Source-linked AI summary
StrokeGuard: A Multi-Agent Guided System for Prehospital Stroke Assessment
Wentao Yang, Zhenye Xu, Ruoyi Li, Musen Zhang, Yao Guo
TL;DR
StrokeGuard addresses interaction, input-quality, workflow-order, and clinician-handoff gaps that assessment models or general conversational agents do not fully solve alone. It uses controlled multi-agent guidance for mobile FAST screening and reports a 10.83-point MATES-9 increase over a paper-form baseline in a simulated prehospital scenario.
Problem
Assessment models and general conversational agents do not fully address interaction, multimodal input-quality, workflow-order, or auditable handoff problems before usable prehospital stroke-assessment input is obtained.
Method
StrokeGuard separates formal facial, arm, and speech assessment from procedural support through dual-channel multi-agent guidance, state control, local fallback recovery, constrained video modules, and evidence-aware handoff.
Results
10.83 points: StrokeGuard increased the MATES-9 total score over the paper-form group, from 45.50±8.22 to 56.33±5.61, a 23.8% relative increase.
Takeaways & Limitations
StrokeGuard provides a more standardized and executable mobile FAST assessment process for low-expertise caregivers, with controlled execution and reviewable handoff.
Abstract
from arXiv · showhide
Prehospital stroke assessment aims to accurately identify stroke symptoms and make rapid decisions through standardized procedures within an extremely narrow time window, thereby saving valuable time for subsequent treatment. In clinical practice, FAST-based scales are widely used for prehospital stroke assessment by issuing instructions that guide subjects to perform specific actions to screen facial, arm, and speech functions. However, in home and community settings, non-clinical users often encounter challenges such as inaccurate descriptions, incomplete symptom observation, and difficult operational procedures, which may lead to inaccurate or biased assessment results. To address these challenges, this paper presents StrokeGuard: a multi-agent guided system designed for prehospital stroke assessment that makes mobile FAST screening more standardized and executable. Specifically, to overcome the limitations of traditional single-agent systems in terms of procedural fault tolerance and user guidance capability, StrokeGuard adopts a dual-channel agent mechanism that separates formal assessment (i.e., facial palsy, arm weakness, speech impairment) from procedural support (e.g., step prompts, error correction, and real-time feedback). It guides the assessment process through multi-agent collaboration, dual-channel interaction, state-machine control, and stage-local fallback recovery mechanisms. Stage-specific scoring is delegated to constrained pretrained video assessment modules, while evidence source records are integrated with structured report generation. The user evaluation uses MATES-9, an exploratory scale for measuring user experience in multistep AI-guided tasks. In a simulated prehospital scenario, StrokeGuard improves the MATES-9 total score over a paper FAST-style form by 10.83 points, corresponding to a 23.8% relative increase.
I. INTRODUCTION
StrokeGuard addresses the difficulty of executing FAST-based prehospital stroke assessment reliably outside clinical settings. It combines separated assessment and support functions with controlled orchestration, recovery, and evidence-preserving handoff.
- Motivation: FAST-based tools structure prehospital screening around observable facial, arm, and speech signs.They guide subjects through specific actions to screen these functions.
- Motivation: Low-expertise users may provide inaccurate descriptions, observe symptoms incompletely, or struggle with operational procedures.These challenges are reported in home and community settings, alongside limits in stroke recognition accuracy.
- Research Gap: AI stroke systems support early triage but do not fully solve input capture, workflow-order, or clinician-handoff problems.A general conversational agent likewise does not guarantee controlled scoring, adequate multimodal input quality, or an auditable handoff.
- System: StrokeGuard separates formal assessment of facial palsy, arm weakness, and speech impairment from procedural support such as prompts, error correction, and feedback.The system uses multi-agent collaboration, dual-channel interaction, state-machine control, stage-local fallback recovery, constrained pretrained video modules, and structured evidence records.
- Contributions: The paper frames FAST screening for low-expertise users as a multi-agent assessment-execution problem with explicit responsibility boundaries.These boundaries include formal assessment, procedural support, stage progression, local fallback, and source-preserving handoff.
- Evaluation: The evaluation combines mechanism-ablation case studies with MATES-9, an exploratory measure of multistep AI-guided task experience.The design examines dual-channel constraints, evidence-aware orchestration, and stage-local recovery.
A. Multimodal AI Algorithms for Stroke Assessment
Prior work establishes multimodal signals as useful for stroke-related assessment but often assumes usable input and focuses on recognition performance. StrokeGuard instead wraps pretrained assessment algorithms in a controlled, evidence-aware workflow that addresses capture and state-management problems.
- Prior Multimodal Algorithms: Facial dynamics, arm motion, speech signals, and wearable sensing provide useful evidence streams for stroke-related assessment.Prior studies examine feature extraction, modality fusion, classification, and functional estimation across these streams.
