Passive Brain-Computer Interfaces

• Passive BCIs are systems that infer cognitive, perceptual, or affective states from unstructured neural activity without explicit user commands.
• They utilize advanced hardware and preprocessing techniques, including EEG/iEEG setups and deep learning models, to decode continuous brain signals.
• Challenges include low signal-to-noise ratios and temporal misalignments, with future directions toward multimodal fusion and improved artifact rejection.

Passive brain-computer interfaces (BCIs) constitute a class of systems that infer cognitive, perceptual, or affective states from spontaneous neural activity without requiring explicit user intent or voluntary modulation. Unlike active BCIs, where subjects deliberately perform mental tasks to generate identifiable neural signatures, passive BCIs monitor brain signals during naturalistic or implicit behavior, such as listening, observing, or simply engaging in daily activities. The principal objective of passive BCIs is to decode or interpret ongoing mental states, perceptual experiences, or functional states in real-time, thereby enabling adaptive technologies, monitoring systems, or new investigative paradigms in neuroscience.

1. Foundations and Paradigm Distinctions

Passive BCIs diverge fundamentally from active BCIs in both operational design and neuroscientific underpinnings. In the passive paradigm, the user is not required to engage in overt actions (e.g., imagined movement, motor imagery, self-initiated speech production). Instead, neural signals are acquired during unstructured or externally driven experiences. For example, during passive listening paradigms, subjects are exposed to continuous auditory stimuli (e.g., speech, music) and their brain responses are recorded without any explicit behavioral output or task engagement (Fodor et al., 2024). This distinguishes passive BCIs from the conventional protocols where controlled, repetitive, or intentional mental activities are central.

The motivation underlying passive BCI research includes the prospect of monitoring perceptual or attentional states in real-world contexts, developing assistive or surveillance applications that rely on what a user actually experiences rather than consciously communicates, and opening avenues for brain-based inference without user effort. However, the subtlety and diffuse spatial distribution of neural responses in passive paradigms impose considerable challenges for signal detection and interpretation, necessitating advanced algorithms and noise-robust methodologies (Fodor et al., 2024, Hamza-Lup et al., 2020).

2. Signal Acquisition and Hardware Setups

The implementation of passive BCIs spans a range of hardware, from fully non-invasive electroencephalography (EEG) to invasive intracranial EEG (iEEG) and stereo-EEG (sEEG). For speech perception decoding, electrode sites typically cover perisylvian regions—superior temporal gyrus (STG), posterior superior temporal sulcus (pSTS), Broca's area, and adjacent cortices. Intracranial studies acquire high-resolution data using ECoG grids, strips, and sEEG depth electrodes with channel counts ranging from 32 up to 173, and sampling rates up to 2048 Hz. Standard referencing includes mastoid or common average techniques, with subsequent preprocessing and artifact rejection implemented using pipelines such as MNE-Python (Fodor et al., 2024).

In non-invasive settings, commercially available EEG headsets with 1 to 14 channels (e.g., Neurosky MindWave, Emotiv EPOC+) are employed for continuous frontal or midline monitoring. These setups typically use dry or semi-dry electrodes, frontal placements (e.g., AF7/AF8), and reference leads as per manufacturer design. Sampling rates in passive attention studies range from 128 to 512 Hz (Hamza-Lup et al., 2020).

3. Preprocessing and Feature Extraction

Passive BCI applications rely on rigorous preprocessing to enhance the SNR given the low-amplitude, overlapping signals characteristic of passive paradigms. Common steps include:

• Selection of functional channels (ECoG/sEEG or EEG), exclusion of non-responsive or noisy electrodes.
• Notch filtering at 50 Hz and harmonics to suppress power line interference.
• Common average referencing and linear detrending to normalize spatial baselines.
• Band-pass filtering (typically 1–120 Hz for iEEG/EEG) to isolate salient spectral components associated with perception or attention (Fodor et al., 2024, Hamza-Lup et al., 2020).

For intracranial speech decoding, the signal $X(t)\in\mathbb{R}^N$ is segmented into 50 ms overlapping windows shifted in 10 ms increments. Analytic signals are computed via Hilbert transform to yield instantaneous amplitude envelopes for each electrode and time window. Channel-wise mean power $p_i(k)$ is calculated and aggregated into feature vectors $\mathbf{p}(k)$ for subsequent inference (Fodor et al., 2024).

In non-invasive attention detection, single-channel EEG is normalized (zero-mean, unit variance), smoothed with spline filters to extract delta–alpha rhythms ($<13$ Hz), processed via non-linear transformations, and sampled into fixed-length temporal feature vectors without explicit spectral decomposition (Hamza-Lup et al., 2020).

