This study explores a novel approach to emotion recognition through machine learning, addressing the limitations of previous methods. While deep learning has shown promise in this field, it often requires significant computational resources and time. In response, we propose a multimodal approach utilizing Feature-level Fusion (FLF) and Decision-level Fusion (DLF) to enhance performance while reducing complexity. The study focuses on integrating electroencephalogram (EEG), electromyography (EMG), and electrooculogram (EOG) signals. Signal preprocessing involves extracting statistical features, power spectral density (PSD), and incremental entropy analysis. Recursive Feature Elimination (RFE) is employed as a feature selector, facilitating the fusion of different signal features. Three fusion strategies are explored: EEG with EOG, EEG with EMG, and a combination of EEG with EOG and EMG. For classification, the Bagging Classifier and K-Nearest Neighbors Algorithm are chosen. Results demonstrate promising accuracy rates, with 95.7% for arousal, 96.41% for valence in subject-dependent classification, 93.68% for arousal, and 93.23% for valence in subject-independent classification. This approach presents a viable alternative to deep learning methods, offering improved performance with reduced computational burden.