PaperlayerView exact revision on arXiv

Source-linked AI summary

OptoAgent: A Trustworthy Multi-Agent Framework for Opportunistic Vision Micro-Screening in Classroom Environments

Toqeer Ali Syed, Ali Akarma, Adeel Ahmad, Hammad Muneer

arXiv:2609.14514v1cs.AI

TL;DR

School screening is episodic, while children’s visual difficulty can persist between screening events. SightSentinel embeds calibrated probes in ordinary reading and accumulates quality-filtered evidence against each child’s baseline for referral; the paper reports no measured performance and makes earlier lead time over scheduled screening its central test.

  • Problem

    Screening is episodic despite continuous visual difficulty, and existing calibrated digital tests still require someone to initiate a session.

    Source: §I · p. 1, A. Interaction design · p. 4

  • Method

    SightSentinel embeds brief probes in ordinary reading, filters encounters, accumulates evidence against a child’s baseline, and limits retained video through local processing and short-lived raw-frame buffers.

    Source: C. Accumulation against a personal… · p. 3, A. Interaction design · p. 4, D. Privacy by architecture · p. 4

  • Results

    No measured performance is reported; the central proposed test is whether repeated ambient encounters raise a referral band before yearly or sporadic screening would detect the same decline.

    Source: §V · p. 6, §VII · p. 6

  • Takeaways & Limitations

    The system reports repeated distance-vision concern and recommends a formal acuity and eye examination, leaving diagnosis to the clinician.

    Source: B. Worked example · p. 4

    Validation must be stratified across child, stimulus, and environment factors that shift measurement, including age, language, optotype set, display size, lighting, and viewing distance.

    Source: §VI · p. 6

Abstract

from arXiv · show

A child with reduced distance vision often does not know that anything is wrong. Children adapt, move closer, and rarely report the difficulty, so the problem can survive years of schooling before an adult notices. School screening addresses part of this, but it runs on a schedule, depends on staffing, and is separated from the classroom moments where the difficulty appears. Smartphone and web-based acuity tests have widened access, yet every one of them still needs somebody to start a test. We present SightSentinel, an architecture that turns a wall display a child already reads from into a recurring screening site. Ordinary educational content carries short calibrated optotype probes, and eight specialized agents divide the work. Perception agents recover viewing distance, recognition accuracy, approach behavior, gaze stability, response latency, and interocular difference from each encounter. A quality agent discards observations taken under bad geometry, poor lighting, or inattention. A longitudinal agent accumulates only the surviving evidence against the child's own baseline, and an orchestrator reports a Vision Concern Score routed through a safety gate whose output range excludes diagnosis, refraction, prescription, and reassurance. The design question is whether many cheap, noisy, well-gated encounters can reach a referral decision that one scheduled test reaches late or misses. We state the formulation, the architecture, a four-stage validation protocol against clinical reference standards, and the conditions under which the approach should be rejected.

I. INTRODUCTION

Children may adapt to reduced distance vision without recognizing a problem, while scheduled screening and self-started digital tests remain episodic. SightSentinel proposes embedding calibrated probes in ordinary reading and accumulating quality-filtered encounters against a child’s own history to support referral, not diagnosis.

I. INTRODUCTION

  • Children may not recognize reduced distance vision, and common signs such as squinting or moving closer have many innocent explanations.The difficulty may therefore escape informal observation.Source: §I · p. 1
  • Screening reviews report uneven coverage and parental awareness, while the difficulty itself is continuous rather than confined to screening events.Referral and follow-up matter alongside the screening event.Source: §I · p. 1
  • Digital acuity tools have shown promise in children, but each still depends on a person deciding to start a test.The cited examples include phone, web-based, at-home, and objective testing –.Source: §I · p. 1
  • The paper identifies a gap between calibrated tests that require deliberate sessions and continuous behavioral observation without stimulus control.Habitual reading distance is informative but cannot alone support a screening decision.Source: §I · p. 1
  • SightSentinel proposes turning ordinary reading at a wall display into measurement through estimated viewing distance and recognition of known symbols.Its central claim to test is that many weak measurements, filtered for validity and compared with a child’s history, could support referral earlier than an annual test.Source: §I · pp. 1–2
  • The system is designed for referral rather than diagnosis, because non-cycloplegic refraction can underestimate children’s refractive error, particularly hyperopia.The introduction also cites unresolved safety, workflow, and equity questions for AI-enabled pediatric ophthalmic systems.Source: §I · p. 2
  • The paper contributes an eight-agent architecture, personal-baseline evidence accumulation, a bounded Vision Concern Score, and a four-stage validation protocol, but reports no measured performance.Quality-weighted temporal decay is intended to make an isolated poor encounter non-actionable.Source: §I · p. 2

