A decoder for a brain-computer interface (a system that turns recorded brain activity into commands) is usually a single-case result. It is trained on one person, works for that person, and tells us little about the next user. Clinical medicine knows this problem well: a single case report shows that something is possible, but it cannot show that it is a pattern. We propose to make that boundary measurable for imagined-movement brain-computer interfaces. We train a decoder on each person, test it on every other person, and ask what predicts the success of each transfer. Our candidate explanation comes from embodied cognition, the finding that mental spatial manipulation reuses the body's motor machinery. The specific predictor is how fast a person can mentally shift their own viewpoint onto a human body, as opposed to how fast they rotate an abstract object.
The claim fits in two sentences. Transfer of a decoder from person A to person B depends on how similar A and B are in egocentric, body-based spatial processing, not only on how "good" each is at brain-computer interface control. If that holds, a five-minute behavioral test given before any EEG is recorded can pick which existing users' data should train a new user's decoder. It also tells a lab which of its single-case successes are likely to transfer, and to whom.
Core Claim
Earlier work asks a per-person question: can a person's traits predict their own decoding accuracy? We ask a relational question: does the distance between two people's embodied spatial profiles predict the accuracy of a decoder trained on one and tested on the other, once we account for how good a source and a target each are on their own?
The profile has three parts. The first is the slope from egocentric body rotation tasks, which measures the speed of mentally rotating oneself into a displayed figure's pose. The second is the corresponding object-based slope. The third is the non-rotational intercept, which captures perception and response time. We model the cross-subject transfer score as
$$T_{ij} = \mu + a_i + b_j + \beta_p\, d_p(i,j) + \beta_e\, d_e(i,j) + \varepsilon_{ij}.$$
Here $T_{ij}$ is Cohen's kappa when a decoder trained on subject $i$ is tested on subject $j$. The term $a_i$ is a random effect for how good a source $i$ is, and $b_j$ is a random effect for how decodable target $j$ is. The term $d_p(i,j)$ is the distance between the two subjects' embodied spatial profiles. The term $d_e(i,j)$ is the distance between their resting-state EEG covariance matrices, a purely neural similarity measure. The final term $\varepsilon_{ij}$ is residual error.
The prediction is that $\beta_p$ is negative (closer profiles transfer better) and stays negative with $d_e$ in the model. It should also be carried by the egocentric slope rather than the abstract-cube slope.
What Prior Work Establishes and Leaves Open
Three embodied-cognition studies motivate the choice of predictor. They show that bodily training reshapes mental rotation in a way that depends on the stimulus. Selective Effects of Sport Expertise on the Stages of Mental Rotation Tasks With Object-Based and Egocentric Transformations splits reaction time into slope and intercept and finds that elite divers rotate faster only in the egocentric body task ($t(42)=2.389$, $p<.05$), with no slope advantage on abstract cubes. Axial rotation affects the cognitive characteristics of spatial ability shows that sports with multi-axis body rotation confer a specific advantage on human-body rotations. It also finds that the male advantage appears only in low-spatial athletes, and that rotation difficulty depends on the axis because of limb "sheltering" behind the torso. The Time Course of Event-Related Brain Potentials in Athletes' Mental Rotation With Different Spatial Transformations finds that the size of a rotation-related EEG wave tracks both rotation speed and years of training.
Together these papers suggest that egocentric body processing differs between people in a way tied to motor history, and that it has neural correlates. That is the kind of variable that could explain why one person's motor-imagery EEG resembles another's. None of the three papers touches decoding. All of them use mostly adolescent cohorts, which we deliberately avoid.
Benaroch et al. both motivate and constrain the proposal. Are users' traits informative enough to predict/explain their mental-imagery based BCI performances? found that a paper-and-pencil Mental Rotation Test was the strongest predictor of motor-imagery performance within one cohort ($r=0.6$, mean absolute error 3.03%). The same models fell to chance when the three studies were pooled. We take that collapse seriously. It could mean that trait–performance links are fragile across contexts, in which case our relational version might fail the same way. We test this directly with a context-shift session. We differ in two ways: the predictor is a relational distance rather than a per-person score, and the spatial measure is embodied and egocentric rather than abstract.
