Hyperdimensional Computing for ADHD Classification using EEG Signals

Can artificial intelligence identify patterns associated with Attention Deficit Hyperactivity Disorder (ADHD) by analyzing brain activity, even when only a small amount of data is available? ADHD is a neurodevelopmental disorder characterized by difficulties involving attention, impulsivity, and hyperactivity. Its assessment generally relies on clinical observations, behavioral evaluations, and standardized questionnaires. However, the complexity of its symptoms and their overlap with other conditions can make the diagnostic process challenging. Electroencephalography (EEG) offers an additional perspective by recording the brain’s electrical activity through electrodes positioned on the scalp. These signals contain information about neural dynamics, but extracting meaningful patterns from them is far from straightforward. Artificial intelligence (AI) can help identify subtle differences in EEG recordings that might otherwise remain difficult to recognize. The challenge is developing models that are accurate, computationally efficient, and capable of learning from the relatively small datasets commonly available in healthcare. A promising solution comes from Hyperdimensional Computing (HDC), a brain-inspired computational paradigm that represents information in extremely high-dimensional mathematical spaces.

Traditional deep learning approaches, including Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, learn patterns by repeatedly adjusting large numbers of parameters. Although effective in many applications, these architectures can require substantial computational resources and extensive training datasets. HDC follows a fundamentally different principle. Instead of learning through the iterative optimization of neural network weights, HDC transforms information into mathematical structures called “hypervectors”. These vectors contain thousands of dimensions, allowing complex relationships to be represented through relatively simple operations. The key advantage lies in how information is distributed across the entire representation. Rather than depending on individual features, HDC captures patterns collectively, making the resulting representations relatively robust to noise and small variations in the input data. For EEG analysis, this means that the electrical activity recorded from different brain regions can be combined into a compact mathematical representation while preserving important characteristics of the original signals. This principle forms the foundation of ADHDC, a classification framework designed to distinguish EEG recordings associated with ADHD from those of individuals without the condition.

Brain signals are inherently complex because they evolve over time and originate from different anatomical regions. Recognizing meaningful patterns therefore requires understanding both when particular electrical activities occur and where they are recorded. ADHDC addresses this challenge through a spatio-temporal encoder, which transforms EEG recordings into high-dimensional representations. The temporal component captures how electrical signals change over time. EEG recordings are divided into short, non-overlapping segments, and each segment is encoded using mathematical operations that preserve the order of its signal values. The spatial component associates these temporal representations with the electrodes from which the signals originate. By combining information from multiple channels, the system creates a unified representation of the recorded brain activity. The resulting hypervectors are compared with reference representations associated with ADHD and control subjects. Classification is performed by measuring mathematical similarity, without requiring the iterative training procedures typical of deep neural networks. Another important characteristic is the limited amount of preprocessing required. The EEG recordings undergo basic operations such as removing the initial noisy seconds, reducing the sampling frequency, and quantizing signal amplitudes. This approach simplifies data preparation while preserving information that may be relevant for distinguishing the two groups.

The framework is evaluated using EEG recordings from 79 individuals, including 37 participants with ADHD and 42 control subjects. The analysis focuses on resting-state recordings collected while participants kept their eyes open. Two EEG channels, F4 and Cz, are selected to capture relevant brain activity without introducing the effects of additional cognitive tasks. The dataset is divided into training and testing groups, with 59 participants used for training and 20 for evaluation. To assess the effectiveness of the approach, ADHDC is compared with three established machine learning architectures: a Support Vector Machine (SVM), a CNN, and an LSTM network. ADHDC outperforms the other models in terms of average accuracy, F1-score, and recall, indicating a strong ability to identify participants belonging to the ADHD group. However, the most significant finding concerns the amount of training data required. When trained using EEG recordings from only 7 participants, corresponding to approximately 8.9% of the complete dataset, the hyperdimensional model already achieves an accuracy of approximately 72.9%. This performance exceeds that of the LSTM benchmark trained using the full available training set. The result demonstrates how high-dimensional representations can extract useful discriminatory information even when the number of training examples is severely limited. Computational efficiency is equally notable. Training requires approximately 15 seconds on a conventional CPU-based computer, while classifying an individual EEG segment takes approximately 0.001 seconds. By comparison, the CNN and LSTM benchmarks require more than 4 minutes for training on the same hardware. These findings suggest that effective EEG classification does not necessarily require increasingly complex neural architectures or specialized computing infrastructure.

The combination of high classification accuracy, limited training requirements, and lightweight computation creates interesting opportunities for AI-assisted neurological assessment. In clinical environments, collecting large quantities of labeled biomedical data is often difficult. Patient privacy, limited access to specialized examinations, and the complexity of medical data acquisition can restrict the development of conventional machine learning systems. HDC offers an alternative by making more efficient use of the information already available. A lightweight EEG classification system could potentially support clinicians during preliminary screening, helping identify cases that warrant further investigation and reducing the computational burden associated with automated signal analysis. Importantly, such a system would complement rather than replace established clinical diagnostic procedures. ADHD cannot be diagnosed from EEG classification alone, and the reported results require further validation across larger and more diverse populations before clinical deployment. Additional developments could focus on improving the interpretability of hyperdimensional representations, identifying which EEG characteristics contribute most strongly to classification, and extending the approach to other neurodevelopmental conditions. Integration with embedded or mobile computing platforms could also make automated EEG analysis more accessible in environments without specialized hardware. The broader significance extends beyond ADHD detection. It demonstrates that AI can become more effective not only by increasing model complexity, but also by fundamentally rethinking how information is represented and processed. By combining brain-inspired computation with the analysis of brain activity, HDC points toward a new generation of biomedical AI systems designed to learn efficiently, operate quickly, and extract meaningful information from limited data.