II. RELATED WORK AND POSITIONING

The related work spans established screening, digital acuity tests, naturalistic reading signals, and multi-agent systems. SightSentinel positions itself at their intersection: ambient pediatric triage using calibrated stimuli, behavioral features, validity filtering, and a safety gate.

A. Screening programs and instrument-based screening

  • School and instrument-based screening can identify children who might otherwise go undetected, but the encounter remains episodic,,.A child who passes one round may develop or reveal a problem before the next.Source: A. Screening programs and… · p. 2

B. Digital and objective acuity measurement

  • Digital and objective acuity tests are credible in children –, though pragmatic trial results remain heterogeneous and warrant more real-world validation.These systems still produce no data unless someone initiates a test.Source: B. Digital and objective acuity… · p. 2

C. Naturalistic reading signals

  • Reading distance and gaze features offer naturalistic signals, but reading studies target skill rather than acuity, and children move closer for many reasons,.The paper therefore treats these features as priors on validity and concern, not evidence of refractive error.Source: C. Naturalistic reading signals · p. 2

D. Agentic AI and multi-agent coordination

  • Prior agentic systems address planning, disability inclusion, infrastructure, and interactive environments –, not micro-calibrated screening embedded in everyday learning.SightSentinel distributes sensory geometry, stimulus calibration, behavioral priors, validity filtering, and governance across specialized agents under a bounded safety gate.Source: D. Agentic AI and multi-agent… · p. 2

E. What this leaves open

SightSentinel combines controlled micro-stimuli with everyday behavior and compensates for noisy encounters through filtering, arbitration, and repetition. Its encounter model records privacy-limited signals, calibrates optotype size to viewing distance, and reports an interval when only a few probes are available.

E. What this leaves open

  • SightSentinel combines calibrated micro-stimuli with uncontrolled everyday behavior, using validity filtering, multi-agent arbitration, and repetition to handle the resulting noise.This is the proposed combination left open by the preceding work.Source: E. What this leaves open · pp. 2–3

A. Encounter model

  • Each encounter uses a privacy-limited feature vector covering recognition, distance, head pose and approach, gaze, latency, quality and confidence, and display context.A lightweight edge-deployed vision backbone extracts geometric and behavioral features on-device, and raw image buffers expire after processing.Source: A. Encounter model · p. 3
  • Optotype size is set from visual angle and measured distance, so distance error transfers directly into angular-size error.A 10% distance error mislabels the tested acuity level by roughly 10%; the system suppresses encounters when distance uncertainty exceeds tolerance.Source: A. Encounter model · p. 3
  • A staircase adapts optotype difficulty, while the small probe count per encounter yields a screening interval rather than a point acuity.Source: A. Encounter model · p. 3
  • When validated separate-eye interaction is available under the deployment protocol, interocular acuity difference raises referral priority because binocular viewing can mask unilateral loss.Source: A. Encounter model · p. 3

B. Validity before evidence

An encounter contributes evidence only when its combined quality conditions meet a threshold. The multiplicative gate makes any failed condition suppress the encounter’s contribution, preventing a distracted response at uncertain distance from raising vision concern.

B. Validity before evidence

C. Accumulation against a personal baseline

The longitudinal stage discounts older surviving encounter evidence and compares each child with their own record. Its score combines multiple concern signals into one of four referral bands, none of which diagnoses or clears a child.