Two decoding papers supply the methods and the bar to beat. EEG-DG: A Multi-Source Domain Generalization Framework for Motor Imagery EEG Classification aligns feature distributions across source subjects and weights sources by their similarity to the target. It reaches 81.79% accuracy (kappa 0.7572) on BCI Competition IV-2a without any target calibration. Its authors note that source domains are divided heuristically. Toward Real-World BCI: CCSPNet, A Compact Subject-Independent Motor Imagery Framework shows that on 54 subjects, a subject-independent model (74.28%) nearly matches a per-subject model (74.41%). Both treat between-subject variation as statistical noise to be aligned away. We ask whether some of that variation has a measurable cognitive source.
A Personalized and Adaptable User Interface for a Speech and Cursor Brain-Computer Interface sets the single-case frame. It reports a 22-month, 4,000-hour deployment with one participant, and the authors state that the co-designed system may not generalize. The clinical case literature makes the same point from outside neuroscience. Laparoscopic pancreaticoduodenectomy in A-92-older Chinese patient for cancer of head of the pancreas credits a nonagenarian's surgical success to that patient's exceptional baseline reserve. That is exactly the kind of hidden individual covariate we try to measure rather than guess at.
Research Program
We recruit about 60 adults in four groups chosen to spread the egocentric profile widely. The groups are closed high-spatial athletes (gymnasts, divers, aerial skiers), open high-spatial athletes (basketball and soccer players), closed low-spatial athletes (runners), and non-athletes. This follows the sport categories of Sport training of axial rotation affects spatial ability: Evidence from behavior and fNIRS, but with adults.
Each participant first completes a profiling battery. It includes the object-based cube, object-based body, and egocentric body tasks from the diver study at several angles, plus left–right, up–down, and front–back axis rotations. From these we fit slopes and intercepts, and the battery also yields the axis-anisotropy pattern. We record resting EEG before any imagery task.
Participants then complete two sessions of four-class motor imagery (left hand, right hand, feet, tongue) with the 22-channel montage of BCI Competition IV-2a. Matching the montage lets us validate our pipeline against published numbers. Session two changes the feedback display and instructions, deliberately recreating the kind of context difference that broke Benaroch et al.'s pooled models.
We fit each per-subject decoder with Filter Bank Common Spatial Patterns (FBCSP) to fill the full 60×60 transfer matrix. We then estimate the crossed random-effects model above. Finally, we turn the result into a source-selection rule for new users: rank existing subjects by profile distance to the newcomer, then train EEG-DG on the nearest $k$.
Baselines
The do-nothing reference point has two parts: chance kappa of zero, and a decoder trained on a random $k$ source subjects. The random-$k$ decoder isolates whether choosing sources matters at all.
The strongest existing competitor is EEG-DG trained on all available sources with its built-in similarity weighting, using the authors' released code. It is the strongest because it already solves zero-calibration transfer and already weights sources by similarity. We first reproduce it on IV-2a against its reported 81.79% accuracy and kappa 0.7572, to confirm our setup is faithful. We then run it on our cohort with leave-one-subject-out evaluation. It gets the same subjects, the same preprocessing, and the same hyperparameter search budget as our profile-selected variant, which differs only in which sources enter training.
CCSPNet serves as a second subject-independent reference, since it reaches 74.28% subject-independent accuracy on the 54-subject database. Per-subject FBCSP with a short calibration run is the subject-dependent reference; it reports kappa 0.569 on IV-2a in Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b.
Four ablations isolate the key ingredient by keeping the selection rule fixed and swapping the distance used to rank sources:
- the Benaroch-style abstract cube score alone;
- sport-category labels alone;
- resting-EEG covariance distance alone;
- the egocentric body profile.
