Résumé : Mental stress is a largely prevalent condition known to affect many people and could be a serious health concern. The quality of human life can be significantly improved if mental health is properly managed. Towards this, we propose a robust method for perceived stress classification, which is based on using multimodal data, acquired from forty subjects, including three (electroencephalography (EEG), galvanic skin response (GSR), and photoplethysmography (PPG)) physiological modalities. The data is acquired for three minutes duration in an open eyes condition. A perceived stress scale (PSS) questionnaire is used to record the stress of participants, which is then used to assign stress labels (two- and three classes). Time (four from GSR and PPG signals) and frequency (four from EEG signal) domain features are extracted. Among EEG based features, using a frequency band selection algorithm for selecting the optimum EEG frequency subband, the theta band was selected. Further, a wrapper-based method is used for optimal feature selection. Human stress level classification is performed using three different classifiers, which are fed with a fusion of the selected set of features from three modalities. A significant accuracy (95% for two classes, and 77.5% for three classes) was achieved using the multilayer perceptron classifier.