<p>Advances in upper-limb prosthetic control increasingly rely on artificial intelligence (AI) methods, particularly machine learning (ML) and deep learning (DL), for decoding motor intent from biosignals. This review presents a structured synthesis of approaches based on surface electromyography (sEMG), electroencephalography (EEG), and hybrid multimodal systems. The evolution of learning paradigms is analysed from classical ML methods to modern DL architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), Transformers, and graph-based models, with emphasis on their ability to capture spatial, temporal, and non-stationary characteristics of biosignals. Signal acquisition and pre-processing pipelines are examined in relation to electrode configurations, noise suppression, and their impact on robustness and generalization. Multimodal strategies are systematically categorized into early, late, and hybrid fusion frameworks, and evaluated in terms of accuracy, noise resilience, and computational trade-offs. A key observation is that EMG-EEG integration exploits complementary neural and muscular information, resulting in improved stability under signal variability. The review further consolidates performance of AI models across standard benchmark datasets and identifies critical limitations in current evaluation practices, including dataset heterogeneity, absence of unified cross-modal benchmarks, and inconsistent reporting protocols, which hinder reproducibility and fair comparison. System-level considerations, including edge deployment, real-time inference, and hardware integration, are analysed alongside explainability, regulatory, and ethical requirements for clinically viable systems. The paper highlights open challenges and outlines directions for developing scalable and generalizable multimodal ML/DL frameworks for biosignal-driven prosthetic control.</p>