4. Machine Learning Architectures and Objectives

Passive BCIs depend heavily on machine learning for state inference due to the inherently low SNR and high variability of spontaneous signals. Architectures include:

• Fully Connected Deep Neural Networks (Fc-DNN): Mapping high-dimensional feature vectors to target states (e.g., speech mel-spectrogram frames) using dense layers with ReLU activations; output layer sizes matching regressed targets (e.g., 80 mel bins) (Fodor et al., 2024).
• 2D Convolutional Neural Networks (2D-CNNs): Exploiting spatiotemporal context by operating on $N\times W$ feature maps, combining convolutional layers, non-linear activations (swish), dropout, max-pooling, and dense regression outputs (Fodor et al., 2024).
• Feedforward and Recurrent Neural Networks (MLPs, RNNs): For attention decoding, MLPs receive temporal feature vectors, process with multiple dense layers (ReLU), and output scalar probabilities (via sigmoid), interpreted as engagement or attention indices. RNN approaches are discussed but not fully elaborated in specific reports (Hamza-Lup et al., 2020).

Optimization is typically performed using Adam with early stopping, and relevant loss functions include mean squared error for regression (e.g., mel-spectrogram reconstruction)

$$\mathcal L(\theta) = \frac{1}{K}\sum_{k=1}^K\|y(k)-\hat{y}(k)\|_2^2$$

and binary cross-entropy for categorical (e.g., Like/Dislike) attention classification

$$L = -\frac{1}{N}\sum_{i=1}^N\Big[y_i\log(\hat{y}_i) + (1-y_i)\log(1-\hat{y}_i)\Big].$$

5. Experimental Design and Performance Metrics

Passive BCI experiments typically employ naturalistic or implicit paradigms to elicit brain signals. In speech perception decoding, subjects listen passively to continuous audiovisual stimuli (e.g., a 6.5-minute movie with interleaved speech and music blocks) with no explicit response or task (Fodor et al., 2024). Attention detection experiments involve subjects viewing randomized image sequences and self-reporting preferences, while neural activity is continuously recorded (Hamza-Lup et al., 2020).

Performance evaluation depends on task: for speech decoding, validation and test mean squared error (MSE) between predicted and true spectrogram frames is standard. For attention inference, accuracy and binary cross-entropy loss are typical. Informal listening tests are noted in speech decoding studies to qualitatively assess output perceptibility, revealing that reconstructed signals preserve global features (e.g., silences) but fail to reach intelligibility (Fodor et al., 2024).

6. Current Challenges and Limitations

Passive BCI accuracy is constrained by several factors:

• Neural signatures evoked during passive paradigms are weaker and distributed across broader cortical territories, yielding lower SNR compared to explicit production tasks (Fodor et al., 2024).
• Temporal alignment of neural data with perceptual streams (e.g., audio) is challenging due to variable processing delays and inter-subject heterogeneity.
• For speech decoding, even with dense intracranial coverage, current models achieve best validation MSEs on the order of 0.65–0.80 (for mel-spectrogram regression), which is insufficient for reconstructing intelligible signal (Fodor et al., 2024).
• Invasive iEEG and sEEG, though higher fidelity, are not viable for population-scale or out-of-lab applications; non-invasive approaches face even lower SNR and limited spatial resolution (Fodor et al., 2024, Hamza-Lup et al., 2020).
• Attention-state decoding remains predominantly binary, with engagement indices being interpreted only indirectly.

7. Future Directions and Implications

Research into passive BCIs is expanding toward several technical and conceptual frontiers:

• Signal Alignment and Modeling: Integration of synchrony estimation methods (e.g., transformer-style alignment layers, lag compensation), and advanced temporal models such as temporal convolutional networks and transformer architectures to capture long-range dependencies (Fodor et al., 2024).
• Multimodal Data Fusion: Combining EEG/iEEG with fMRI, MEG, or peripheral biosignals (e.g., EMG, eye tracking) to enrich the feature space and improve inference reliability (Fodor et al., 2024).
• Interpretability and Feature Attribution: Developing model-agnostic sensitivity analyses to determine which spatial, spectral, or temporal features are most discriminative for passive perceptual states (Fodor et al., 2024).
• Hybrid BCIs: Fusing passive decoding of perceptual/affective states with active volitional control to enable adaptive interfaces and closed-loop systems that respond both to intent and subconscious experience (Fodor et al., 2024).
• Non-Invasive Expansion: Improving artifact rejection and feature extraction for dry-electrode or wearable EEG platforms to extend passive BCI applicability beyond clinical/invasive contexts (Hamza-Lup et al., 2020).
• Dataset Scalability: Systematic collection of larger, more diverse datasets spanning naturalistic listening/viewing to support statistical robustness and generalization.

The methodological advances in passive BCI preprocessing, feature extraction, and end-to-end deep modeling are foundational for future systems capable of monitoring or reconstructing natural mental states in real time. While key technical hurdles, especially regarding signal interpretability and usability, persist, evidence to date demonstrates that passive paradigms elicit decodable neural signatures that future BCIs can exploit for both fundamental neuroscience and translational applications (Fodor et al., 2024, Hamza-Lup et al., 2020).