- Research Gap: Existing multimodal stroke studies mainly evaluate perception, fusion, classification, or functional-estimation performance after acceptable-quality input is available.This assumption leaves input availability and capture quality less directly addressed.
- StrokeGuard Approach: StrokeGuard does not retrain a stroke classifier; it wraps pretrained Face, Arm, and Speech algorithms as constrained modules.The modules are integrated into a staged assessment process rather than presented as a new recognition model.
- StrokeGuard Approach: The system addresses input availability through staged capture guidance, video-first processing, and single-stage text fallback.Capture failure is limited to the affected stage rather than propagating through the full assessment process.
- Multi-Agent Control: Role division alone is insufficient for reliable medical assessment because state changes, evidence ownership, and failure effects require explicit workflow boundaries.StrokeGuard combines multi-agent roles with state control, evidence-aware orchestration, and task-oriented dialogue control.
D. Human-Computer Interaction Evaluation in Medical Systems
StrokeGuard’s evaluation framework combines user-experience measurement with a layered architecture for interaction, deterministic state control, dedicated assessment, and auditability. MATES-9 focuses on dimensions relevant to multistep AI-guided tasks while requiring interpretation alongside behavioral evidence.
- MATES-9 Evaluation: MATES-9 measures task completion, step orientation, support continuity, workload, result understandability, and action-decision support.Reverse wording is used for Q4, Q5, Q6, and Q7 to reduce acquiescence bias.
- Overall Architecture: The User Interface Layer presents the current stage, capture action, and supportive communication to the caregiver.This layer is the caregiver-facing component of the layered architecture.
- Overall Architecture: StrokeGuard’s architecture links natural-language interaction with dedicated multimodal assessment modules and deterministic flow control.The system organizes these components into a unified assessment process for reviewable caregiver-led assessment.
- Overall Architecture: The Multi-Agent Coordination Layer routes input and natural-language interaction while coordinating deterministic assessment-state control.The Assessment and Audit Layer calls dedicated Face, Arm, and Speech modules and stores evidence sources and processing traces.
B. Dual-Channel Interaction and Assessment-State Separation
StrokeGuard separates clinical assessment from procedural support, allowing only validated assessment events to affect clinical state while support interactions remain state-neutral.
- Rationale: The separation addresses confusion caused when clinical evidence and procedural dialogue share one conversational control path.For example, a statement about arm weakness can be assessment evidence, whereas a question about holding the phone should only support capture.
- Channel separation: The dual-channel design routes formal assessment input separately from support requests such as camera guidance, explanations, reassurance, and emergency-response questions.Assessment events include media submission, fallback completion, processing failure, and stage completion; support produces natural-language replies only.
- Permission boundaries: Support-channel interactions cannot write assessment evidence or advance the assessment stage.The support channel can read session context and generate replies, but it has no state-transition permissions.
- System architecture: Deterministic orchestration routes inputs, controls stage progression and fallback, and preserves auditable channel and state effects.The architecture connects user input routing with deterministic control and source-preserving assessment outputs.
- State validation: State updates occur only when assessment events pass orchestration checks for stage consistency, completion conditions, and evidence conflicts.The orchestrator validates assessment events before changing clinical state.
C. Evidence-Aware Orchestration and Stage-Local Recovery
StrokeGuard manages assessment as ordered stages with evidence-aware updates and local recovery, preserving video priority while containing media failures to the affected step.
- Evidence-aware orchestration: The orchestrator uses a finite-state machine covering patient information, Face, Arm, Speech, processing, and done stages.Each stage has an independent evidence slot, processing state, and stage context.
- Evidence-aware orchestration: Video is preferred because it preserves direct cues for facial movement, arm posture, and speech performance, while structured text fallback handles local capture failures.Fallback records the lower evidence level and the reason for using text.
- Evidence-aware orchestration: Evidence priority increases from pending to text and then to video, and new evidence is accepted only after session, stage, and task identity checks.The stage evidence slot records conclusion, modality, processing state, and temporal information.
- Evidence-aware orchestration: Delayed external results are checked for session, stage, and task identity, while duplicate callbacks are handled idempotently.These checks protect evidence writes when asynchronous assessment results arrive.
- Stage-local recovery: Stage-local recovery keeps the current stage active for rerecording, enables text fallback only at that stage, and prevents rollback or failure inheritance.This converts a session-level media failure into stage-level compensation while maintaining assessment continuity.
D. Source-Preserving Structured Handoff
StrokeGuard generates handoff reports as structured projections of the assessment record, linking each conclusion to its evidence source, processing history, and fallback status for clinician review.
- Structured handoff: Each stage conclusion is bound to evidence modality, source, processing state, fallback status, and event history rather than presented as a final label alone.The report preserves how each result was formed and the level of evidence supporting it.
- Source preservation: Stage records store result fields, source fields, and event fields covering routing, submission, failure, fallback, callbacks, and stage progression.Immediate and delayed results enter the same stage record and are merged by deterministic rules.
