Understood: MR Support for Adult ADHD

• Understood is a mixed reality communication-support system for adult ADHD that provides real-time assistance during face-to-face conversations.
• It leverages AI-driven summarization, context-aware word suggestions, and topic-shift detection to mitigate working-memory deficits, speech disfluency, and attentional drift.
• Empirical evaluations indicate that Understood reduces average pause and off-topic durations, complementing traditional therapeutic interventions.

Understood is a Mixed Reality (MR) communication-support system for adults with Attention Deficit Hyperactivity Disorder (ADHD), implemented on Microsoft HoloLens 2 to provide therapist-free, real-time assistance during face-to-face conversation. It is motivated by the observation that adult ADHD frequently manifests not as overt hyperactivity but as executive dysfunction, working-memory deficits, cognitive overload, and emotional dysregulation, all of which can impair discourse cohesion and self-regulation in live interaction. The system targets three recurrent difficulties identified in formative work—forgetting recent conversational content, becoming disfluent during speech, and drifting off topic—and addresses them through real-time conversation summarization, context-aware subsequent word suggestions, and topic-shift reminders. A within-subjects study and clinician review indicate high usability and significant reductions in average pause duration and off-topic duration, positioning the system as a complement to therapist-mediated interventions rather than a replacement for them (Zhang et al., 24 Jul 2025).

1. Clinical and interactional context

The problem addressed by Understood is situated in the adult presentation of ADHD, where communication difficulties are linked to executive dysfunction, working-memory deficits, cognitive overload, and emotional dysregulation rather than primarily to childhood-style overt hyperactivity. In conversational settings, these impairments manifest as difficulty retaining recent utterances, difficulty selecting and sequencing what to say, and difficulty sustaining attentional anchoring to the shared topic. Participants described forgetting what the other person had said while they were still speaking, experiencing memory “blanking” during their own turns, and abruptly shifting from the current topic to unrelated material. Clinicians in the formative study framed these failures as disruptions to maintaining context, making inferences, and preserving discourse cohesion (Zhang et al., 24 Jul 2025).

The system focuses on three specific challenge classes. The first is deficits in working memory, where the speaker cannot keep enough recent conversational material active to respond coherently. The second is disfluency and pauses, arising either from losing the thread of an utterance or from cognitive overload created by too many simultaneously activated candidate ideas. The third is attentional anchoring failure, expressed as topic shifting and excessive elaboration driven more by internal associations than by the conversational goal. Although impulsive turn-taking and interruption are acknowledged, the paper notes that adults often learn to self-regulate these more successfully over time, which is why Understood concentrates on the other three domains.

Existing interventions are described as effective but poorly transferred into everyday live conversation. Cognitive Behavioral Therapy (CBT) can support self-monitoring and cognitive restructuring, but it presupposes the very real-time monitoring capacities that ADHD can compromise. Social Skills Training (SST) can improve performance in role-play and scripted scenarios, yet its benefits often do not generalize to uncontrolled natural interaction. Self-regulation training and mindfulness can increase awareness of dysregulation without necessarily supplying the linguistic content needed at the moment of breakdown. Technology-mediated interventions such as VR, biofeedback, and neurofeedback are similarly limited by their emphasis on offline practice, intrusiveness, or lack of direct linguistic support. The paper summarizes these shortcomings as three gaps: lack of sustainable transfer, limited autonomy, and high disruption in real-world use (Zhang et al., 24 Jul 2025).

2. Formative research and design goals

Understood was derived from a two-stage formative process. The first stage consisted of semi-structured interviews designed to identify communication challenges and current coping strategies. Four licensed ADHD specialists—two psychiatrists and two therapists, with an average of 11.75 years of experience—were interviewed using 15 communication statements from the DIVA-5 adult ADHD interview. Twelve adults with DSM-5-diagnosed ADHD, five male and seven female with mean age approximately 21.5, were interviewed using the same statement set and asked to provide concrete real-life examples, coping methods, and experiences with therapy or self-help. The analysis used affinity diagramming and triangulated clinician and participant perspectives; findings were member-checked with participants (Zhang et al., 24 Jul 2025).

This stage confirmed the three communication challenges and also yielded two broader findings. First, participants and clinicians emphasized the limitations of existing methods, including medication side effects and poor generalization from CBT, SST, and self-regulation strategies. Second, they stressed the need for accessible, therapist-free, real-time support because adults often lack continuous clinical support precisely when conversational problems occur.

