Visualization and workload with implicit fNIRS-based BCI: toward a real-time memory prosthesis with fNIRS

Russell, Matthew; Hincks, Samuel; Wang, Liang; Babar, Amin; Chen, Zaiyi; White, Zachary; Jacob, Robert J. K. · 2025 · Crossref

DOI: 10.3389/fnrgo.2025.1550629

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitation of current functional Near-Infrared Spectroscopy (fNIRS)-based implicit Brain-Computer Interfaces (BCIs), which predominantly focus on detecting mental workload. The authors propose expanding the application of prefrontal cortex measurements to distinguish between different brain network states, specifically the Default Mode Network (DMN) and the Dorsolateral Prefrontal Cortex (DLPFC), to create a prototype "memory prosthesis." This system aims to automatically index and retrieve information based on the user’s passive brain state, acting as an adaptive assistant that presents relevant content without requiring conscious user intent. The study involved eight participants who performed two distinct tasks designed to engage different neural networks: a creative visualization task (Task A) intended to activate the DMN, and a complex furniture selection task involving budget management (Task B) intended to engage the DLPFC. Data was acquired using a Multichannel ISS Imagent fNIRS device with a single probe pad positioned over the left prefrontal cortex (Brodmann area 10). Due to technical issues, only one detector was used, limiting spatial coverage. The experimental design included two training groups and one testing group. For real-time classification, data was segmented into 17.24-second windows, filtered using Recursive Least Squares adaptive filters, and classified using a Support Vector Machine (SVM) with a linear kernel. Offline analyses employed leave-one-out cross-validation (LOO-CV) and tested various classifiers, including K-Nearest Neighbors, Linear Discriminant Analysis, and Random Forests, while also analyzing the contribution of lateral versus medial prefrontal source locations. Results indicated that the two tasks were differentiable, with LOO-CV achieving a performance of 71% across participants. However, real-time online classification showed high variability; the average macro F1-score was 0.516 with a standard deviation of 0.282. While some participants achieved high accuracy (e.g., F1-score of 0.964 for Participant 3), others showed near-zero performance, suggesting the model lacked robustness for reliable real-time application with the current dataset size. Offline analyses revealed that specific source locations in the lateral and medial left prefrontal areas contributed significantly to classification accuracy, highlighting promising approaches for future feature selection. The significance of this work lies in demonstrating the feasibility of using fNIRS to distinguish between distinct cognitive states beyond simple workload, moving toward a brain-based memory assistant. Although the prototype’s real-time performance was inconsistent, the study validates the concept of using passive neural signals as an index for information retrieval. The findings suggest that future systems could leverage specific prefrontal activation patterns to adapt interfaces fluidly to user needs, potentially enhancing knowledge-worker tasks by automatically surfacing relevant information based on the user’s current cognitive context. The authors plan to make the dataset publicly available to facilitate further research, including the application of deep learning methods.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices