Human-Centered Explainable AI for Brain-Computer Interface-driven Rehabilitation
Param Rajpura · 2026 · Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26) · doi:10.1145/3772363.3799222
Summary
This CHI Extended Abstract presents a doctoral research agenda developing human-centered explainable AI (HCXAI) frameworks for brain-computer interfaces (BCIs) used in stroke rehabilitation. Rajpura argues that current BCI systems prioritize technical performance, such as classification accuracy of motor imagery, while neglecting the patient-facing explanations that allow users to self-correct, calibrate trust, and continue practice without a therapist present. The work is situated in resource-constrained Indian healthcare settings, where physiotherapists are scarce (0.36 per 10,000 people) and only 40% of stroke survivors access rehabilitation. The dissertation is organized around four research questions covering how stakeholders conceptualize explainability, what HCXAI principles patient-facing systems require, how XAI can adapt to varied cognitive abilities and communication barriers, and whether explanatory feedback affects adherence and functional recovery. Methods combine systematic review (the XAI4BCI design space, drawn from 84 studies), formative co-design workshops with healthy participants, and video-based scaffolding sessions with three stroke survivors with moderate-to-severe aphasia and three caregivers, facilitated by clinicians. The author also reports on a working motor-imagery BCI prototype using 16-channel wireless EEG and a two-degree-of-freedom upper-limb exoskeleton, intended as the deployment platform for future adaptive XAI prototypes.
Key findings
The XAI4BCI design space identified four dimensions (Who, Why, What, How) and revealed that current BCI XAI research targets developers and clinicians rather than end users. Co-design sessions showed that stroke survivors with aphasia cannot easily articulate XAI preferences without scaffolding; four facilitation techniques worked: analogical bridging (mapping AI states to mobile signal bars), projective personas, binary forcing through A/B choices, and 15–30 second extended response windows. Reflexive analysis surfaced facilitation biases, including expert override, hypothesis confirmation, and authority effects, and proposed directionality and contradiction checks to make these biases analyzable. Mixed-methods work identified four dimensions of "therapeutic trust" needed for sustained engagement: perceptual transparency, motivational calibration, social mediation, and institutional legitimacy, organized using the ICF vocabulary. Stakeholders showed divergent explainability needs across granularity, mediation mode, and trust source, suggesting a single explanatory design is insufficient. Empirical benchmarking work demonstrated that classification accuracy alone is inadequate for evaluating BCI models, and proposed Earth Mover's Distance-based metrics to quantify spatial-domain neurophysiological explanations across EEG datasets.
Relevance
Patient-facing AI explanations are an underexplored area in accessibility, and this work makes the case that explainability for disabled users is not just a UI problem but a therapeutic-ecosystem problem spanning perception, motivation, family mediation, and clinical authority. The scaffolding methods, particularly analogical bridging and binary forcing for participants with aphasia, transfer directly to other participatory research with people who have communication or cognitive impairments, and the reflexive bias checks address a real risk in such work. The reframing of XAI from "justifying predictions" to "enabling action" is a useful lens for any AI system serving disabled users in high-stakes contexts. As an extended abstract, the paper previews future work rather than reporting completed deployment results, and its sample (three survivors, three caregivers) is small; the planned multi-site validation across Indian rehabilitation centers and longitudinal 6–12 week deployment studies will be needed to test generalizability.
Tags: brain-computer interface · explainable AI · stroke rehabilitation · aphasia · human-centered AI · participatory design · co-design · neurotechnology · trust calibration · Global South accessibility
Standards referenced: ICF (International Classification of Functioning, Disability and Health)