The second stage was a design study oriented toward potential interventions. It involved 13 participants: 8 adults with ADHD who had completed the earlier interviews, 3 HCI researchers, and 2 clinicians specializing in ADHD. Through brainstorming, storyboard generation, and a design workshop, the study produced several design findings. MR devices were preferred over phones or laptops because they preserve face orientation and avoid shifting attention to handheld screens; audio-only prompting was seen as potentially overloading attention. Participants preferred concise visual summaries as an “external memory,” short word or phrase suggestions rather than pictorial cues, subtle visual reminders for off-topic speech, rapid high-contrast presentation, minimalistic on-demand interfaces, positive feedback, and feature-level customization. These findings were consolidated into five design goals: accessible and customizable real-time support; conversational content summarization; predictive word suggestions for speech fluency; off-topic detection with subtle reminders; and an intuitive, minimalistic, encouraging interface (Zhang et al., 24 Jul 2025).

3. System architecture and computational pipeline

Understood is implemented on Microsoft HoloLens 2 using Unity 2022 and MRTK3 for the MR interface, Azure AI Speech for real-time speech recognition, and OpenAI GPT-4o for summarization, phrase suggestion, and off-topic classification. The system captures conversation audio through the HoloLens microphone, streams it to Azure AI Speech, and receives transcripts tagged by speaker as far as possible under a single-microphone configuration. These transcripts are organized as a dialogue history

$H = [(s_1, u_1), (s_2, u_2), \dots, (s_t, u_t)],$

where $s_i$ denotes speaker identity and $u_i$ the utterance text. This history is then passed to three separate GPT-4o prompt pipelines (Zhang et al., 24 Jul 2025).

For summarization, the system supplies the recent utterance $u_t$ and the previous summary $S_{t-1}$, and asks GPT-4o to produce a 4–12 word concise, high-level summary using keywords. The result is an incremental summary $S_t$ for either the user or the conversation partner. For word suggestions, the system uses the current user utterance and the dialogue history to request a single phrase of fewer than six words that would naturally continue the conversation. The paper frames this as functionally equivalent to a short-span continuation objective,

$$\text{Suggestion} \approx \arg\max_{\text{phrase } p,\ |p|<6} P(p \mid H, u_t).$$

For topic-shift detection, GPT-4o receives the current utterance and dialogue history and outputs exactly “Yes” or “No” according to whether the utterance deviates from the main conversational thread. The paper presents this as a binary classifier

$f(H, u_t) \in \{\text{"Yes"}, \text{"No"}\}.$

The three support features and their target functions can be summarized as follows.

Feature	Target challenge	Output form
Real-time conversation summarization	Working-memory deficits	4–12 word keyword summaries
Context-aware subsequent word suggestions	Disfluency and pauses	Single phrase of fewer than 6 words
Topic shifting detection and reminding	Attentional anchoring failure	Binary off-topic judgment mapped to trigger color

The computational design deliberately favors compact outputs over verbose LLM responses. Summaries are incrementally updated per utterance, while suggestions are refreshed approximately every second. Privacy is handled by streaming audio to Azure and OpenAI without local retention by Understood itself; the paper states that speech content is not archived by the system after the session (Zhang et al., 24 Jul 2025).

4. Mixed Reality interaction model

The interaction design is structured into three phases: functionality selection, conversation, and feedback. In the functionality selection phase, the user chooses which supports to activate through icon toggles corresponding to self-summarization, other-summarization, word suggestions, off-topic detection, and optional pop-up animation. Opacity indicates the active or inactive state of each feature, and the selected configuration persists into the live conversation.

During conversation, the central interaction element is a small trigger positioned near the bottom of the field of view and animated with gentle oscillation. When the user glances at this trigger, the selected support panels appear in front of them for 5 seconds. These can include separate panels for self-summarization, other-summarization, and word suggestions. Eye-gaze is also used to indicate focus: when the user looks at one panel, other panels dim and become semi-transparent, and if pop-up animation is enabled the focused panel slightly enlarges. This interface logic operationalizes the design preference for on-demand visual assistance rather than always-on overlays, thereby limiting distraction and preserving conversational continuity (Zhang et al., 24 Jul 2025).

The topic-shift reminder is encoded not as a separate textual panel but in the color state of the trigger itself. When GPT-4o classifies the current utterance as off-topic, the trigger gradually shifts toward deep red. This produces a peripheral and nonverbal cue rather than an explicit corrective message. The feedback phase occurs at the end of the interaction, when tapping the trigger opens a window showing how many times assistance was requested; an additional tap produces confetti animations as a form of positive reinforcement. The absence of overt scoring or admonishment is consistent with the design goal of supporting autonomy and reducing shame.

Several UI properties were chosen specifically for cognitive accessibility. The system uses BBC Reith font, high-contrast text, and semi-transparent dark panels to maintain legibility against arbitrary backgrounds. Eye-gaze controls focus, while hand gestures are reserved for explicit confirmation and session closure. Auditory cues are limited to simple interaction feedback sounds, since participants considered spoken prompts or additional concurrent audio streams potentially disruptive. Taken together, these choices make the MR interface less a general heads-up display than a controlled, low-density support channel for brief consultation during live interaction (Zhang et al., 24 Jul 2025).

