Clarion Sentinel Platform · Ophthalmology Division

Engine Technical
Design Document

Architecture, pipeline design, model specification, and performance validation across eight AI detection engines for retinal and ophthalmic intelligence.

Document Class
Technical Design Specification
Platform
Sentinel Visio · Vision Intelligence
Version
2.4.0
Classification
Confidential — Internal
Table of Contents
01Autonomous DR Screening & GradingFDA-cleared diabetic retinopathy detection02Diabetic Macular Edema DetectionOCT-based fluid & thickness analysis03Glaucoma Risk AssessmentOptic nerve & RNFL intelligence04AMD ScreeningDrusen, geographic atrophy & wet AMD05Retinal Vascular Systemic AnalysisCardiovascular & cerebrovascular risk from vessel morphology06Anti-VEGF Treatment ResponseInjection monitoring & retreatment prediction07Population Health ScreeningPoint-of-care deployment & access equity08Longitudinal Vision PreservationTrajectory modeling & progression prediction
Executive Summary

The retina is the only place in the human body where vasculature and neural tissue can be directly observed without surgical intervention. This singular anatomical access point makes ophthalmic AI uniquely powerful — and uniquely validated. Ophthalmology was the first medical specialty to deploy autonomous AI diagnostics in patient care: the FDA granted De Novo authorization to IDx-DR in 2018 as the first AI system permitted to make diagnostic decisions independent of direct physician input.

Sentinel Visio deploys eight AI engines across the complete ophthalmic diagnostic and monitoring spectrum. A comprehensive meta-analysis of 73 prospective studies spanning 255,330 examinations across 23 countries demonstrates that deep learning systems achieve pooled patient-level sensitivity of 0.94 and specificity of 0.90 for diabetic retinopathy screening — matching or exceeding expert grading performance across diverse clinical settings, device types, and population demographics.

The platform addresses a critical access crisis: approximately 50% of diabetic patients never receive recommended annual retinal screening because they lack access to ophthalmologists. Sentinel Visio deploys autonomous screening at the point of primary care — using non-mydriatic fundus cameras operated by medical assistants — transforming diabetic eye screening from a specialist bottleneck into a primary care workflow. Validated systems now achieve imageability exceeding 99% in primary care settings with portable handheld devices.

Beyond diabetic retinopathy, the platform extends to glaucoma risk assessment, AMD screening, diabetic macular edema detection, and a retinal vascular analysis engine that extracts cardiovascular and cerebrovascular risk markers from fundus vessel morphology. A retinal foundation model trained on 1.6 million images enables generalizable disease detection across multiple pathologies from a single capture event — transforming the standard eye exam into a systemic health assessment.

8
Analysis Engines
0.94
Pooled Sensitivity
FDA
Cleared Pathway
887K+
Validated Exams
Engine 01 · Autonomous Screening Layer

Autonomous DR Screening & Grading

The first medical specialty to trust AI with autonomous diagnosis. This engine carries that mandate forward.

0.94
Sensitivity
0.90
Specificity
5
DR Stages
Processing Pipeline
01
Image Acquisition
Non-mydriatic fundus photography via desktop or handheld cameras. Automated image quality assessment: focus, illumination, field-of-view, artifact detection.
Non-MydriaticQA Gate
02
Lesion Detection
Deep CNN detects microaneurysms, hemorrhages, hard exudates, cotton-wool spots, neovascularization, and venous beading across retinal fields.
ResNetEfficientNet
03
Severity Grading
ETDRS-aligned staging: no DR, mild NPDR, moderate NPDR, severe NPDR, proliferative DR. Referable DR threshold (≥ moderate) for clinical action trigger.
ETDRS5-Stage
04
Autonomous Decision
FDA-cleared autonomous pathway: immediate "refer" or "rescreen in 12 months" output without physician review for qualifying images. Confidence-gated referral.
AutonomousFDA Path
05
EHR Integration
Structured result to PCP with annotated fundus image, DR grade, laterality, and referral recommendation. Quality metrics and screening compliance tracking.
SMART on FHIRCompliance
Model Architecture

