Each of the 3936 trials was segmented into labeled windows corresponding to:

Segments were extracted from \(\:_\), representing the single-channel sensorimotor input.

Evaluation scenarios:

Within-session: Random train/test split per session.

Cross-session: Train on one session, test on a different session.

Adaptive few-shot: 5 labeled calibration trials provided to the proposed model.

For the few-shot recalibration experiments, the calibration trials were selected as the first chronological trials from the target session to simulate a realistic setup phase in practical BCI deployment. After recalibration using these initial trials, all remaining trials in the session were reserved for evaluation. No random sampling across the session was performed.

Gamma-burst detection

Figure 1 shows the single-channel EEG data for the grasp-and-lift motor task, along with the output of the proposed gamma burst detection algorithm.

Fig. 1

Single-Channel EEG and Detected Gamma Bursts

The top panel shows the simulated EEG trace recorded over the sensorimotor cortex, \(\:_\). The bottom panel displays the gamma-band envelope (30–80 Hz) extracted using the Hilbert transform after bandpass filtering. The dashed horizontal line denotes the adaptive threshold, enabling robust burst isolation under real-world noisy conditions. The shaded regions highlight the detected gamma bursts, which cluster tightly around movement onset – consistent with corticospinal engagement during grasp and lift execution.

Fig. 2

Time-Frequency Representation with Gamma Bursts

The heatmap in Fig. 2 shows Morlet-based wavelet power (5–80 Hz) over time. The dashed vertical line indicates the movement onset, and shaded bands correspond to intervals where the gamma-band envelope (30–80 Hz) exceeds the adaptive threshold and is classified as a gamma burst.

Fig. 3

Effect of Few-shot adaptation on Cross-Section Decoding

In Fig. 3, we can observe the effect of few-shot adaptation on cross-section decoding accuracy for the proposed gamma burst-based deep learning model. Accuracy is shown as mean \(\:\pm\:\) standard deviation across subjects for 0, 5, and 10 labeled calibration trials from the target session. The proposed model exhibits substantial improvements with even minimal calibration data whereas the baseline model shows only marginal gains, indicating limited ability to correct session-specific distributed shifts without adaptive mechanisms.

Fig. 4

t-SNE embeddings of features

Two-dimensional t-SNE embeddings of target-session feature representations before (left) and after (right) adaptation are shown in Fig. 4. Each point corresponds to a single trial segment and is colored according to its motor state (Rest, Preparation, Execution). After adaptation, the class clusters become more compact and better separated, indicating that the adaptive normalization and recalibration layers align the feature space to the target session and enhance discriminability.

Table 1 provides the performance metrics of the proposed architecture in comparison with other state-of-the-art techniques in single-channel decoding.

Table 1 Classification Performance (Mean \(\:\pm\:\) SD Across Subjects)

The proposed model achieved 89% accuracy within-session, demonstrating strong decoding ability from a single channel. The cross-section accuracy of 78% represents a > 10% absolute gain over the best non-adaptive baseline. Th ablation analysis shows:

Gamma bursts alone outperform classical ERD/ERS.

Adaptation contributes ~ 7% in cross-session generalization.

Full model synergy yields the highest performance.

Repeated-measures ANOVA confirms a significant improvement (p < 0.01).

Few-shot adaptation performance

The adaptive recalibration layer was evaluated using N = 5 labeled calibration trials from the test session.

The adaptive recalibration layer was evaluated using N = 5 labeled calibration trials from the test session, with the results shown in Table 2.

Table 2 Few-Shot Adaptation

It can be seen that with 5 calibration trials, there is an improvement in the generalization. At 10 calibration trials, performance approaches within-session levels. This demonstrates the practical viability of low-effort calibration for real-world BCI use.

Effect of burst features on decision dynamics

We evaluated the discriminative power of burst descriptors via Mutual Information (MI) against class labels.

Table 3 Average Mutual Information (bits)

Gamma bursts provide the highest motor-stat information content, reinforcing neuroscientific literature suggesting bursts reflect short-lived cortical activation.

Fig. 5

Classification Accuracy by Method

Figure 5 shows the comparison of within-session and cross-session classification accuracy across baseline models and the proposed framework. The proposed gamma burst-based adaptive model consistently outperforms all baselines, with the largest gains observed in cross-session generalization.

Fig. 6

Figure 6 shows the confusion matrices of the proposed gamma-burst adaptive deep learning framework under three conditions: (left) within-session, (middle) cross-session without adaptation, and (right) cross-session with few-shot adaptation. Adaptation reduces misclassifications, particularly between Rest and Preparation, leading to improved cross-session generalization.

Contextual comparison with multi-channel methods

Although the primary objective of this study is to investigate decoding performance under a strict single-channel constraint, it is informative to contextualize the results relative to representative multi-channel approaches evaluated on the same WAY-EEG-GAL dataset.

Table 4 summarizes selected multi-channel studies alongside the proposed method [22]. employed a 32-channel EEG configuration with convolutional neural networks for six-event grasp-and-lift detection and reported an average AUC of 0.944 [23]. used full multi-channel EEG with CSP and continuous wavelet transform features combined with GoogLeNet and achieved 78.08% intersubject four-class accuracy.

In comparison, the proposed single-channel burst-centric adaptive framework achieves 89.0% within-session accuracy and up to 81.0% cross-session accuracy with limited calibration. While multi-channel systems retain advantages due to spatial filtering and improved signal-to-noise ratio, the relatively small performance gap highlights the discriminative strength of temporally structured burst descriptors combined with adaptive normalization.

