Subject-Specific Personalisation
The Problem
A Brain FM pre-trained on population-level data produces a general encoder that captures the average relationship between physiological signals and cognitive states. When deployed for a specific individual, performance is degraded by that individual's unique signal characteristics.
The degree of degradation can be severe: the correlation between population-level EEG workload markers and individual-level NASA-TLX scores varies widely, with some individuals showing strong correspondence and others showing nearly no detectable relationship using population-level models.
Personalisation addresses this by adapting the model to the specific individual using a small amount of subject-specific data - ideally from a calibration session before or alongside deployment.
Design Constraints
Any personalisation approach must satisfy three operational constraints:
Minimal calibration data. An operator cannot be asked to perform a 2-hour EEG experiment before their shift. Personalisation must work with as little as 5–15 minutes of labelled calibration data, or even with unlabelled data only.
No catastrophic forgetting. Personalising the model for one individual must not degrade performance on other individuals. The base model should remain intact; only individual-specific components should change.
Privacy-preserving. The personalisation process must not allow one individual's data to leak into another individual's adapted model. Individual-specific parameters should be stored and used in an isolated, access-controlled manner.
Approaches
Unsupervised Domain Adaptation (PULSE)
Wang et al. (2025) - Expert Systems with Applications
PULSE frames personalisation as unsupervised domain adaptation: the population-level pre-trained model is the source domain, and the new individual is the target domain. Adaptation uses unlabelled target-domain data - no cognitive state labels are required for the calibration session.
The domain adaptation aligns the individual's signal distribution with the population-level representation learned during pre-training, without using any labels. This is the most practical approach for operational deployment where labelled calibration data is unavailable.
Efficient Personalised Adaptation (LoRA / Adapters)
Wu et al. (ICML 2025) - Efficient Personalised Adaptation for Physiological Signal FM
Insert lightweight adapter modules (or apply Low-Rank Adaptation / LoRA) on top of the frozen foundation model. The base model weights are never changed; only a small set of individual-specific parameters (typically < 1% of total parameters) are updated during personalisation.
In LoRA, each weight matrix \(W\) in the model is decomposed as:
where \(A \in \mathbb{R}^{d \times r}\) and \(B \in \mathbb{R}^{r \times d}\) are low-rank matrices with \(r \ll d\). Only \(A\) and \(B\) are trained during personalisation; \(W\) remains frozen.
Advantages:
- Individual-specific parameters are tiny - easy to store one set per operator.
- The base model is unchanged - no forgetting, no shared interference between individuals.
- Works with very small labelled calibration datasets due to the low-rank constraint.
Self-Adaptive Learning
An online adaptation mechanism where individual-specific adapter layers are continuously updated from incoming data during deployment, using pseudo-labels derived from the model's own confident predictions. This allows the model to track slow changes in the individual's physiological state over time (e.g. gradual fatigue accumulation) without explicit labels.
Meta-Learning (MAML-style)
Train the base model with a meta-learning objective (e.g. Model-Agnostic Meta-Learning, MAML) such that the base model can be quickly adapted to any new individual with just a few gradient steps. The meta-training objective explicitly optimises for few-shot adaptation speed.
Advantage: Maximally data-efficient at personalisation time.
Disadvantage: More complex training procedure; requires diverse individual data during meta-training.
Comparison
| Approach | Labelled Calibration Data | Compute at Personalisation | Inference Overhead |
|---|---|---|---|
| Unsupervised DA (PULSE) | Not required | Moderate (DA optimisation) | None |
| LoRA / Adapters | Small labelled set needed | Low (few gradient steps) | Negligible |
| Self-Adaptive Learning | Not required | Continuous (online) | Negligible |
| Meta-Learning (MAML) | Small labelled set needed | Very low (few gradient steps) | None |
Relationship to Differential Privacy
Differential privacy during pre-training prevents the base model from encoding individual identity information. Personalisation then adds back individual-specific knowledge in a controlled, structured way. These two mechanisms work together to achieve both generalisation and per-individual accuracy simultaneously.