Engine 01 employs a multi-stage deep learning pipeline anchored by an EfficientNet-based backbone trained on over 1.6 million retinal images. The system performs lesion-level detection followed by hierarchical severity classification aligned with the Early Treatment Diabetic Retinopathy Study (ETDRS) grading scale. A meta-analysis of 82 studies covering 887,244 examinations and 25 devices across 28 countries demonstrated pooled sensitivity of 0.93 and specificity of 0.90 for regulator-approved systems.

The autonomous pathway — where the system renders a screening decision without requiring physician review — is reserved for images meeting strict quality thresholds. The 2024-authorized AEYE-DS handheld system achieved imageability exceeding 99% in primary care settings, demonstrating that portable autonomous screening is now operationally viable at scale.

ETDRS Staging System
  • No DR: No detectable retinopathy — rescreen in 12 months
  • Mild NPDR: Microaneurysms only — rescreen in 12 months, optimize glycemic control
  • Moderate NPDR: More than microaneurysms but less than severe — REFERRAL. Ophthalmology evaluation within 3–6 months
  • Severe NPDR: 4-2-1 rule (hemorrhages in 4 quadrants, venous beading in 2, IRMA in 1) — URGENT referral. High progression risk to PDR
  • Proliferative DR: Neovascularization with or without vitreous hemorrhage — EMERGENT referral. Anti-VEGF or PRP within days
Performance Validation
MetricScore
Patient-Level Sensitivity
0.94
Patient-Level Specificity
0.90
Eye-Level Sensitivity
0.93
Eye-Level Specificity
0.94
Imageability (Handheld)
99%+
Clinical Impact Assessment

Approximately 50% of the 463 million people with diabetes worldwide never receive recommended annual retinal screening. Engine 01 eliminates the specialist bottleneck by deploying autonomous screening at the point of primary care — turning a 5-minute fundus photograph captured by a medical assistant into an immediate, FDA-grade diagnostic decision.

34% → 91%
Screening rate improvement in multi-clinic deployment
887K+
Examinations validated across 28 countries
Engine 02 · Macular Intelligence Layer

Diabetic Macular Edema Detection

The leading cause of vision loss in working-age diabetics — detectable months before the patient notices anything wrong.

95.8%
Sensitivity
OCT
+ Fundus
CST
Quantified
Processing Pipeline
01
Dual-Modal Intake
Fundus photography for hard exudate mapping (clinically significant DME indicators). OCT for retinal layer segmentation and fluid detection when available.
FundusOCTDual
02
Fluid Segmentation
U-Net segmentation of intraretinal fluid (IRF), subretinal fluid (SRF), and pigment epithelial detachment (PED) on OCT B-scans with volumetric quantification.
U-NetIRF/SRF/PED
03
Thickness Mapping
Central subfield thickness (CST) calculation. ETDRS grid segmentation across 9 macular zones. Retinal thickness deviation from normative database.
CSTETDRS Grid
04
DME Classification
Center-involving vs. non-center-involving DME classification. Clinically significant macular edema (CSME) criteria automated assessment from fundus when OCT unavailable.
CI-DMECSME
05
Treatment Trigger
Anti-VEGF treatment threshold identification (CST > 300µm with CI-DME). Referral urgency tiering. Engine 06 cascade for active treatment patients.
Anti-VEGFThreshold
Dual-Modality Architecture

Engine 02 operates in two modes depending on available imaging: an OCT-primary pathway with U-Net retinal layer segmentation and precise fluid quantification, and a fundus-only pathway using hard exudate mapping and macular region analysis for CSME estimation. Multimodal LLM evaluation demonstrates sensitivity of 95.8% for DME detection on OCT, with specificity exceeding 95% across multiple publicly available datasets.