C. Accumulation against a personal baseline

  • Surviving encounter evidence accumulates with temporal decay, and each child is compared against their own record.The decay factor λ sets how quickly older encounters lose weight.Source: C. Accumulation against a personal… · p. 3
  • The Vision Concern Score balances acuity, distance-dependent performance, behavioral, gaze, latency, interocular, and trend concerns against an uncertainty penalty.The orchestrator is intended to keep fleeting behavioral variance from triggering inappropriate clinical escalation.Source: C. Accumulation against a personal… · p. 3
  • The score maps to four referral bands: Insufficient Evidence, Routine Recheck, Vision Screening Recommended, and Prompt Clinical Review.Source: C. Accumulation against a personal… · p. 3
  • No band names a condition or clears a child; “Insufficient Evidence” describes the data, not the eye.Source: C. Accumulation against a personal… · p. 3

IV. ARCHITECTURE

Eight specialized agents produce candidate outputs, which the orchestrator resolves into a single decision. The ordering makes validity and clinical-safety constraints part of the architecture, not merely score-weight choices.

IV. ARCHITECTURE

  • Eight agents contribute candidate outputs that the orchestrator resolves into a single decision.Source: §IV · p. 4
  • The validity agent sits upstream of accumulation, so unusable behavioral observations cannot outvote the fact that they failed the gate.Source: §IV · p. 4
  • The safety gate sits downstream of the score, so changing score weights cannot produce a clinical claim.Source: §IV · p. 4

A. Interaction design

The wall alternates ordinary reading with brief calibrated probes, making screening evidence part of existing classroom activity. A worked example shows repeated, valid distance-related misses contributing to concern, while diagnosis remains with an eye-care professional.

A. Interaction design

  • Natural Reading Mode yields passive distance, pose, gaze, and approach features, while Micro-Screening Mode inserts a brief calibrated challenge and adds recognition accuracy.Probe frequency is a scheduling choice because probing too often turns the reading wall back into a test.Source: A. Interaction design · p. 4
  • A single encounter contributes only a handful of responses, keeping per-encounter burden near zero while statistical strength comes from accumulation.Source: A. Interaction design · p. 4

B. Worked example

  • In the example, repeated misses at standard classroom distance and success after moving closer survive quality filtering and contribute to rising trend and interocular concern.The agents distinguish a repeated distance change from geometry error or an isolated observation.Source: B. Worked example · p. 4
  • The system reports repeated distance-vision concern and recommends a formal examination; an optometrist or ophthalmologist determines the diagnosis.Source: B. Worked example · p. 4

C. The safety gate

The Clinical Safety Gate bounds the system to referral-oriented output rather than diagnostic claims or clearance. Privacy constraints are structural, limiting what is retained and who receives referral wording.

C. The safety gate

  • The Clinical Safety Gate excludes condition names, dioptric values, and “eyes normal” from system output; confirmation remains clinical.Cycloplegic refraction remains the reference when accommodation would bias a non-cycloplegic estimate.Source: C. The safety gate · p. 4

D. Privacy by architecture

  • Privacy controls include local video processing where hardware allows, a short-lived raw-frame buffer, and durable storage of calibrated features rather than classroom footage.Source: D. Privacy by architecture · p. 4

C. The safety gate

  • Figure 1 places the validity gate before accumulation and the Clinical Safety Gate between the score and human-facing output.Only surviving evidence reaches the accumulator, and a referral band is the only permitted human-facing output.Source: Figure 1 · p. 5

D. Privacy by architecture

  • Identity is stored separately from analytic features, teachers receive no medical inference, and referral wording goes to a guardian through a sanctioned route.The design also rules out attention scoring, behavior monitoring, academic profiling, and advertising from screening data.Source: D. Privacy by architecture · p. 5

E. Distributed checkpoints

With guardian consent and secure identity matching, evidence can merge across distributed checkpoints, raising encounter frequency without lengthening individual interactions. Local validation and temporal state support decentralized observations while keeping raw sensory streams from a central repository.

