Poster De Conférence Année : 2024

Investigating evolution of features in Brain Computer Interface experimentation for robustness

Résumé

Motor imagery brain-computer interfaces face a significant hurdle: many subjects struggle to control devices effectively. Performance hinges on the features employed, typically power spectral density (PSD) in the α or β bands, influenced modulated by different mental tasks like such as motor imagery (MI) and rest. Variations in features stem from Event-Related Desynchronization in the sensorimotor cortex. Monitoring feature evolution within a session is crucial for maintaining consistent performance that follows users ‘intent. In this work, we had 15 subjects (8 F, 25 +/- 1.5 yo) coming for 3 sessions where they controlled a robotic arm using MI for the gripper closing and an eye tracker for the position to reach . During a session, they went through 8 runs of 10 MI and Rest trials. 3 runs were of calibration (no feedback) then 3 runs of control based on a LDA trained on features of the calibration and then 2 runs of control with the LDA trained on the previous control runs. We investigated MI and resting state β band PSD in the left sensorimotor cortex (postcentral and precentral area) using source reconstruction at the group level with all sessions confounded to see if patterns rose between blocks. We explored the beneficial impact of feedback in BCI experiments and examined how specific features contributed to this effect. We also investigated whether calibration features are linked to online performance. We discovered that patterns of motor imagery (MI) and rest differed over runs. We then examined how the variability of MI-Rest PSD during calibration correlated with overall performance. We found a significant correlation (p = 0.006, r = 0.402), suggesting that initial differences in the ERD between conditions during calibration leads to better performance. However, despite MI being consistent across control blocks, rest showed block-specific behavior, indicating its intrinsic variability in the experimental design. Here, we explored features’ evolution used in a BCI experimentation, we could correlate their early profile to performances. From this, we can draw different conclusions, i) features ‘variation forces to adapt the algorithm either by retraining it or by using adaptive classifier, ii) As MI features were stable through the control period, they could be selected from the start as they seem robust, oppositely, rest features could gain in being retrained. To go further, it might be necessary to investigate other features more robust throughout experimentation such as connectivity .

Mots clés

Domaines

Fichier principal
Vignette du fichier
PosterSfnv2.pdf (1.05 Mo) Télécharger le fichier
Origine Fichiers produits par l'(les) auteur(s)
Licence

Dates et versions

hal-04890567 , version 1 (16-01-2025)

Licence

Identifiants

  • HAL Id : hal-04890567 , version 1

Citer

Tristan Venot, Camile Bousfiha, Fabrizio de Vico Fallani, Marie-Constance Corsi. Investigating evolution of features in Brain Computer Interface experimentation for robustness. SFN 2024 - Annual meeting of the Society for Neuroscience, Oct 2024, Chicago (IL), United States. ⟨hal-04890567⟩
171 Consultations
84 Téléchargements

Partager

  • More