Human stress is a serious concern in this era of an ever-increasing number of challenges. Stress causes detrimental effects on the health of an individual and imparts a heavy financial burden on society. Stress can be a forebringer to reduced working ability, a weak immune system, anxiety, depression, blood pressure and cardiovascular diseases. Traditional psychological methods to measure stress require trained experts and human intervention. In developing countries, there is a shortage of health facilities especially, related to mental health. A large number of stress patients remain undiagnosed due to a shortage of health facilities and stigma associated with a mental checkup. With the advancement of brain-computer interfaces, it is possible to make an objective measure of stress using electroencephalography (EEG). Machine learning and statistical methods have been successfully used for the quantification and analysis of human stress. Conventionally, EEG analysis is performed using a clinical dense electrode system, which could consist of up to 132-channels and difficult to wear and operate. This thesis presents methods for quantification and analysis of human stress from EEG recordings using commercially available EEG headsets, which are cost-effective, wearable and provide good temporal resolution. Three different sets of experiments are conducted to acquire EEG datasets using three different commercially available EEG headsets. The first study uses a classification and regression model to identify the longterm stress and its key neural oscillatory features using a single-channel EEG headset. Perceived stress scale (PSS) scores are used for labelling participants into stress and control group. The Naïve Bayes algorithm classifies chronic stress with an accuracy of 71.4% using the single-channel EEG headset. Computation time is reduced by a factor of 7% when the low beta is used as a sole significant feature, which is selected by multiple linear regression. The correlation-based feature subset selection method selected three features low beta, high beta and low gamma oscillations and improved the classification accuracy up to 78.59% using a support vector machine algorithm. The second study assesses the long-term stress level with multiple machine learning algorithms using the five channel EMOTIV Insight headset. A dataset consisting of EEG, PSS scores and expert’s evaluation of stress is collected. In the multichannel analysis, overall alpha asymmetry, beta and low gamma in the frontal region are selected as statistically significant features by using Student’s t-test over forty-five different features. Only four active electrodes are required for the classification of human stress with 85% accuracy using the sole feature of overall alpha asymmetry. The third study utilizes the identified feature of overall alpha asymmetry for expertnovice classification of a video game player. Only a single feature of overall alpha asymmetry classifies expertise level of a game player with an accuracy of 82.35%. It verifies the applicability of identified features to a stimulus-based scenario such as video game playing. For further improvement of classification results, the temporal, frontal and overall alpha asymmetries are used as features. With only four electrodes, the expertise level of a game player is classified with an accuracy of 94.1% and it proved to be better than the accuracy of 14- and 4-channel classifications. This points to the fact that stress research has potential applications, where cognitive aspects of an individual are considered for performance improvement as stress based features are used In the future, the daylong monitoring of stress using EEG in an out-of-lab environment will be considered. More participants will be employed for long-term stress monitoring to generalize results. Video game stimulus will be compared with standard stress- inducing tasks like a mental arithmetic task to get a detailed insight regarding stress assessment.