E. Distributed checkpoints

  • Evidence from multiple checkpoints merges under guardian consent and secure identity matching, increasing encounter frequency without lengthening individual interactions.The distributed design uses local validation, temporal state, and secure communication without exposing raw sensory streams to a central repository.Source: E. Distributed checkpoints · p. 5
  • Distributed agents execute local validation, maintain temporal state, and securely communicate observations to an orchestrator without exposing raw sensory streams centrally.Perceptual and geometric consistency across checkpoints is intended to preserve stimulus geometry, display calibration, and interaction-log integrity.Source: E. Distributed checkpoints · p. 5

V. EVALUATION PROTOCOL

The protocol proposes four stages to test calibration, multimodal and longitudinal fusion, school deployment, and whether repeated ambient encounters detect decline earlier than infrequent screening. The discussion identifies validity, confounding, equity, and security risks that validation and deployment must address.

VI. DISCUSSION AND LIMITATIONS

  • The framework remains untested: the authors state that no experiments have been run and that the protocol describes how it should be tested.Validity gating is identified as load-bearing but least proven; only Stage 1 data can settle the calibration risk.Source: §V · p. 5, §VI · p. 6

V. EVALUATION PROTOCOL

  • Stage 1 compares display-based micro-screening with standardized monocular acuity and clinical refraction, measuring gate rejection, error, reliability, sensitivity, specificity, and AUC.The protocol treats gate rejection rate as a feasibility measure because discarding most observations would require an implausible encounter rate.Source: §V · p. 6
  • Stage 2 compares optotype-only, multimodal, and longitudinal models; if optotype-only matches the full model, the added terms and fusion weights should be removed.The protocol frames that outcome as evidence against retaining the extra components.Source: §V · p. 6
  • Stage 3 compares consented classroom referrals with conventional screening and clinical examination, tracking clinical, operational, workload, cooperation, and privacy outcomes.Clinical outcomes include missed cases and false-positive referrals; operational measures include coverage and cost per confirmed referral.Source: §V · p. 6
  • Stage 4 tests whether repeated ambient encounters detect deterioration earlier than yearly or sporadic screening; a null result refutes the paradigm.The protocol treats this lead-time comparison as decisive regardless of the outcomes of Stages 1–3.Source: §V · p. 6
  • Privacy utility and acceptability require parallel evaluation, including on-device extraction versus permissive retention and stakeholder study across affected groups.The proposed stakeholder study includes children, families, teachers, administrators, and receiving eye-care professionals.Source: §V · p. 6

VI. DISCUSSION AND LIMITATIONS

  • Missed symbols have multiple possible causes, and classroom behavior cannot distinguish attention, literacy, fatigue, glare, ocular disease, and refractive error.Validity-gate miscalibration could either discard observations or accumulate noise, while safe supervised occlusion is required for the asymmetry agent.Source: §VI · p. 6
  • Validation must be stratified across child, stimulus, and environmental factors; the framework also complements established screening and does not replace professional eye care.The listed factors include age and refractive status, reading language and optotype set, display size, lighting, and viewing distance; security risks include denial-of-service, spoofing, and tampering.Source: §VI · p. 6

VII. CONCLUSION

SightSentinel changes when screening happens by embedding calibrated probes in ordinary reading and accumulating valid evidence against each child’s record. Its central hypothesis—that repeated gated encounters detect deterioration earlier than infrequent screening—remains untested, with a specified protocol and rejection criterion.

VII. CONCLUSION

  • SightSentinel embeds calibrated probes in ordinary reading, filters encounters through a validity gate, and accumulates evidence against each child’s record until referral is justified.Its output is constrained to a referral band rather than a diagnosis.Source: §VII · p. 6
  • The central hypothesis is that many weak gated encounters detect deterioration earlier than an infrequent strong test.The specified protocol includes outcomes that would tell the authors to abandon the approach.Source: §VII · p. 6
  • Until the proposed studies run, SightSentinel remains an architecture and a hypothesis rather than an experimentally evaluated system.The paper identifies earlier detection relative to infrequent screening as the empirical question on which the paradigm stands or falls.Source: §VII · p. 6
Loading 2609.14514v1…