The proposal beats EEG-DG if profile-selected training raises mean leave-one-out kappa over all-source EEG-DG by a margin we pre-register, tested with a paired test across targets, and especially if it helps the bottom quartile of targets. If it matches EEG-DG but with far fewer sources, that counts only as an efficiency result.
Why not just…?
Why not just run EEG-DG on everyone and let alignment handle the differences? This is a serious threat. EEG-DG already weights sources and reports strong zero-calibration accuracy. If profile-selected training does not beat all-source EEG-DG on our cohort, we will say so and drop the decoding-performance claim. The explanatory claim could still stand: a significant $\beta_p$ in the transfer matrix would say something about which single-case decoders generalize, even if alignment already captures it implicitly. The two experiments are reported separately so that one cannot hide the other.
Why not just use neural similarity, such as resting EEG distance, to pick sources? This is the most dangerous alternative, because neural similarity sits closer to the outcome than any behavioral test. The mixed model settles it by including $d_e$ next to $d_p$. If the profile term vanishes once EEG distance is in the model, the profile is merely a cheap stand-in for EEG. The claim then narrows to pre-recording triage, meaning choosing sources and flagging poor transfer before a headset is fitted. We would still report that honestly as the result.
Why not just use the standard Mental Rotation Test, or the athlete's sport category? Benaroch et al. found the abstract test to be the top single-cohort predictor. It would be simpler, and so would a sport label. The diver study, however, found no rotation-speed difference on abstract cubes, only on egocentric body rotation. That predicts the cube-distance ablation should carry little transfer information. Sport category is coarse and cannot separate, say, a slow gymnast from a fast one. Both ablations are run directly. If either matches the full profile, we adopt the simpler measure.
Why not just calibrate each new user? Per-subject calibration works, but CCSPNet's near-parity between subject-independent and subject-dependent models, together with the fatigue it cites, makes zero-calibration the relevant target. Calibration remains in the baselines as the ceiling for what a user-specific decoder buys.
Decisive Checks
The central test is whether $\beta_p$ is reliably negative in the crossed model, judged against 10,000 permutations that shuffle profiles across subjects while keeping the transfer matrix fixed. It must survive the inclusion of $d_e$, and it must be driven mainly by the egocentric slope rather than the cube slope. The second test is the context shift: we refit the model on session-two data. If the profile–transfer relation collapses the way Benaroch et al.'s pooled trait models did, the relational framing offers no protection against context fragility, and we would report that as a negative finding. The third test is held-out prospective selection: the last ten recruited subjects are profiled, get sources chosen by the frozen rule, and are compared against random-$k$ and all-source EEG-DG. No reported result was found on Lacuna for cross-subject transfer matrices predicted from embodied spatial profiles, so none of these outcomes is prejudged by existing numbers.
References
- Selective Effects of Sport Expertise on the Stages of Mental Rotation Tasks With Object-Based and Egocentric Transformations
- Axial rotation affects the cognitive characteristics of spatial ability
- The Time Course of Event-Related Brain Potentials in Athletes’ Mental Rotation With Different Spatial Transformations
- Are users' traits informative enough to predict/explain their mental-imagery based BCI performances?
- EEG-DG: A Multi-Source Domain Generalization Framework for Motor Imagery EEG Classification
- Toward Real-World BCI: CCSPNet, A Compact Subject-Independent Motor Imagery Framework
- A Personalized and Adaptable User Interface for a Speech and Cursor Brain-Computer Interface
- Laparoscopic pancreaticoduodenectomy in A-92-older Chinese patient for cancer of head of the pancreas
- Sport training of axial rotation affects spatial ability: Evidence from behavior and fNIRS
- Filter Bank Common Spatial Pattern Algorithm on BCI Competition IV Datasets 2a and 2b