Brain-signal decoders — whether driven by scalp electroencephalography (EEG) or by intracranial recordings — are today reliable mainly for the individual they were trained on: between-person differences in neurophysiology and in electrode placement break decoders transferred to a new person, so practical brain-computer interfaces (BCIs) typically require per-user calibration data. Whether decoding can generalize across individuals without such recalibration — calibration-free inter-individual decoding — remains an open research problem. As the MIBRAIN framework for intracranial neural decoding (arXiv:2506.12055) states in its abstract, inter-individual heterogeneity has 'constrained current approaches to neural decoding within single individuals, rendering interindividual neural decoding elusive'. The problem remains open: transfer-learning and foundation-model approaches achieve partial cross-subject gains but no demonstrated calibration-free parity with within-individual decoders, and 2025-2026 work shows electrode-montage mismatch and subject-identity confounds remain unresolved obstacles that even complicate evaluating progress on the problem.

Background

A brain-computer interface (BCI) reads electrical brain activity and decodes it into commands or predictions, for example for motor rehabilitation, communication prostheses, or speech decoding. Two recording settings dominate. In scalp electroencephalography (EEG), electrodes on the scalp surface record the summed activity of cortical neurons at low spatial resolution and low signal-to-noise ratio; a common experimental paradigm is motor-imagery (MI) decoding, in which the user imagines moving a limb and the decoder infers which limb from sensorimotor rhythms. In intracranial recording — electrocorticography (ECoG) or stereo-electroencephalography (sEEG), in which electrode strips or depth electrodes are implanted inside the skull — signals are far cleaner, but electrode positions are dictated by clinical need (for example, localizing epilepsy) and are therefore unique to each patient.

Decoding means training a statistical model (classical machine-learning pipelines, or more recently deep networks) to map segments of recorded activity to labels such as the imagined movement, the perceived stimulus, or the articulated phoneme. Standard practice is within-subject decoding: the model is trained and evaluated on data from the same person, which requires collecting calibration data from that person first. When a model trained on one person is applied to another, accuracy drops sharply, for two structurally distinct reasons. First, individuals differ physiologically: the same mental state produces different signal patterns across people. Second, the sensor geometry differs: scalp montages vary across systems and sessions, and implanted electrode sets never coincide across patients, so the input channels of one individual's model do not even correspond to those of another. In machine-learning terms, inter-subject variability induces a covariate shift — the input distribution changes between training (source subjects) and deployment (target subject) — which is why classical pipelines 'faced challenges in generalizing across subjects and sessions', as a 2025 review of BCI decoding techniques puts it (Zhang, Theoretical and Natural Science, 2025), and why inter-subject variability is repeatedly named a principal obstacle in EEG-based BCI (for example, Zhao et al., Digital Signal Processing, 2026).

A first line of attack is transfer learning and domain adaptation: reusing or aligning models across subjects, surveyed by Jayaram et al. (IEEE Computational Intelligence Magazine, 2016) and by Saha and Baumert (Frontiers in Computational Neuroscience, 2020), who frame intra- and inter-subject variability as covariate shift impeding parameter transferability. These methods mitigate but do not eliminate the gap. A second, current line is pretraining biosignal foundation models — large neural networks first trained by self-supervised learning (training on unlabeled data with artificially masked or corrupted inputs that the network must reconstruct, requiring no labels) on pooled multi-subject corpora, then fine-tuned to downstream tasks. Examples include Brant for intracranial signals (NeurIPS, 2023), BIOT for heterogeneous biosignals with mismatched channel sets (arXiv:2305.10351), LaBraM for cross-dataset EEG (arXiv:2405.18765), and EEGPT (NeurIPS, 2024), whose abstract still names 'high inter-subject variability, and channel mismatch' as the core unresolved challenges.

The most direct recent assault on the intracranial side is MIBRAIN (Multi-individual Brain Region-Aggregated Network; Wu et al., arXiv:2506.12055, 2025), which integrates sEEG recordings across 11 patients with different electrode implantations into one functional brain-network model using self-supervised region prototypes, and decodes Mandarin phoneme articulation — including silent articulation — online and offline. Under leave-one-subject-out cross-validation (train on all participants except one, test on the held-out individual), decoding is significantly above chance but, by the authors' own report, lower than that of models trained on the complete dataset including the test subject's own data. Moreover, decoding a new subject currently runs through an ensemble: each existing participant contributes a participant-specific neuro-embedding encoder, and the final prediction is a majority vote across these per-subject models. The authors' limitations section states that these 'participant-specific embeddings ... necessitate ensemble-based decoding for predictions on unseen subjects' and that developing a unified encoding strategy yielding a single subject-agnostic model remains future work. Two 2026 diagnostic studies show the obstacles persist even where foundation models appear to succeed: a systematic channel-adaptation benchmark (Kokate et al., arXiv:2604.23091) finds that pooling data across heterogeneous electrode montages — a prerequisite for scaling — remains architecture-dependent, with severe negative-transfer cases (adaptation that leaves the model worse than no adaptation), and the 'Identity Trap' audit (Lin et al., arXiv:2606.06647) shows that subject-identity features dominate the representations of EEG foundation models, so high accuracy under subject-disjoint evaluation can reflect who the subject is rather than what the decoded neural process is. In short: cross-individual decoding above chance is demonstrated in several settings, but a single calibration-free decoder that approaches within-individual performance — the explicitly stated goal of this literature — does not yet exist.

