# Can Stroke Survivors With Aphasia Tell Us What They Want From BCI Explainability?

A formative study with six participants — three stroke survivors (two with moderate-to-severe aphasia) and three caregivers — demonstrates that structured scaffolding protocols can successfully surface explainable AI (XAI) requirements from users who cannot reliably navigate standard elicitation methods. The work, by Param Rajpura and Yogesh Kumar Meena (arXiv:2607.25423), addresses one of the least-discussed design bottlenecks in rehabilitation [brain-computer interfaces](https://bciintel.com/glossary/brain-computer-interface): we don't actually know what patients want from the algorithms making decisions about their recovery, partly because we've been asking the wrong way.

The core finding: a video-based scaffolding protocol, employing four specific facilitation techniques, can elicit heterogeneous and sometimes conflicting XAI preferences from stroke survivors with acquired communication disorders. The same protocol also exposed three systematic biases that facilitators themselves introduce — normative bias, hypothesis confirmation bias, and presence effect — each capable of corrupting the requirements data that eventually shape device design. The authors propose these biases as protocol risk guidelines for practitioners.

For the BCI field, this matters immediately. Motor-imagery and EEG-based rehabilitation BCIs are heading toward broader clinical deployment, and regulatory expectations around algorithm transparency are tightening. Building trustworthy human-machine systems without grounding XAI design in actual patient preferences is a design failure that will surface at the FDA, in device recalls, or in non-adoption.

---

## Why Standard Requirements Elicitation Fails Aphasia Patients

Conventional user research — interviews, surveys, card-sorting, think-aloud protocols — implicitly assumes users can articulate abstract preferences in real time under moderate cognitive load. Stroke survivors with aphasia cannot reliably do this. The communication disorder doesn't reflect cognitive capacity, but it does structurally block the linguistic channels that standard elicitation relies on.

The authors frame this as a "dyad-to-triad" problem. Rehabilitation BCI research has long recognized the patient-caregiver dyad as the functional unit of care. This paper argues that requirements elicitation must expand to a triad: patient, caregiver, and facilitator — with the facilitator's role formalized and the biases that role introduces systematically documented rather than ignored.

This is methodologically honest in a way that much BCI human-factors research is not. The facilitator is always present; pretending neutrality doesn't eliminate the presence effect, it just makes it invisible.

---

## The Four Scaffolding Techniques

The protocol developed and piloted in this formative study employs four distinct facilitation approaches, each targeting a specific barrier:

**1. Analogical bridging** maps AI system states to familiar everyday systems, reducing the conceptual gap for users who lack a framework for algorithmic transparency. If a patient cannot reason about "why the decoder classified this EEG segment as motor intent," they may be able to reason about why a familiar machine behaves unexpectedly — and that analogical transfer can surface genuine preferences.

**2. Projective personas** depersonalize sensitive topics by asking participants to respond on behalf of a fictional character rather than themselves. This is a standard technique in participatory design but its application to aphasia rehabilitation contexts is less documented.

**3. Binary forcing** reduces cognitive load by constraining responses to two options, lowering the working memory and language-production demands on aphasia patients without eliminating meaningful preference expression.

**4. Extended response time** acknowledges that aphasia patients' processing and production timelines differ from neurotypical users, and that standard interview pacing systematically disadvantages them.

The study reports that these approaches "successfully surfaced heterogeneous, sometimes conflicting XAI needs across participants." The conflicting nature of the findings is itself important data — it argues against a one-size-fits-all transparency layer in rehabilitation BCI software, and toward personalized or configurable explainability interfaces.

---

## Three Facilitator Biases That Corrupt Requirements Data

The reflexive analysis component of this work is arguably its most practically useful contribution. The authors identify three biases introduced by facilitators themselves:

- **Normative bias**: Facilitators inadvertently signal what a "correct" or socially acceptable preference looks like, steering responses toward perceived norms rather than genuine patient preferences.
- **Hypothesis confirmation bias**: Facilitators with prior beliefs about what patients need tend to probe in directions that confirm those beliefs, suppressing contradictory signals.
- **Presence effect**: The physical or social presence of the facilitator shapes how participants respond, independent of the questions asked — a confound that cannot be eliminated, only acknowledged and mitigated.

These three are presented as protocol risk guidelines — explicit cautions for any practitioner attempting to replicate or extend this methodology. This framing is appropriate: naming a bias as a known risk is more useful to the field than claiming the protocol eliminates it.