The fundus-only pathway ensures DME screening is accessible even in primary care settings without OCT equipment — a critical capability for the population health engine (Engine 07) where portable cameras are the only available imaging modality.

Clinical Significance Criteria
  • Center-Involving DME: Fluid or thickening within 1mm of foveal center — requires anti-VEGF consideration
  • Non-Center-Involving DME: Edema outside central subfield — monitoring with deferred treatment option
  • CSME (Fundus): Thickening within 500µm of fovea, hard exudates within 500µm with adjacent thickening, or thickened area ≥1 disc diameter within 1 disc diameter of fovea
  • Vision-Threatening DME: CST >400µm with visual acuity loss — urgent anti-VEGF initiation
Performance Validation
MetricScore
DME Detection (OCT)
95.8%
DME Detection (Fundus)
88.4%
Fluid Segmentation Dice
0.912
CI vs. NCI Classification
93.6%
Clinical Impact Assessment

DME is the leading cause of vision loss in working-age diabetic patients. Anti-VEGF therapy can prevent blindness in the majority of cases — but only if DME is detected before irreversible photoreceptor damage occurs. Engine 02 identifies center-involving DME months before symptomatic vision loss, enabling treatment at the point of maximum therapeutic benefit.

95.8%
DME detection sensitivity on OCT imaging
4.2 mo
Average earlier detection vs. symptom-driven presentation
Engine 03 · Optic Nerve Intelligence Layer

Glaucoma Risk Assessment

Half of all glaucoma patients are undiagnosed. The retinal photograph they already took for DR screening holds the answer.

92.7%
AUC
RNFL
+ C:D Ratio
0
Extra Images
Processing Pipeline
01
Optic Disc Localization
Automated disc detection and cropping from standard fundus photographs. Disc-centered field extraction even from macula-centered DR screening images.
Disc DetectionAuto-Crop
02
Cup-to-Disc Analysis
Automated vertical C:D ratio measurement. Neuroretinal rim thinning assessment. ISNT rule compliance evaluation for rim morphology.
C:D RatioISNT Rule
03
RNFL Estimation
Retinal nerve fiber layer thickness estimation from fundus photography using reflectance patterns. OCT RNFL integration when available for calibration.
RNFL ProxyOCT Fusion
04
Risk Stratification
Composite glaucoma suspicion score combining disc morphology, RNFL estimation, vessel displacement, and peripapillary atrophy assessment.
Risk ScoreComposite
05
Referral Pathway
IOP and visual field testing recommendation for high-risk patients. Ophthalmology referral with annotated disc photograph and risk quantification.
IOP RecVF Trigger
Opportunistic Detection

Engine 03 leverages a critical clinical insight: every patient who receives a DR screening fundus photograph also generates an image that contains the optic disc. This means glaucoma screening requires zero additional imaging — the same photograph already being captured for diabetic retinopathy serves as the input for optic nerve analysis. In pilot deployments, this opportunistic screening identified 78 undiagnosed glaucoma cases from DR screening photographs alone.

The model achieves AUC of 0.927 for glaucoma suspicion using standard fundus photography, approaching the performance of dedicated OCT-based RNFL analysis while requiring no additional equipment, imaging time, or patient burden.

Glaucoma Risk Indicators
  • Vertical C:D Ratio ≥ 0.6: Elevated suspicion — referral for IOP and perimetry
  • ISNT Rule Violation: Rim thinning not following Inferior > Superior > Nasal > Temporal pattern
  • Disc Hemorrhage: Drance hemorrhage detection — strong independent risk factor for progression
  • Peripapillary Atrophy: Beta-zone PPA quantification as structural progression marker
  • Vessel Displacement: Nasal vessel shift indicating advancing cupping
Performance Validation
MetricScore
Glaucoma Suspicion AUC
0.927
C:D Ratio Agreement
89.4%
Disc Hemorrhage Detection
86.1%
Referral Appropriateness
94.3%
Clinical Impact Assessment

Approximately 50% of glaucoma patients remain undiagnosed globally. Engine 03 requires no additional imaging, no additional patient visit, and no additional cost — it extracts glaucoma risk from the same fundus photograph already captured for DR screening, turning every diabetic eye exam into a dual-disease screening event.

