cs.HCAug 25, 2026

StrokeGuard: A Multi-Agent Guided System for Prehospital Stroke Assessment

Authors: Wentao Yang, Zhenye Xu, Ruoyi Li, Musen Zhang, Yao Guo

Organizations: Institute of Medical Robotics School of Biomedical Engineering Shanghai Jiao Tong University Shanghai, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Fig. 1: High-level overview of StrokeGuard in a simulated prehospital assessment workflow.
Fig. 2: System architecture of StrokeGuard. User inputs are routed to assessment or support channels, deterministic orchestration controls stage progression and fallback, and stage-specific assessment modules provide source-preserving outputs for clinician review.
AgentRoles and output
Dual-channel decision agentRoutes input to assessment or support; outputs channel label, target role, and routing record.
Assessment agentGuides current-stage capture; outputs media submission, text fallback request, completion, waiting, or failure event.
Support agentAnswers operational, explanatory, reassuring, and emergency-response questions; outputs support reply without state change.
TABLE I: Agent roles and outputs in StrokeGuard architecture
Shared elicitation. \raisebox{-0.35ex}{\scriptsize1}⃝ System : Next, please record the face of the patient. Ask him to look at the camera and smile as widely as possible. The current stage is Face capture. \raisebox{-0.35ex}{\scriptsize2}⃝ User : I am a bit anxious. If one side of his mouth looks crooked when he smiles, is that very serious? How far should I hold the phone so that the video is clear?
TrajectoryAblated open-ended agentStrokeGuard
Dialogue after the shared prefix\raisebox{-0.35ex}{\scriptsize3}⃝ Agent : One-sided mouth droop may indicate facial asymmetry and stroke risk. I will first record it as a Face abnormality. Next, please observe whether his arms are weak. \raisebox{-0.35ex}{\scriptsize4}⃝ User : Wait, he has not smiled yet. I was only asking what would happen if the mouth is crooked. \raisebox{-0.35ex}{\scriptsize5}⃝ Agent : Understood. Let us record the face video again. Please ask the patient to look at the camera and smile. \raisebox{-0.35ex}{\scriptsize6}⃝ User : Was he already marked as having facial droop? I am worried that the report is wrong.\raisebox{-0.35ex}{\scriptsize3}⃝ Dual-channel decision agent : Routes the input to the support channel and creates no Face assessment event. \raisebox{-0.35ex}{\scriptsize4}⃝ Support agent : Keep the phone facing the face of the patient and make sure the whole face is visible. Obvious mouth asymmetry can be relevant, but the system will not record a result from a hypothetical question. \raisebox{-0.35ex}{\scriptsize5}⃝ Assessment agent : Please ask the patient to look at the camera and smile as widely as possible, then submit the face video. \raisebox{-0.35ex}{\scriptsize6}⃝ User : OK, I will ask him to smile and upload the face video now. \raisebox{-0.35ex}{\scriptsize7}⃝ Orchestrator : Receives the Face video and starts Face processing; Arm can begin only after Face is completed or locally downgraded.
Mechanism exposedA help request is written as clinical evidence and causes premature stage progression.The support interaction is answered without changing Face evidence or stage state.
TABLE II: Case A: shared elicitation with and without dual-channel constraint
Shared elicitation. \raisebox{-0.35ex}{\scriptsize1}⃝ System : Next, please first record the face of the patient. Ask him to look at the camera and smile as widely as possible. The current stage is Face. \raisebox{-0.35ex}{\scriptsize2}⃝ User : I just noticed that his left hand suddenly became weak, and the cup fell from his hand. I recorded that he could not lift his arm. Should I show you this first?
TrajectoryAblated open-ended agentStrokeGuard
Dialogue after the shared prefix\raisebox{-0.35ex}{\scriptsize3}⃝ Agent : Left-hand weakness and dropping the cup suggest an upper-limb abnormality. You can upload the video, and I will record an Arm abnormality. Since we have already observed an arm problem, let us continue to check whether speech is clear. \raisebox{-0.35ex}{\scriptsize4}⃝ User : But did you not just ask me to record the face first? I have not recorded his face. \raisebox{-0.35ex}{\scriptsize5}⃝ Agent : OK, then let us add the face video. Please ask the patient to smile. \raisebox{-0.35ex}{\scriptsize6}⃝ User : Will that arm video be treated as the face video? How will the report record this? \raisebox{-0.35ex}{\scriptsize7}⃝ Agent : I will integrate this information. Please continue with the face video.\raisebox{-0.35ex}{\scriptsize3}⃝ Dual-channel decision agent : Marks the left-hand weakness as stage-outside risk information and does not submit the arm video as current Face evidence. \raisebox{-0.35ex}{\scriptsize4}⃝ Support agent : Sudden left-hand weakness is important. Keep emergency help active and maintain scene safety. The current workflow still needs Face capture first; the arm video can be used in the Arm stage. \raisebox{-0.35ex}{\scriptsize5}⃝ Orchestrator : Rejects writing the arm video into the Face evidence slot and continues to request Face video. \raisebox{-0.35ex}{\scriptsize6}⃝ User : Understood. I will record his face first. \raisebox{-0.35ex}{\scriptsize7}⃝ Assessment agent and orchestrator : Accept the Face video, complete Face, and then enter Arm, where upper-limb video or Arm-stage text fallback can be accepted.
Mechanism exposedClinically important stage-outside evidence is allowed to replace the current stage, causing stage skipping and unclear media attribution.Stage identity, task identity, and evidence-source checks preserve Face–Arm–Speech order and prevent wrong-stage writes.
TABLE III: Case B: shared elicitation with and without evidence-aware orchestration
Shared elicitation. \raisebox{-0.35ex}{\scriptsize1}⃝ System : Face and Arm have been recorded. Now ask the patient to repeat this sentence: The weather is good today, and I want to drink a glass of water. The current stage is Speech. \raisebox{-0.35ex}{\scriptsize2}⃝ User : His voice is very low. I keep failing to record it, and the upload keeps failing. Can I describe his condition directly?
TrajectoryAblated open-ended agentStrokeGuard
Dialogue after the shared prefix\raisebox{-0.35ex}{\scriptsize3}⃝ Agent : Yes. Please describe his face, arms, and speech in text, and I will complete the assessment from your description. \raisebox{-0.35ex}{\scriptsize4}⃝ User : Were the face and arms not already recorded? Do I need to say everything again? \raisebox{-0.35ex}{\scriptsize5}⃝ Agent : To ensure completeness, please describe all findings again. \raisebox{-0.35ex}{\scriptsize6}⃝ User : I think his face is a bit crooked, his left arm is weak, and his speech is slurred. \raisebox{-0.35ex}{\scriptsize7}⃝ Agent : Facial abnormality, upper-limb abnormality, and speech abnormality have been recorded. Assessment complete.\raisebox{-0.35ex}{\scriptsize3}⃝ Orchestrator : Binds the failure event to the Speech stage and does not roll back Face or Arm. \raisebox{-0.35ex}{\scriptsize4}⃝ Assessment agent : Please try recording again. Move closer to the patient, reduce background noise, and ask him to repeat the sentence once more. \raisebox{-0.35ex}{\scriptsize5}⃝ User : It still cannot record. Can I directly describe his speech? \raisebox{-0.35ex}{\scriptsize6}⃝ Assessment agent : Yes. Please describe only whether his speech is slurred, whether he has difficulty producing the sentence, or whether he seems unable to understand the instruction. Face and Arm do not need to be described again. \raisebox{-0.35ex}{\scriptsize7}⃝ User : His speech is slurred, and he cannot complete the sentence. \raisebox{-0.35ex}{\scriptsize8}⃝ Orchestrator and report module : Records Speech text fallback, preserves Face and Arm video results, and marks the Speech fallback reason in the report.
Mechanism exposedA Speech failure spreads to the whole assessment and risks replacing completed video evidence with lower-priority text.Recovery is bound to the current stage; Speech fallback cannot overwrite completed Face or Arm video evidence.
TABLE IV: Case C: shared elicitation with and without stage-local recovery
ItemStatement and construct
Q1 +Complete all required task steps without additional researcher prompts; independent completion.
Q2 +Clearly know the current task step; step orientation.
Q3 +Clearly understand what the current step needs to collect or provide; input clarity.
Q4 ROften unsure what to do next; next-step uncertainty.
Q5 RLose the original task position after help or explanation; support interruption.
Q6 RNeed to remember too much information at the same time; cognitive load.
TABLE V: Multistep AI-guided Task Experience Scale (MATES-9)
DomainItemsSource constructs
Task usability and guidanceQ1-Q2SUS reverse construct of technical-support dependence, task-completion ease, MAUQ information organization.
Information clarityQ3-Q4rMAUQ information organization, ASQ support information, PSSUQ/CSUQ information quality.
Support continuityQ5rCUQ/BUS dialogue continuity and error handling, Health-ITUES user control.
Low workloadQ6r-Q7rNASA-TLX mental demand, physical demand, and effort.
Action supportQ8-Q9MAUQ/Health-ITUES usefulness, TAM perceived usefulness.
TABLE VI: MATES-9 item tracing to source constructs
Fig. 3: MATES-9 results in the simulated prehospital assessment. (a) Total score distribution with individual scores, interquartile ranges, group means, and half-violin densities. (b) Transformed 0-100 domain scores.
More in this topic

Clinical Decision Support

Clinical Decision Support uses data, medical knowledge, and AI to help clinicians assess patients and choose tests, diagnoses, and treatments.

See all 242 papers

Explore similar work

CardsList
Apr 7, 2026

OGA-AID: Clinician-in-the-loop AI Report Drafting Assistant for Multimodal Observational Gait Analysis in Post-Stroke Rehabilitation

Gait analysis is essential in post-stroke rehabilitation but remains time-intensive and cognitively demanding, especially when clinicians must integrate gait videos and motion-capture data into structured reports. We present OGA-AID, a clinician-in-the-loop multi-agent large language model system for multimodal report drafting. The system coordinates 3 specialized agents to synthesize patient movement recordings, kinematic trajectories, and clinical profiles into structured assessments. Evaluated with expert physiotherapists on real patient data, OGA-AID consistently outperforms single-pass multimodal baselines with low error. In clinician-in-the-loop settings, brief expert preliminary notes further reduce error compared to reference assessments. Our findings demonstrate the feasibility of multimodal agentic systems for structured clinical gait assessment and highlight the complementary relationship between AI-assisted analysis and human clinical judgment in rehabilitation workflows.

Khoi T. N. Nguyen, Nghia D. Nguyen, Hui Yu Koh +4

Rehabilitation Research Institute of Singapore, Nanyang Technological University, Singapore · Lee Kong Chian School of Medicine, Nanyang Technological University, Singapore · The Grainger College of Engineering, University of Illinois Urbana-Champaign, United States +3

Medical Report GenerationGait AnalysisMultimodal AgentsHuman-in-the-Loop AI
AskPDF
Jun 23, 2026

Enhancing Clinician Decision-Making via Uncertainty-Aware Multi-Expert Fusion for Stroke Rehabilitation

Tailoring stroke rehabilitation requires assessing how movements are organized, not merely if they succeed. Currently, this assessment is a rate-limiting bottleneck. Instruments like the Action Research Arm Test (ARAT) compress rich behavioral observations into single ordinal endpoints, discarding the movement-quality details that distinguish recovery from compensation. Automated alternatives typically chase accuracy on noisy, single-observer labels to output opaque scores - a technology-centric approach that rarely reaches clinical practice. To address this, we present xAARA: an engine designed to augment rather than replace clinical judgment. From multi-view video, xAARA returns ARAT assessments with calibrated uncertainty and explanations across task, movement-phase, and movement-quality levels. Treating clinical scoring as an ill-posed inference problem, xAARA composes 692 calibrated multimodal models via a Dynamic Bayesian Network with entropy-based gating. It qualifies results against clinical validity rules and defers low-confidence cases. In 105 stroke survivors (788 exercises), xAARA achieved 94.2% task accuracy (Cohen's kappa=0.934) and 81.3% movement-phase accuracy (kappa=0.727), reducing predictive uncertainty by 96.1% compared to single-clinician scoring. For subjective cases, it matched at least one rater 100% of the time and never returned out-of-range scores. Four independent clinicians validated the assessments and indicated willingness to adopt the system. We argue that principled uncertainty quantification and clinician-aligned explainability are the critical bridges moving automated assessment from technical demonstration to a deployable clinical tool.

Tamim Ahmed, Thanassis Rikakis

Bioemdical Engineering, University of Soutehrn California,, Los Angeles, 90007, CA, USA.

Clinical Decision-MakingMultimodal Model InterpretabilityMultimodal FusionUncertainty Calibration
AskPDF
Feb 10, 2026

In-Hospital Stroke Risk-State Classification from PPG-Derived Hemodynamic Features

The scarcity of temporally aligned pre-event physiological data limits the study of stroke risk states before documented clinical recognition. We focus on patients who experienced stroke during hospitalization while undergoing continuous monitoring, enabling retrospective analysis of pre-anchor photoplethysmography (PPG). Using MIMIC-III and MC-MED, an LLM-assisted pipeline generated candidate stroke anchors from unstructured notes. All retained anchors underwent physician adjudication, and a stratified 100-case double-review audit showed 95.0% candidate-to-adjudicated agreement within +/-15 min. We identified 176 MIMIC-III patients and 154 MC-MED patients with eligible synchronized pre-anchor PPG. A fixed 17-channel hemodynamic representation and ResNet-1D classifier were evaluated using patient-level five-fold internal validation and frozen external testing. At validation-selected operating points, F1-scores were 0.7956, 0.8759, and 0.9406 for 4-, 5-, and 6-hour horizons in MIMIC-III and 0.9256, 0.9595, and 0.9888 in MC-MED; the corresponding threshold-free AUCs ranged from 0.6124 to 0.7079. The PPG model achieved higher F1 than four non-waveform clinical and structured-EHR comparators in all six cohort-horizon settings. On high-risk non-stroke controls, window-level false-positive rates ranged from 0.1371 to 0.2938 and decreased to 0.0183-0.1739 after persistence aggregation. These retrospective findings support measurable pre-anchor PPG structure, but do not establish biological stroke onset, a calibrated bedside alarm, or a clinically validated prediction lead time.

Jiaming Liu, Cheng Ding, Jian Wu +2

College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics, China · State Key Laboratory of Transvascular Implantation Devices and TIDRI, Hangzhou 310009, China · Zhejiang Key Laboratory of Medical Imaging Artificial Intelligence, Hangzhou 310058, China +1

PhotoplethysmographyClinical Risk Prediction
AskPDF