5. Empirical evaluation

The main evaluation used a within-subjects design with 10 adults with ADHD, 4 male and 6 female, with mean age 21.3. Each participant completed two approximately 7-minute conversation tasks, one with Understood and one without it, with condition order counterbalanced. Topics were sampled from five everyday themes such as favorite place, weekend routine, go-to meal, relaxation, and phone apps. Each conversation included a 45-second supplementary passage read by the experimenter to increase working-memory demand, as well as divergent and convergent phases to elicit both exploratory talk and summarization. After the tasks, participants completed 5-point Likert questionnaires on effectiveness and usability and took part in semi-structured interviews. Two clinicians later reviewed the system and selected session videos to assess clinical alignment and risks (Zhang et al., 24 Jul 2025).

Objective communication measures showed that Understood primarily affected recovery rather than event frequency. The number of pauses and the number of off-topic segments did not change significantly, but both average pause duration and average off-topic duration were reduced.

Metric	With Understood	Without Understood
Number of pauses	$15.60 \pm 5.28$	$18.90 \pm 7.37$
Average pause duration	$s_i$0 s	$s_i$1 s
Number of off-topic segments	$s_i$2	$s_i$3
Average off-topic duration	$s_i$4 s	$s_i$5 s

The reduction in average pause duration was significant under a Wilcoxon test with $s_i$6, and the reduction in average off-topic duration was significant with $s_i$7. This pattern supports the interpretation that the system helps users re-establish fluency and topical anchoring more quickly rather than preventing every breakdown from occurring (Zhang et al., 24 Jul 2025).

The system’s LLM-mediated outputs were also quantified. Summaries averaged $s_i$8 words with response time $s_i$9 seconds, while suggestions averaged $u_i$0 words with response time $u_i$1 seconds. Participants rated summary length $u_i$2, summary response time $u_i$3, and summary correctness $u_i$4. For suggestions, they rated length $u_i$5, response time $u_i$6, and correctness $u_i$7. Perceived effectiveness was similarly high, including concentration on conversation at $u_i$8, naturalness during conversation at $u_i$9, and overall helpfulness at $u_t$0. Usability measures were also strong: ease of use $u_t$1, learnability $u_t$2, low interruption $u_t$3, low mental demand $u_t$4, and willingness for frequent use $u_t$5 (Zhang et al., 24 Jul 2025).

6. Clinical interpretation, limitations, and implications

Clinicians interpreted the system as aligning with several established therapeutic strategies while remaining distinct from them. Conversation summarization was understood as an externalization of working memory, analogous to CBT-like note-taking or organizational scaffolds. Word suggestions resembled the therapist prompts used in SST role-play, but delivered in real time during natural conversation. Topic-shift reminders were seen as a form of self-regulation support, making attentional drift visible without direct correction. At the same time, clinicians emphasized that Understood mainly addresses surface-level manifestations of impairment—pauses, lost context, topic drift—rather than the underlying executive dysfunction itself. On that basis, they treated it as a complement to therapy, not as a substitute for CBT or SST (Zhang et al., 24 Jul 2025).

The paper identifies several technical and methodological limitations. HoloLens 2 is heavy, expensive, and visually conspicuous, which constrains ecological validity and raises the possibility of device-related stigma. GPT-4o was chosen because smaller models such as GPT-3.5 and Phi-4-Mini produced poorer and often verbose outputs, but GPT-4o still exhibits latency variability and occasional hallucinations. Inaccurate suggestions can increase confusion and cognitive effort, and clinicians explicitly warned about over-reliance on suggestion prompts. The system attempts to mitigate this through the active-trigger design and automatic panel fading after 5 seconds, but the tension between support and self-reliance remains central. Speaker separation is limited by the single-microphone setup, and the evaluation sample was small, young, university-based, and short-term, with no longitudinal evidence וועגן transfer to everyday life.

Ethical considerations are equally central. Because Understood processes live speech, privacy and partner consent are structural concerns rather than peripheral implementation details. The system states that it does not store audio or transcripts locally after the session, but real-world deployment would still require careful governance of streamed conversational data. Autonomy is also a recurring concern: the system is explicitly designed not to impose a “normal” communication style, and its suggestions are presented as optional cues rather than directives. Feature-level customization, user-triggered invocation, and non-corrective visual feedback are therefore not merely interface conveniences but mechanisms for preserving agency.

Despite these constraints, the work establishes a concrete model for real-time AI-mediated communication support in neurodivergent adulthood. Its broader contribution lies in showing that MR can be used to externalize conversational context, lexical continuation, and topic monitoring without requiring a therapist to be present. A plausible implication is that future systems on lighter AR platforms could extend the same design logic to other populations, including aphasia, autism spectrum disorder, pragmatic language impairment, or second-language communication, provided that the balance between cognitive offloading, autonomy, privacy, and long-term skill development is preserved (Zhang et al., 24 Jul 2025).