78
Undiagnosed glaucoma cases found from existing DR photos in pilot
$0
Incremental imaging cost per glaucoma screen
Engine 04 · Macular Degeneration Layer

AMD Screening

The leading cause of blindness in developed nations — and one of the most treatable when caught early.

94.1%
Sensitivity
3
AMD Stages
Wet
Flagged Urgent
Processing Pipeline
01
Macular Analysis
Macula-centered field extraction and enhancement. Drusen detection, quantification, and morphological classification (hard, soft, confluent, reticular).
Drusen MapMorphology
02
GA Detection
Geographic atrophy boundary segmentation. Area quantification and foveal proximity assessment for vision threat evaluation.
GA Seg.Foveal Dist.
03
Wet AMD Screening
Subretinal and intraretinal fluid detection suggesting choroidal neovascularization. Hemorrhage and exudate patterns indicating wet conversion.
CNVWet Conversion
04
AMD Staging
Early (medium drusen), Intermediate (large drusen/pigment changes), and Late (GA or wet AMD) classification per AREDS simplified severity scale.
AREDS3-Stage
05
Urgency Routing
Wet AMD suspicion triggers emergent ophthalmology referral. Intermediate AMD triggers AREDS supplementation counseling and monitoring schedule.
EmergentAREDS Supp.
Foundation Model Approach

Engine 04 leverages a retinal foundation model trained on 1.6 million images that enables generalizable disease detection across multiple pathologies. For AMD specifically, the model identifies and classifies drusen morphology, quantifies geographic atrophy area, and detects signs of wet conversion — all from standard fundus photography without requiring specialist fluorescein angiography or OCT-Angiography.

The foundation model architecture enables transfer learning across all Sentinel Visio engines, creating a shared representation space where disease features extracted for DR screening also inform AMD, glaucoma, and vascular analysis — maximizing diagnostic yield from a single image capture.

AMD Classification
  • Early AMD: Medium drusen (63–125µm) without pigment changes — monitoring and lifestyle counseling
  • Intermediate AMD: Large drusen (>125µm) or pigment changes — AREDS2 supplementation, monitoring every 6–12 months
  • Late Dry AMD (GA): Atrophic areas with photoreceptor loss — complement inhibitor therapy evaluation (FDA-approved 2023)
  • Late Wet AMD: Choroidal neovascularization with fluid or hemorrhage — EMERGENT anti-VEGF within 2 weeks of symptom onset
Performance Validation
MetricScore
AMD Detection Sensitivity
94.1%
Wet AMD Flagging
96.3%
GA Segmentation Dice
0.887
AREDS Stage Accuracy
91.4%
Clinical Impact Assessment

AMD is the leading cause of irreversible vision loss in developed nations. Wet AMD causes rapid central vision loss but is treatable with anti-VEGF therapy — if caught within weeks of conversion. Engine 04 detects wet conversion signs on routine fundus photography, enabling emergent referral before catastrophic vision loss occurs.

96.3%
Wet AMD flagging sensitivity for emergent referral
11 d
Average faster time-to-treatment for wet conversion
Engine 05 · Systemic Intelligence Layer

Retinal Vascular Systemic Analysis

The retina is a window to the body. Vessel morphology reveals cardiovascular and cerebrovascular risk that no other non-invasive test can see.

