Skip to content

Explainability - Modified LASTS Framework

Why Explainability Matters

A cognitive monitoring system that produces an alert - "this individual is approaching cognitive overload" - is only valuable if a supervisory entity can understand why the system reached that conclusion.

Without explainability:

  • Predictions are not trusted. Individuals who cannot understand the basis for an alert will dismiss it, particularly when it contradicts their own intuition.
  • Errors cannot be diagnosed. False positives and false negatives are inevitable; without transparency, it is impossible to understand why they occurred or how to fix the system.
  • Scientific value is limited. The physiological features that drive predictions are valuable scientific findings; an opaque model cannot contribute to understanding of cognitive neuroscience.

The Original LASTS Framework

LASTS is an XAI framework designed for time series classifiers. It generates explanations by:

  1. Building a surrogate model - a simpler, interpretable model (e.g. a shallow decision tree) that locally approximates the complex classifier's behaviour around a specific prediction.
  2. Identifying counterfactuals - minimal changes to the input signal that would flip the prediction (e.g. "if frontal theta power had been 20% lower, the system would have predicted normal workload").

Both explanation types are complementary: the surrogate tells you what features matter; the counterfactual tells you what would need to change.

Surrogate Model Explanations

A surrogate model is trained to mimic the foundation model's predictions in a local neighbourhood around the input of interest. The neighbourhood is defined by generating slightly perturbed versions of the input window and querying the foundation model for their predictions.

The surrogate is constrained to be interpretable - typically a shallow decision tree (depth ≤ 4) or a linear model over extracted EEG features (band power, HRV features, gaze statistics).

Example output:

Prediction: HIGH WORKLOAD (confidence: 87%)

Surrogate explanation (decision tree depth 2):
  IF frontal_theta_power (F3, F4) > 15.3 µV²
  AND alpha_suppression_occipital > 0.42
  THEN → HIGH WORKLOAD (87% of surrogate cases)

This explanation is:

  • Actionable - identifies specific EEG features that drove the prediction.
  • Scientific - frontal theta and alpha suppression are known workload markers, providing a sanity check.
  • Accessible - a decision tree is interpretable by researchers and developers without ML expertise.

Counterfactual Explanations

A counterfactual explanation finds the closest possible input signal to the current window that would have received a different prediction. It answers: "What would need to change for the prediction to be different?"

Counterfactuals are found by minimising the change to the input (e.g. in L1 or L2 norm over extracted features) subject to the constraint that the BFM predicts a different class.

Example output:

Counterfactual explanation:
  Current state  → HIGH WORKLOAD (confidence: 87%)

  If frontal_theta_power decreased by 3.1 µV² (-20%)
  → prediction would change to MODERATE WORKLOAD

  Minimum change required: 3.1 µV² in frontal theta

Counterfactuals are often more actionable than feature importance for operational personnel because they describe a concrete path to a different outcome - even if that path (reducing the operator's frontal theta power) is not directly controllable, it tells the supervisor what the physiological threshold is.

Implementation Notes

Feature Extraction for Surrogate Models

The surrogate model operates over extracted physiological features rather than raw time series, to ensure interpretability:

EEG features:

  • Band power per electrode: delta, theta, alpha, beta, gamma (5 bands × C electrodes)
  • Frontal asymmetry (F4 − F3 alpha power)
  • Frontal midline theta (Fz theta power)
  • Alpha suppression index (ratio of task alpha to baseline alpha)

HRV features (from ECG/PPG):

  • SDNN, RMSSD, pNN50
  • LF power, HF power, LF/HF ratio

Eye-tracking features:

  • Mean fixation duration, scan path entropy, AOI coverage
  • Mean pupil diameter, max TEPR

Counterfactual Search Algorithm

Counterfactuals are found using a gradient-based or genetic optimisation procedure over the feature space:

\[\mathbf{x}^* = \arg\min_{\mathbf{x}'} \left[ \lambda_1 \cdot d(\mathbf{x}, \mathbf{x}') + \lambda_2 \cdot \mathcal{L}(\hat{y}(\mathbf{x}'), y_{\text{target}}) \right]\]

where \(d(\mathbf{x}, \mathbf{x}')\) penalises large changes from the original features, \(\mathcal{L}\) is a loss function that rewards predicting the target class, and \(\lambda_1, \lambda_2\) balance proximity against target achievement.

Relationship to Safety Applications

All cognitive state predictions feeding into safety applications (overload monitoring, error prediction) are required to produce LASTS explanations alongside the prediction. The explanation is logged with the prediction timestamp and the operator identifier, enabling post-shift review of all alerts and their physiological basis.