CogniQ: An MCP-Driven Architecture for Hybrid Classical-Quantum Brain-Computer Interfaces

Abstract

Brain–Computer Interfaces (BCIs) enable direct neural interaction with external systems, yet their development is hindered by rigid software architectures that struggle to integrate emerging computational paradigms. This paper proposes CogniQ, a scalable Machine Learning Operations (MLOps) framework designed to provide a standardized environment for the execution and comparative evaluation of classical and quantum-enhanced EEG processing pipelines. The framework abandons monolithic scripts in favor of a distributed multiagent architecture orchestrated by an asynchronous ZeroMQ message broker and a custom Model Context Protocol (MCP).We establish a robust classical baseline by dynamically injecting two denoising autoencoders—a convolutional baseline (CDAE) and our Phase-Locked Spatiotemporal Residual Autoencoder (PLSRAE)—alongside parameter-efficient classifiers (EEGNet, EEGNet Fusion). To circumvent the strict dependency conflicts between classical deep learning libraries and quantum simulators, CogniQ introduces a ZeroMQ "Sidecar" deployment pattern. This allows complex hybrid models—such as an 8-qubit Hybrid Quantum PLSRAE, an 8-qubit Quantum Kernel SVM, and a Quantum-Enhanced EEGNet (QEEG-Net)—to execute in isolated microservices. Evaluated subject-independently over 106 subjects with 5-fold cross-validation and three seeds, and replicated on a second 51-subject cohort, the quantum-enhanced pipelines show no accuracy advantage over their classical counterparts at NISQ-era simulation scales. Fully classical configurations complete end-to-end in under 600 ms.