[Paper Review] Minimizing inter-subject variability in fNIRS based Brain Computer Interfaces via multiple-kernel support vector learning

Berdakh Abibullaev, Jinung An+3 more|arXiv (Cornell University)|Sep 25, 2012

EEG and Brain-Computer Interfaces12 references20 citations

TL;DR

This paper proposes a multiple-kernel support vector machine (MK-SVM) framework to minimize inter-subject and inter-session variability in fNIRS-based brain-computer interfaces (BCIs), enabling robust classification without lengthy calibration. By integrating subject- and session-specific feature spaces into a unified, richer feature space with optimized decision boundaries, the method achieves high generalization across unseen subjects and sessions, significantly reducing calibration needs while maintaining strong performance on oxy-hemoglobin signals.

ABSTRACT

Brain signal variability in the measurements obtained from different subjects during different sessions significantly deteriorates the accuracy of most brain-computer interface (BCI) systems. Moreover these variabilities, also known as inter-subject or inter-session variabilities, require lengthy calibration sessions before the BCI system can be used. Furthermore, the calibration session has to be repeated for each subject independently and before use of the BCI due to the inter-session variability. In this study, we present an algorithm in order to minimize the above-mentioned variabilities and to overcome the time-consuming and usually error-prone calibration time. Our algorithm is based on linear programming support-vector machines and their extensions to a multiple kernel learning framework. We tackle the inter-subject or -session variability in the feature spaces of the classifiers. This is done by incorporating each subject- or session-specific feature spaces into much richer feature spaces with a set of optimal decision boundaries. Each decision boundary represents the subject- or a session specific spatio-temporal variabilities of neural signals. Consequently, a single classifier with multiple feature spaces will generalize well to new unseen test patterns even without the calibration steps. We demonstrate that classifiers maintain good performances even under the presence of a large degree of BCI variability. The present study analyzes BCI variability related to oxy-hemoglobin neural signals measured using a functional near-infrared spectroscopy.

Motivation & Objective

• To address the major challenge of inter-subject and inter-session variability in fNIRS-based BCI systems, which hinders generalization and necessitates lengthy, subject-specific calibration.
• To reduce the dependency on time-consuming and error-prone calibration sessions by enabling a single classifier to generalize across diverse subjects and sessions.
• To develop a learning framework that captures and models subject- and session-specific spatio-temporal neural signal variability in the feature space.
• To improve BCI performance by integrating multiple subject- and session-specific feature representations into a unified, optimized decision boundary.
• To demonstrate the effectiveness of the proposed method on oxy-hemoglobin signals measured via functional near-infrared spectroscopy (fNIRS).

Proposed method

• The method employs linear programming support vector machines (LP-SVM) extended into a multiple-kernel learning (MKL) framework to handle multiple feature spaces.
• Each subject or session is represented by a distinct feature space derived from oxy-hemoglobin signal patterns, capturing their unique spatio-temporal variability.
• The MKL framework combines these individual feature spaces into a unified, richer feature space using optimized kernel weights to learn a single, robust decision boundary.
• The optimization process uses linear programming to determine the optimal combination of kernels, minimizing classification error while accounting for inter-subject and inter-session variability.
• The resulting classifier generalizes well to new, unseen subjects and sessions without requiring retraining or calibration.
• The approach explicitly models variability in the feature space rather than in raw signal space, enhancing adaptability and performance.

Experimental results

Research questions

• RQ1 Can a single fNIRS-BCI classifier generalize across different subjects and sessions without requiring subject-specific calibration?
• RQ2 How effectively can multiple-kernel learning reduce inter-subject and inter-session variability in fNIRS-based BCI systems?
• RQ3 To what extent does integrating subject- and session-specific feature spaces improve classification accuracy and generalization in fNIRS-BCI?
• RQ4 Can the proposed MK-SVM framework maintain high performance under high levels of neural signal variability?
• RQ5 What is the impact of optimizing kernel weights across multiple feature spaces on classifier robustness and calibration reduction?

Key findings

• The proposed multiple-kernel learning framework significantly reduces inter-subject and inter-session variability in fNIRS-based BCI systems, enabling high-performance classification without calibration.
• The classifier generalizes effectively to new, unseen subjects and sessions, demonstrating strong robustness across diverse neural signal patterns.
• The method maintains high classification accuracy even under large degrees of BCI variability, indicating resilience to neural signal fluctuations.
• By integrating subject- and session-specific feature spaces into a unified framework, the approach eliminates the need for repeated calibration for each individual.
• The use of linear programming in the MK-SVM framework enables efficient optimization of kernel weights, improving decision boundary quality and generalization.
• The framework achieves reliable performance on oxy-hemoglobin signals, confirming its suitability for real-world fNIRS-BCI applications.