Table 4 Comparative Study with Multi-Channel Inputs

Comparison with continuous gamma envelope input

A natural question is whether the explicit extraction of burst onset, duration, and recurrence statistics is necessary, or whether a convolutional neural network could implicitly learn burst-like temporal patterns directly from the continuous gamma-band envelope.

To investigate this, we conducted an ablation experiment in which the explicit burst descriptor stream was removed. Instead, the analytic gamma envelope \(\:_\left(t\right)\) was provided directly as a continuous time-series input to the same temporal CNN–GRU backbone. All architectural components, training procedures, and adaptive normalization mechanisms were kept identical to the proposed model.

Table 5 summarizes the results.

The results indicate that while the continuous gamma envelope provides substantial discriminative information, explicit burst-dynamics descriptors yield consistent performance gains of approximately 4–5% absolute accuracy in both within-session and cross-session settings. This suggests that although deep networks can learn local amplitude fluctuations, the structured representation of burst onset timing, duration, inter-burst intervals, and recurrence statistics provides additional higher-order temporal information that is not fully captured through implicit convolutional feature learning alone.

In particular, burst recurrence and inter-burst interval statistics encode event-level temporal organization across hundreds of milliseconds, which may require deeper receptive fields or larger datasets to be reliably inferred directly from raw envelope signals. The explicit descriptor representation therefore serves as an inductive bias that stabilizes learning in the data-limited single-channel regime.

Sensitivity to threshold multiplier and minimum duration

Burst detection relies on two key parameters: the adaptive threshold multiplier \(\:\alpha\:\), which scales the baseline standard deviation, and the minimum duration constraint \(\:_\), which suppresses short transient events.To evaluate robustness, we varied \(\:\alpha\:\) within ± 20% of its nominal value and tested \(\:_\) across a range corresponding approximately to 1–3 gamma cycles (e.g., 15–45 ms for 30–80 Hz). For each parameter configuration, burst features were re-extracted and the full decoding model retrained under identical conditions.

Table 6 summarizes the results.

Table 6 Sensitivity analysis

The results indicate that decoding performance varies smoothly across the tested parameter range, with fluctuations remaining within approximately 2–3% absolute accuracy. Lower \(\:\alpha\:\) values increase burst density, while higher \(\:\alpha\:\) values yield more conservative detections; similarly, shorter \(\:_\) values admit brief events, whereas longer durations emphasize sustained bursts. Importantly, no abrupt collapse in performance was observed, suggesting that the proposed burst-detection framework is robust to moderate parameter variations.

Consideration of EMG contamination in the gamma band

A potential limitation of gamma-band EEG analysis is contamination from cranial muscle activity, particularly during overt movement execution. EMG artifacts are typically broadband, high-amplitude, and short-duration events that may overlap spectrally with cortical gamma oscillations.

Several aspects of the present framework mitigate this concern. First, preprocessing incorporated band-limited filtering and wavelet-domain attenuation to suppress broadband transient components. Second, burst detection required not only threshold crossing but also a minimum duration and envelope continuity constraint, reducing sensitivity to isolated high-frequency spikes characteristic of muscle artifacts.

Importantly, burst modulation was observed not only during movement execution but also during the motor preparation interval (− 300 to 0 ms relative to lift-off), where overt muscle activation is minimal. The presence of structured gamma burst dynamics during preparation suggests a cortical contribution beyond purely peripheral muscle activity.

Furthermore, gamma bursts were defined within a bounded frequency range (30–80 Hz), excluding higher-frequency broadband components more typical of EMG contamination. While single-channel analysis cannot completely eliminate the possibility of residual EMG influence, the structured temporal patterns and cross-session reproducibility observed in this study support the interpretation that the extracted features capture meaningful neural dynamics.

Nevertheless, future work may incorporate simultaneous EMG-informed artifact rejection or source-separation techniques to further isolate cortical gamma activity in single-sensor systems.

Contribution of individual burst descriptors

To examine the relative importance of individual burst descriptors, we performed a structured ablation study. Starting from the full burst-dynamics feature set, descriptor groups were selectively removed while keeping all other components of the decoding pipeline unchanged. The evaluated descriptor groups included:

Table 7 summarizes representative within-session and cross-session decoding accuracies.

Table 7 Comparison of Individual Burst Descriptors

The results suggest that recurrence and inter-burst interval (IBI) statistics provide the largest contribution to decoding performance, particularly in cross-session scenarios. Removal of these features resulted in the greatest performance degradation. Burst duration and peak amplitude also contributed meaningfully, while onset timing features alone had comparatively smaller impact.

These findings indicate that the temporal organization and rhythmic recurrence of gamma bursts carry substantial discriminative information beyond instantaneous amplitude measures. In particular, recurrence patterns may reflect structured cortical activation during motor preparation and execution, supporting the hypothesis that burst dynamics encode event-level neural coordination rather than isolated oscillatory peaks.

Computational latency and real-time feasibility

For practical BCI deployment, computational latency is a critical consideration. The proposed framework consists of three primary stages: gamma-envelope computation, burst feature extraction, and neural network inference.

In our implementation (Python with GPU acceleration), the forward pass of the CNN–GRU model requires approximately 6–12 ms per trial window. Gamma-envelope computation and burst detection for a 500 ms segment require approximately 10–20 ms. Thus, the cumulative computational latency for processing a single window is approximately 20–35 ms.

It is important to distinguish computational latency from window-induced temporal delay. Reliable burst detection requires a short temporal context (e.g., 300–500 ms). In an online setting, predictions can be generated using overlapping sliding windows updated every 50–100 ms, allowing quasi-continuous decoding with sub-100 ms update intervals. Therefore, while the intrinsic temporal context introduces a short buffering delay, the computational overhead of the proposed method is compatible with near real-time BCI operation.