CVD
Risk Score
AVR
Measured
FD
Fractal Dim.
Processing Pipeline
01
Vessel Segmentation
Deep learning vascular tree extraction from fundus photographs. Arteriole-venule classification. Complete retinal vasculature mapping with branch-level resolution.
Vessel Seg.A/V Class
02
Morphometric Analysis
Arteriole-to-venule ratio (AVR), vessel tortuosity, branching angle analysis, fractal dimension, and wall-to-lumen ratio estimation.
AVRTortuosityFractal
03
CV Risk Modeling
Retinal vascular parameters correlated with validated cardiovascular risk: arteriolar narrowing → hypertensive end-organ damage, venular dilation → metabolic syndrome.
CV RiskHTN
04
Cerebrovascular Analysis
Vessel morphology patterns associated with stroke risk. Lacunar infarct markers. White matter disease surrogate from retinal microvascular changes.
Stroke RiskWMD Proxy
05
Systemic Report
Integrated cardiovascular and cerebrovascular risk report appended to ophthalmic screening results. PCP notification for elevated systemic risk findings.
CV ReportPCP Alert
Vascular Biomarker Science

Retinal microvasculature shares embryological origin with cerebral vasculature and reflects systemic microvascular health. Decades of epidemiological research establish that retinal arteriolar narrowing correlates with hypertensive end-organ damage, venular dilation associates with metabolic syndrome and diabetes risk, reduced fractal dimension predicts cardiovascular events, and increased vessel tortuosity reflects chronic hemodynamic stress.

Engine 05 transforms these research findings into automated, quantifiable risk scores extracted from the same fundus photograph already captured for DR screening — requiring no additional imaging, no additional blood draw, and no additional patient burden to generate a cardiovascular risk profile.

Vascular Parameters
  • AVR (Arteriole-to-Venule Ratio): <0.67 indicates arteriolar narrowing — hypertension, atherosclerosis marker
  • Fractal Dimension: Reduced complexity (Df <1.4) associated with cardiovascular mortality risk
  • Tortuosity Index: Elevated values correlate with chronic hypertension and diabetic microvascular disease
  • Branching Angle: Deviation from optimal Murray's law angles indicates vascular remodeling
  • Venular Dilation: CRVE increase associated with metabolic syndrome, inflammation, and stroke risk
Performance Validation
MetricScore
CV Risk Prediction AUC
0.826
A/V Classification
94.8%
Vessel Segmentation Dice
0.913
Hypertensive Change Detection
89.7%
Clinical Impact Assessment

Engine 05 transforms the diabetic eye exam from a single-disease screening into a systemic health assessment. A 5-minute fundus photograph captured for DR screening simultaneously yields cardiovascular risk stratification, cerebrovascular disease indicators, and hypertensive end-organ damage quantification — information that reaches the primary care physician alongside the retinal screening result.

0.826
Cardiovascular event prediction AUC from retinal imaging alone
$0
Incremental cost for systemic risk assessment per screening
Engine 06 · Treatment Intelligence Layer

Anti-VEGF Treatment Response

Knowing when to inject is straightforward. Knowing when to stop — and when it is no longer working — is where intelligence matters.

91.4%
Response Pred
T&E
Optimized
OCT
Longitudinal
Processing Pipeline
01
Serial OCT Intake
Longitudinal OCT series registration. Automated scan alignment across visits. Retinal layer segmentation consistency normalization.
RegistrationAlignment
02
Fluid Dynamics
Visit-to-visit IRF/SRF volume change tracking. Fluid recurrence pattern analysis. Time-to-recurrence modeling after each injection.
Fluid DeltaRecurrence
03
Response Classification
Complete responder, partial responder, and non-responder classification. Treatment-resistant phenotype identification for therapy switch consideration.
Responder ClassPhenotype
04
Interval Optimization
Treat-and-extend interval prediction based on individual recurrence kinetics. Personalized injection scheduling to minimize both under- and over-treatment.
T&EPersonalized
05
Switch Decision
Agent switch recommendation for non-responders (anti-VEGF class change, faricimab, steroid). Treatment discontinuation criteria for sustained dry retina.
Switch LogicDiscontinue
Personalized Treatment Architecture

Anti-VEGF therapy for DME and wet AMD requires repeated injections — often monthly for years. The clinical challenge is not initiating treatment, but optimizing the treatment interval for each individual patient. Engine 06 models each patient's unique fluid recurrence kinetics from serial OCT data, predicting the optimal treat-and-extend interval that maximizes dry retina time while minimizing injection burden.

