Ang 978-1-5386-3646-6/18/$31.00 ©2018 IEEE 1984