Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spatial Adaptation Layer: Interpretable Domain Adaptation For Biosignal Sensor Array Applications

Joao Pereira, Michael Alummoottil, Dimitrios Halatsis, Dario Farina

TL;DR

The paper tackles electrode shift causing intersession performance drops in biosignal sensor arrays by introducing Spatial Adaptation Layer ($A_S$) and Learnable Baseline Normalization (LBN). SAL provides a 7-parameter, interpretable affine transform decomposed into translation, rotation, scaling, and shear, implemented via a differentiable sampling operator to map new-session data back to the original spatial frame, while LBN subtracts a learnable per-channel baseline. Across two HD-sEMG datasets with regular and irregular grids, SAL+LBN outperforms standard fine-tuning on regular grids and remains competitive on irregular grids, using orders of magnitude fewer parameters and yielding interpretable, physically meaningful adaptations. Ablation confirms that circumferential translations largely drive improvements, highlighting SAL’s potential for practical, interpretable domain adaptation in biosignal applications.

Abstract

Machine learning offers promising methods for processing signals recorded with wearable devices such as surface electromyography (sEMG) and electroencephalography (EEG). However, in these applications, despite high within-session performance, intersession performance is hindered by electrode shift, a known issue across modalities. Existing solutions often require large and expensive datasets and/or lack robustness and interpretability. Thus, we propose the Spatial Adaptation Layer (SAL), which can be applied to any biosignal array model and learns a parametrized affine transformation at the input between two recording sessions. We also introduce learnable baseline normalization (LBN) to reduce baseline fluctuations. Tested on two HD-sEMG gesture recognition datasets, SAL and LBN outperformed standard fine-tuning on regular arrays, achieving competitive performance even with a logistic regressor, with orders of magnitude less, physically interpretable parameters. Our ablation study showed that forearm circumferential translations account for the majority of performance improvements.

Spatial Adaptation Layer: Interpretable Domain Adaptation For Biosignal Sensor Array Applications

TL;DR

The paper tackles electrode shift causing intersession performance drops in biosignal sensor arrays by introducing Spatial Adaptation Layer () and Learnable Baseline Normalization (LBN). SAL provides a 7-parameter, interpretable affine transform decomposed into translation, rotation, scaling, and shear, implemented via a differentiable sampling operator to map new-session data back to the original spatial frame, while LBN subtracts a learnable per-channel baseline. Across two HD-sEMG datasets with regular and irregular grids, SAL+LBN outperforms standard fine-tuning on regular grids and remains competitive on irregular grids, using orders of magnitude fewer parameters and yielding interpretable, physically meaningful adaptations. Ablation confirms that circumferential translations largely drive improvements, highlighting SAL’s potential for practical, interpretable domain adaptation in biosignal applications.

Abstract

Machine learning offers promising methods for processing signals recorded with wearable devices such as surface electromyography (sEMG) and electroencephalography (EEG). However, in these applications, despite high within-session performance, intersession performance is hindered by electrode shift, a known issue across modalities. Existing solutions often require large and expensive datasets and/or lack robustness and interpretability. Thus, we propose the Spatial Adaptation Layer (SAL), which can be applied to any biosignal array model and learns a parametrized affine transformation at the input between two recording sessions. We also introduce learnable baseline normalization (LBN) to reduce baseline fluctuations. Tested on two HD-sEMG gesture recognition datasets, SAL and LBN outperformed standard fine-tuning on regular arrays, achieving competitive performance even with a logistic regressor, with orders of magnitude less, physically interpretable parameters. Our ablation study showed that forearm circumferential translations account for the majority of performance improvements.
Paper Structure (14 sections, 6 equations, 4 figures, 1 table)

This paper contains 14 sections, 6 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Relevant Work
  3. Methods & Datasets
  4. Spatial Adaptation Layer
  5. Datasets
  6. Simulated Spatial Perturbations
  7. Activity Images & Learnable Baseline Normalization
  8. Preprocessing & Classification
  9. Experimental Analysis
  10. Intrasession Performance
  11. Simulated Spatial Perturbations
  12. Intersession Performance
  13. Model Ablation & Interpretability Assessment
  14. Conclusion

Figures (4)

  • Figure 1: Schematic of SAL architecture and usage. SAL weights are frozen and the chosen model is optimized on session #1. On session #2, given a few examples, the model is frozen and the parameters of SAL and LBN are optimized for adaptation. SAL is similar to an STN module from spatial-transformers, with the key difference that rather than modelling affine coefficients $A_{s}$ as a function of the input, $A_{s}$ are treated as learnable parameters.
  • Figure 2: Majority voting classification accuracies for the intrasession protocol across different models and experimental conditions for (i) CSL and (ii) Capgmyo. For intrasession analysis, we train the classifier on all repetitions except one held out repetition per gesture, and obtain test accuracies on this held-out set. Across conditions, datasets and classifiers, high accuracies were obtained.
  • Figure 3: Majority voting classification accuracies and average electrode displacement for the simulated spatial perturbations across different models and conditions for CSL data. With SAL and input dropout, distances are significantly decreased, improving performance from near chance-level to over 60%.
  • Figure 4: Majority voting classification accuracies for the intersession protocol across different models and experimental conditions for (i) CSL and (ii) Capgmyo. For intersession analysis, we train the classifier on one recording session, then adapt to the second session with one repetition per gesture, using the remaining data from the second session to obtain test performance. We then obtain test accuracies on this held-out set. While SAL + LBN seems to significantly outperform fine-tuning for CSL (regular grid), fine-tuning still performs better on Capgmyo data (irregular grid).