Problem Statement

Can a single, subject-agnostic brain-signal decoder — one model trained on data pooled from multiple individuals, with no per-individual calibration or adaptation at deployment — decode the neural signals of an individual never seen during training at a level approaching that of a decoder trained on that individual's own data? The question spans the two principal recording settings in which it has been posed — scalp EEG (for example, motor-imagery BCI) and intracranial recordings (ECoG/sEEG, where electrode placement differs per patient); it is posed across both settings, and a treatment of one setting alone resolves the question only for that setting. Here decoding performance on held-out individuals is assessed under a subject-disjoint evaluation protocol and compared with the matched within-individual baseline on the same data and task, so that the gap to within-individual decoding is measured rather than asserted, with a subject-identity control that distinguishes decoding of the target neural process from decoding of subject-identity features. The open questions are how close calibration-free inter-individual decoding can come to within-individual decoding and what property of the trained model (unified encoding across heterogeneous electrode layouts, subject-invariant representations) makes that level achievable.

References

  1. Wu, Di; Bu, Linghao; Jia, Yifei; Cao, Lu; Li, Siyuan; Chen, Siyu; Zhou, Yueqian; Fan, Sheng; Ren, Wenjie; Wu, Dengchang; Wang, Kang; Zhang, Yue; Ma, Yuehui; Yang, Jie; Sawan, Mohamad (2025). 'Towards Unified Neural Decoding with Brain Functional Network Modeling' (MIBRAIN). arXiv:2506.12055, DOI 10.48550/arxiv.2506.12055. https://arxiv.org/abs/2506.12055
  2. Zhao, Rongyue; Li, Ruiqing; Li, Yihua; Kong, Xiangzeng (2026). 'A multi-scale statistical pooling neural network with channel attention for motor imagery brain-computer interface'. Digital Signal Processing 182:106308, DOI 10.1016/j.dsp.2026.106308. https://doi.org/10.1016/j.dsp.2026.106308
  3. Zhang, Yiyang (2025). 'The Evolution of Neural Signal Decoding Techniques in Brain-Computer Interfaces from Traditional Methods to Deep Learning'. Theoretical and Natural Science 133:216-223 (HSET proceedings), DOI 10.54254/2753-8818/2025.au25856. https://doi.org/10.54254/2753-8818/2025.au25856
  4. Jayaram, Vinay; Alamgir, Morteza; Altun, Yasemin; Scholkopf, Bernhard; Grosse-Wentrup, Moritz (2016). 'Transfer Learning in Brain-Computer Interfaces'. IEEE Computational Intelligence Magazine 11(1):20-31, DOI 10.1109/MCI.2015.2501545. https://doi.org/10.1109/MCI.2015.2501545
  5. Saha, Simanto; Baumert, Mathias (2020). 'Intra- and Inter-subject Variability in EEG-Based Sensorimotor Brain Computer Interface: A Review'. Frontiers in Computational Neuroscience 13:87, DOI 10.3389/fncom.2019.00087. https://doi.org/10.3389/fncom.2019.00087
  6. Zhang, Daoze; Yuan, Zhizhang; Yang, Yang; Chen, Junru; Wang, Jingjing; Li, Yafeng (2023). 'Brant: Foundation Model for Intracranial Neural Signal'. Advances in Neural Information Processing Systems 36:26304-26321, DOI 10.52202/075280-1144. https://doi.org/10.52202/075280-1144
  7. Yang, Chaoqi; Westover, M. Brandon; Sun, Jimeng (2023). 'BIOT: Cross-data Biosignal Learning in the Wild'. arXiv:2305.10351, DOI 10.48550/arXiv.2305.10351. https://arxiv.org/abs/2305.10351
  8. Jiang, Wei-Bang; Zhao, Li-Ming; Lu, Bao-Liang (2024). 'Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI' (LaBraM). arXiv:2405.18765, DOI 10.48550/arXiv.2405.18765. https://arxiv.org/abs/2405.18765
  9. Wang, Guangyu; Liu, Wenchao; He, Yuhong; Xu, Cong; Ma, Lin; Li, Haifeng (2024). 'EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals'. Advances in Neural Information Processing Systems 37:39249-39280, DOI 10.52202/079017-1239. https://doi.org/10.52202/079017-1239
  10. Kokate, Kuntal; Aristimunha, Bruno; Truong, Dung; Delorme, Arnaud (2026). 'Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes'. arXiv:2604.23091, DOI 10.48550/arXiv.2604.23091. https://arxiv.org/abs/2604.23091
  11. Lin, Jun-You; Wu, Ying Choon; Jung, Tzyy-Ping (2026). 'The Identity Trap in EEG Foundation Models: A Diagnostic Audit'. arXiv:2606.06647, DOI 10.48550/arXiv.2606.06647. https://arxiv.org/abs/2606.06647