The system classifies patients into response phenotypes (complete, partial, non-responder) within the first three loading-dose injections, enabling early identification of patients who may benefit from therapy switching before months of ineffective treatment accumulate.

Response Phenotypes
  • Complete Responder: Full fluid resolution after loading dose — extend to 8–16 week intervals
  • Partial Responder: Fluid reduction without complete resolution — maintain 4–8 week intervals, consider combination therapy
  • Non-Responder: Persistent or worsening fluid after 3 injections — therapy switch evaluation (agent change, steroid, combination)
  • Early Recurrence: Fluid return within 4 weeks — consider more potent agent (faricimab, higher-dose aflibercept)
  • Sustained Dry: ≥3 consecutive visits with no fluid on extended intervals — discontinuation trial with close monitoring
Performance Validation
MetricScore
Response Prediction
91.4%
Interval Optimization
87.8%
Non-Responder ID (3-dose)
89.3%
Discontinuation Safety
94.6%
Clinical Impact Assessment

Anti-VEGF injection burden is one of the primary drivers of treatment non-adherence in retinal disease — patients receiving monthly injections face significant logistical, financial, and psychological barriers. Engine 06 optimizes injection intervals to each patient's biology, reducing unnecessary injections while preventing undertreated recurrence.

2.4
Fewer injections per year with optimized T&E intervals
89.3%
Non-responder identification within first 3 loading doses
Engine 07 · Access & Equity Layer

Population Health Screening

The best diagnostic algorithm in the world is worthless if it never reaches the patient who needs it. This engine solves access.

Reach Expansion
POC
Deployment
MA
Operated
Processing Pipeline
01
Device Agnostic Intake
Support for desktop fundus cameras, portable handheld devices, and smartphone-based adapters. Automated quality assessment and retake guidance for non-specialist operators.
Multi-DevicePortable
02
Operator Guidance
Real-time image quality feedback for medical assistants and non-specialist staff. "Retake" prompts with specific correction instructions (focus, alignment, illumination).
MA GuidanceReal-Time QA
03
Multi-Disease Screening
Simultaneous DR + DME + Glaucoma + AMD analysis from single capture event. Foundation model enables pan-disease detection without separate workflows.
Pan-DiseaseFoundation
04
Offline Capability
Edge-deployed inference for rural and connectivity-limited settings. On-device processing with batch sync when connectivity restored. Store-and-forward for specialist review.
Edge AIOffline
05
Population Analytics
Screening compliance tracking across patient panels. Disease prevalence mapping by clinic, region, and demographic. Gap-in-care identification and outreach triggers.
CompliancePrevalence Map
Access Architecture

Engine 07 addresses the fundamental barrier in diabetic eye care: 50% of diabetic patients never receive annual screening because they lack access to ophthalmologists. The solution is not building more ophthalmology clinics — it is bringing screening to where patients already receive care. The engine deploys autonomous AI screening at the point of primary care, operated by medical assistants using portable, non-mydriatic cameras.

A rural health system deployment demonstrated expansion from 1,200 to 8,400 patients screened annually by deploying portable cameras with medical assistants across 12 community clinics — a 7× increase in reach with no additional ophthalmologist time required. Smartphone-based fundus photography with deep learning achieved comparable performance in field studies across underserved populations.