- Structured handoff: A fixed report template presents each conclusion with its source trace, making system behavior reconstructable when results conflict with on-site observations.The trace identifies input channel, stage advancement, fallback occurrence, and the event producing the current result.
E. Prototype Implementation
The StrokeGuard prototype implements a complete prehospital assessment through caregiver-facing applications, backend orchestration, assessment services, fallback paths, and structured handoff generation.
- Prototype components: The prototype includes caregiver-facing mobile and web applications for conducting the assessment.These applications provide the user-facing interaction layer.
- Prototype components: A backend orchestration service manages stage progression and evidence writing.The service coordinates the assessment workflow and its recorded evidence.
- Prototype components: Supporting services connect stage-specific assessment algorithms, text fallback paths, and structured handoff generation.Together, these components support the complete prehospital assessment workflow.
IV. EXPERIMENTS AND RESULTS
The experiments use mechanism-ablation case studies to examine how StrokeGuard’s control mechanisms address failure modes in prehospital FAST assessment. Comparisons retain open-ended language-model dialogue and task generation while removing one mechanism at a time.
- Experimental Design: Mechanism-ablation case studies examine typical failure modes of open-ended agents in prehospital FAST assessment.Each comparison removes one StrokeGuard control mechanism while retaining the dialogue and task-generation ability of an open-ended language model.
- Experimental Design: Three cases remove the dual-channel constraint, evidence-aware orchestration, or stage-local recovery, respectively.Tables II–IV compare each ablated agent with StrokeGuard under the same input and shared elicitation prefix.
- Findings: StrokeGuard constrains three risks: support requests cannot change clinical state, evidence is written to the correct stage, and local failures cannot contaminate completed stages.The case study provides mechanistic and explanatory evidence rather than statistically estimating causal effects.
B. MATES-9 Scale Design and Simulated Prehospital Evaluation Results
The paper introduces MATES-9 as an exploratory measure of user experience in multistep AI-guided tasks and evaluates it in a simulated prehospital assessment. StrokeGuard users report higher overall experience scores than paper-form users, alongside shorter assessment time.
- MATES-9 Scale Design: MATES-9 measures task guidance, information clarity, support continuity, perceived workload, and action-decision support.Its constructs are adapted and integrated from established usability, workload, mobile-health, health-information, chatbot, conversational-interface, and technology-acceptance scales.
- MATES-9 Scale Design: MATES-9 keeps judgments focused and should be interpreted with task logs, completion outcomes, and open-ended feedback rather than alone.The scale includes task completion, step orientation, support continuity, workload, result understandability, and action-decision support.
- MATES-9 Scale Design: All MATES-9 items use a 7-point Likert scale, with reverse-worded items converted using Qʳ_j = 8 − Q_j for j = 4, . . . , 7.The paper reports raw means and transformed 0-100 scores computed as (mean − 1)/6 × 100.
- Simulated Evaluation: Twelve participants compare a paper FAST-style assessment form with StrokeGuard in the same simulated prehospital scenario.A professional male actor simulates FAST-related symptoms as an apparently 70-year-old patient and is presented as the participant’s elderly relative.
- Simulated Evaluation: Fig. 3 displays individual total-score distributions and transformed 0-100 domain scores, while inferential statistics are treated as exploratory because the sample is small.Panel (a) includes interquartile ranges, group means, and half-violin densities; Panel (b) shows transformed domain scores.
- Results: 10.83 points: StrokeGuard raises the MATES-9 total score from 45.50±8.22 with the paper form to 56.33±5.61, a 23.8% relative increase.The transformed 0-100 total score increases from 67.6 to 87.7; mean assessment time decreases from 116 s to 81 s, a 30.2% reduction.
V. CONCLUSIONS AND DISCUSSION
StrokeGuard combines multi-agent interaction, multimodal assessment, and deterministic state control to make mobile FAST screening more standardized and reviewable for low-expertise caregivers. Its conclusions emphasize controlled evidence handling and source-preserving handoff, while identifying prototype validation and integration as future work.
- Conclusions: StrokeGuard unifies natural-language interaction, dedicated multimodal assessment modules, and deterministic state control for reviewable low-expertise caregiver assessment.The process addresses help seeking, recording, failure recovery, and result handoff within one assessment workflow.
- Contributions: Controlled multi-agent assessment, evidence-aware execution, and source-preserving handoff are central StrokeGuard contributions.Formal assessment guidance and procedural support are assigned to different agents, while clinical-state writing remains under deterministic orchestration.
- Mechanism Comparisons: Tables II and III correspond to comparisons involving dual-channel constraints and evidence-aware orchestration.Their captions identify Case A and Case B as shared-elicitation comparisons with and without those mechanisms.
- Limitations and Future Work: StrokeGuard remains a research prototype, requiring larger and more realistic validation of assessment modules, thresholds, MATES-9, and mechanism-ablation effects.Future work also includes integration with dispatch, electronic health records, and hospital receiving systems.