---

## Implications for BCI Clinical Translation and Device Design

The rehabilitation BCI market — EEG-based motor imagery systems, neurofeedback platforms, and emerging implantable devices for stroke recovery — is at an inflection point. [MindMaze](https://bciintel.com/companies/mindmaze) and [Neurolutions](https://bciintel.com/companies/neurolutions) have demonstrated commercial traction in neurorehabilitation. As these systems incorporate adaptive algorithms and closed-loop feedback, the black-box problem becomes clinically acute: a therapist adjusting stimulation parameters or a caregiver monitoring a patient's session cannot meaningfully oversee what they cannot interpret.

Regulatory pressure is moving in the same direction. The FDA's growing interest in Software as a Medical Device (SaMD) transparency, and the agency's stated concern about AI/ML-based device modifications, creates a near-term compliance driver for rehabilitation BCI developers to formalize XAI frameworks. What Rajpura and Meena are arguing — that patient-facing XAI requirements must be elicited before system design, not retrofitted after — maps directly onto the FDA's own guidance philosophy around user needs documentation in design controls.

The small scale of this formative study (six participants) means no quantitative outcomes should be extrapolated. This is explicitly a methodological contribution: a reusable protocol and a set of practitioner guidelines. Larger validation studies across diverse stroke populations, aphasia severities, and rehabilitation BCI platforms are the necessary next step before these methods can be considered standard practice.

**Analysis:** The field has generated substantial literature on decoding accuracy, electrode biocompatibility, and trial enrollment. It has generated comparatively little on what patients actually understand about — or want from — the algorithms processing their neural signals. This paper is a corrective. Its formative scale is a limitation but not an indictment; every validated protocol begins somewhere. The more significant question is whether rehabilitation BCI developers will integrate this kind of requirements work into their design pipelines before regulatory or market pressures force the issue.

---

## Key Takeaways

- A video-based scaffolding protocol successfully elicited XAI preferences from stroke survivors with aphasia, including two with moderate-to-severe communication disorders, in a six-participant formative study.
- Four techniques — analogical bridging, projective personas, binary forcing, and extended response time — addressed specific barriers imposed by aphasia on standard elicitation methods.
- Participants expressed heterogeneous, sometimes conflicting XAI needs, arguing against uniform transparency designs in rehabilitation BCI software.
- Reflexive analysis identified three facilitator-introduced biases (normative bias, hypothesis confirmation bias, presence effect) that are presented as protocol risk guidelines.
- The authors position patient-facing XAI requirements elicitation as a necessary design prerequisite, not an optional preliminary — a stance with direct implications for FDA design-control documentation.
- This is a small formative study; findings are methodological, not quantitatively generalizable.

---

## Frequently Asked Questions

**What is XAI in the context of rehabilitation BCIs?**
Explainable AI (XAI) refers to techniques that make an algorithm's decisions interpretable to humans. In rehabilitation BCIs — systems that decode neural signals to guide motor recovery or neurofeedback — XAI would allow patients, caregivers, and clinicians to understand why the system classified a neural pattern in a particular way or why it adjusted a feedback parameter.

**Why is eliciting XAI requirements from aphasia patients methodologically difficult?**
Aphasia is an acquired communication disorder, typically following stroke, that impairs language production and/or comprehension. Standard requirements elicitation relies on verbal or written articulation of abstract preferences — cognitive and linguistic tasks that aphasia directly disrupts. The challenge is extracting genuine preferences without the linguistic scaffolding that most user research assumes.

**What is the "dyad to triad" concept in this paper?**
Stroke rehabilitation research has long centered the patient-caregiver dyad as the functional unit. This paper argues that requirements elicitation expands this to a triad by formally including the facilitator — and that the facilitator's own biases must be documented and managed, not ignored.

**How does this research affect rehabilitation BCI developers commercially?**
Companies building adaptive or AI-driven rehabilitation BCI platforms face growing FDA scrutiny of algorithmic transparency in Software as a Medical Device. Formalizing patient-facing XAI requirements as part of design-control documentation — rather than treating explainability as a post-hoc feature — aligns with regulatory expectations and may reduce design-control gaps during PMA or De Novo review.

**Is this protocol ready for widespread clinical adoption?**
No. This is a formative, six-participant pilot study. The protocol and risk guidelines are offered as reusable methodological contributions requiring validation across larger, more diverse stroke populations before they can be considered standard practice.