Source-Free Domain Adaptation for SSVEP-based Brain-Computer Interfaces

TL;DR

This work tackles calibration burden in SSVEP-based BCIs by introducing a source-free unsupervised domain adaptation framework that transfers a DNN trained on source participants to a new user using only unlabeled target data. A novel two-term loss, combining self-adaptation with pseudo-labels ($\mathcal{L}_{sl}$) and a local-regularity term ($\mathcal{L}_{ll}$) that enforces label consistency among neighboring target samples, drives adaptation, with dynamic lambda selection via silhouette clustering. Evaluated on benchmark_Dataset and BETA, the approach achieves state-of-the-art ITRs of $201.15$ and $145.02$ bits/min, substantially improving over pre-trained baselines and competing methods, while preserving user comfort by eliminating calibration. The method is architecture-agnostic, privacy-preserving (no source data retention), and capable of continual improvement as unlabeled target data accumulates, potentially accelerating real-world adoption of SSVEP-BCI spellers.

Abstract

Objective: SSVEP-based BCI spellers assist individuals experiencing speech difficulties by enabling them to communicate at a fast rate. However, achieving a high information transfer rate (ITR) in most prominent methods requires an extensive calibration period before using the system, leading to discomfort for new users. We address this issue by proposing a novel method that adapts a powerful deep neural network (DNN) pre-trained on data from source domains (data from former users or participants of previous experiments), to the new user (target domain) using only unlabeled target data. Approach: Our method adapts the pre-trained DNN to the new user by minimizing our proposed custom loss function composed of self-adaptation and local-regularity terms. The self-adaptation term uses the pseudo-label strategy, while the novel local-regularity term exploits the data structure and forces the DNN to assign similar labels to adjacent instances. Main results: Our method achieves excellent ITRs of 201.15 bits/min and 145.02 bits/min on the benchmark and BETA datasets, respectively, and outperforms the state-of-the-art alternatives. Our code is available at https://github.com/osmanberke/SFDA-SSVEP-BCI Significance: The proposed method prioritizes user comfort by removing the burden of calibration while maintaining an excellent character identification accuracy and ITR. Because of these attributes, our approach could significantly accelerate the adoption of BCI systems into everyday life.

Figures (6)

• Figure 1: The outline of our source-free domain adaptation method and a representative illustration of changes in the DNN's decision boundaries across two consecutive iterations of adaptation are shown. The left panel depicts the state before adaptation, and the right panel shows the state after adaptation. Dashed lines represent decision boundaries, while different shapes denote different classes. The adaptation starts with a DNN pre-trained using data from source domains, and it is carried out by minimizing the loss function $\mathcal{L}_{{total}}$ that we also introduce, which consists of two main terms: self-adaptation $\mathcal{L}_{sl}$ and local-regularity $\mathcal{L}_{{ll}}$. The self-adaptation component utilizes the pseudo-labeling approach, whereas the local-regularity component leverages the inherent structure of the data, forcing the adapted DNN to give similar labels to adjacent instances. Our method is compatible with any DNN architecture and particularly achieves ITR results of 201.15 bits/min and 145.02 bits/min on the benchmark benchmark_Dataset and BETA beta datasets, respectively, when utilizing the DNN architecture from ournetwork. These results show that our proposed method provides significant ITR improvements over the state-of-the-art alternatives.
• Figure 2: The mean classification accuracy on the left and the mean information transfer rate (ITR) on the right are presented across all users in the datasets, together with the standard errors indicated by the bars. The asterisks indicate the results of statistical significance analyses. Paired $t$-tests are applied, and significance levels are reported based on the least significant difference criterion (*$p < \frac{0.05}{4}$, **$p < \frac{0.05}{20}$). The notation 'ns' denotes results that are not statistically significant. The marker $^{\downarrow}$ denotes effects favoring the comparator. (a) The results for the benchmark dataset benchmark_Dataset with 35 users. (b) The results for the BETA dataset beta with 70 users.
• Figure 3: Our $\lambda$ selection is illustrated with the benchmark dataset benchmark_Dataset (a) and the BETA dataset beta (b) for three different signal lengths: $0.2$, $0.6$, and $1$ seconds. The graphs represent the distribution of users into three specific performance intervals of accuracy comparisons between the selected $\lambda$ and the best $\lambda$.
• Figure 4: The mean classification accuracy results are presented across all users in the datasets, together with the standard errors indicated by the bars. These results are for after fine-tuning with calibration data, adaptation using our approach, and the pre-trained versions, using two different DNN architectures. (a) The results for the benchmark dataset benchmark_Dataset. (b) The results for the BETA dataset beta.
• Figure A1: For each signal length $T\in\{0.2,0.4,0.6,0.8,1.0\}\,$, side-by-side boxplots show per-participant performance for each method. Boxes depict the interquartile range (IQR; 25th–75th percentiles) with the median line; whiskers extend to $1.5\times$ IQR; jittered points indicate individual participants. (a) Mean accuracy. (b) Mean information transfer rate (ITR). Results are shown for the benchmark dataset.