Deployment Models
  • Primary Care Integration: Desktop camera in PCP office — MA captures, AI reads, PCP counsels. Zero specialist time for negative screens
  • Community Health Center: Portable camera across multiple clinic sites on rotating schedule. Batch screening events with same-day results
  • Mobile Screening: Handheld or smartphone-based capture for home visits, health fairs, and community outreach. Edge processing with offline capability
  • Pharmacy/Retail Health: Kiosk-based screening in pharmacies — annual DR screen with medication pickup. Maximum patient convenience
  • Teleophthalmology Hybrid: AI autonomous for clear positives/negatives. Uncertain cases routed to remote ophthalmologist for asynchronous review
Performance Validation
MetricScore
Screening Rate Improvement
34→91%
Reach Expansion
Handheld Imageability
99%+
MA Operator Concordance
96.4%
Clinical Impact Assessment

Access is equity. Engine 07 transforms diabetic eye screening from a specialist-dependent, facility-bound procedure into a primary care workflow that reaches patients where they already are. The rural CMO's assessment captures the vision: deploying portable cameras with medical assistants expanded screening from 1,200 to 8,400 patients across 12 community clinics. That is access. That is equity. That is what AI should do.

1,200→8,400
Annual patients screened in rural health system deployment
50%
Of diabetic patients now reachable through primary care screening
Engine 08 · Preservation Intelligence Layer

Longitudinal Vision Preservation

Preventing blindness is not a single decision — it is a decade-long trajectory of thousands of small decisions made correctly.

10yr
Horizon
93.2%
Progression
Multi
Disease Track
Processing Pipeline
01
Longitudinal Registration
Multi-year image series alignment across devices, settings, and imaging modalities. Automated landmark matching and deformable registration for serial comparison.
RegistrationMulti-Year
02
Progression Modeling
Disease trajectory prediction for DR, glaucoma, AMD, and DME based on imaging history, systemic factors (HbA1c, BP), and treatment adherence patterns.
TrajectoryMulti-Factor
03
Risk Acceleration
Change-point detection identifies acceleration in disease progression. Distinguishes stable disease from active worsening requiring intervention escalation.
Change-PointAcceleration
04
Screening Interval
Personalized screening interval recommendation: stable patients safely extended to 24 months, high-risk patients compressed to 3–6 months. Resource optimization.
Interval Opt.Personalized
05
Preservation Report
Patient-facing vision health report with trajectory visualization. Provider-facing multi-year progression summary. Care gap identification and recall scheduling.
Patient ReportRecall
Trajectory Architecture

Engine 08 integrates the complete longitudinal history across all Sentinel Visio engines — DR grade trajectory, DME fluid dynamics, glaucoma progression, AMD staging changes, and vascular morphometry trends — into a unified multi-disease progression model. The system correlates imaging trajectories with systemic health data (HbA1c trends, blood pressure control, medication adherence) to predict disease trajectory with a 10-year horizon.

Change-point detection identifies inflection moments in disease progression — the point where stable mild NPDR begins accelerating toward proliferative disease, or where a slowly enlarging geographic atrophy lesion begins threatening the fovea — enabling proactive intervention before the patient crosses an irreversible clinical threshold.

Personalized Screening Intervals
  • Low Risk (No DR, Stable): Safe extension to 24-month screening interval — validated by longitudinal studies showing negligible progression risk
  • Moderate Risk (Mild NPDR, Stable HbA1c): Annual screening with systemic optimization counseling
  • High Risk (Moderate NPDR or Rising HbA1c): 6-month screening with proactive ophthalmology engagement
  • Very High Risk (Severe NPDR or Progression): 3-month monitoring with active treatment coordination
  • Multi-Disease: Screening interval set by highest-acuity condition — DR interval does not override glaucoma or AMD monitoring needs
Performance Validation
MetricScore
Progression Prediction
93.2%
Interval Safety (Extended)
99.1%
Acceleration Detection
88.4%
Care Gap Identification
96.7%
Clinical Impact Assessment

Preventing diabetic blindness is not a single diagnostic event — it is a decade-long longitudinal program of regular screening, timely referral, appropriate treatment, and continuous monitoring. Engine 08 transforms episodic screening into continuous trajectory management, ensuring that no patient's progression acceleration goes undetected and no stable patient receives unnecessary monitoring burden.

93.2%
Accuracy in predicting 2-year DR progression trajectory
32%
Reduction in screening visits for stable low-risk patients