Inter-Subject Variability
Inter-subject variability is one of the fundamental challenges in EEG-based cognitive monitoring. Two individuals performing the same task at the same workload level can produce dramatically different EEG signals, for reasons that are both physiological and recording-related.
Sources of Variability
Anatomical Differences
Skull thickness, cortical folding patterns, and the geometry of the underlying neural sources all differ between individuals. These anatomical differences affect the way neural signals propagate to scalp electrodes, producing systematic individual-specific signal characteristics that are unrelated to cognitive state.
Cognitive Strategy Differences
Different individuals employ different cognitive strategies to accomplish the same task. Under the same objective workload conditions, one person might rely more heavily on verbal working memory (producing strong frontal theta), while another uses visuospatial strategies (producing more occipital-parietal activity). These strategy differences produce signal differences that are not noise - they reflect genuine individual-level processing differences.
Baseline Arousal and Trait Differences
Individual differences in baseline arousal, trait anxiety, and cognitive capacity all modulate the physiological response to a given workload level. A highly anxious individual may show elevated stress markers under conditions that produce no measurable stress response in a relaxed individual.
Session-to-Session Variability
Even within the same individual, EEG recordings vary between sessions: electrode placement differs slightly, skin conductance changes, and head movements alter electrode-scalp contact. These day-to-day variations add another layer of variability on top of between-subject differences.
Two Complementary Problems
Inter-subject variability creates two challenges that must be addressed with opposite goals:
| Problem | Goal | Approach |
|---|---|---|
| Memorisation during pre-training | Remove individual identity from representations | Differential Privacy |
| Poor generalisation to new individuals | Incorporate individual-specific adaptation | Subject Personalisation |
These are not contradictory: differential privacy prevents the base model from encoding individual identity as a "hidden variable," while personalisation then adds back individual-specific knowledge in a structured, controlled way at deployment time.
Why Both Matter
A model that memorises subject identity during pre-training will perform well on individuals seen during training but fail on new individuals - it has learned "this is person X's EEG pattern under high workload" rather than "high workload EEG looks like this." Differential privacy forces the model to learn population-level cognitive signal features.
A model that lacks any personalisation will produce predictions degraded by inter-subject variability - it can only produce a population-average response. A brief calibration session (e.g. 10 minutes of the target individual performing a known cognitive task) can dramatically improve per-individual performance.
The combination - differentially private base model + lightweight personalisation adapters - achieves both goals simultaneously.