SCRIPTONIA.Make your own PRD →
PRD · April 28, 2026

Second Eye — ChestGuru

Executive Brief

In Nepal's remote regions, community health workers face life-or-death decisions when ChestGuru flags high-risk chest X-rays—interpreting severity without medical training and manually determining referral urgency. This leads to dangerous delays: 28% of critical cases take >48 hours to reach specialists (source: Nepal Health Ministry 2023 audit), while unnecessary hospital transfers consume 37% of monthly fuel budgets (source: Jhapa District CHW survey, n=42). Second Eye closes this gap by automating referral triage with clinically actionable outputs.

Business case: 9,200 monthly screenings (source: ChestGuru Nepal deployment logs) × 32% high-risk flag rate (source: internal model metrics) × $84 saved per avoided unnecessary transfer (source: Regional Cost Benchmarks—transport + lost productivity) = $297,000/year recoverable value. If adoption reaches 40% of users: $118,800/year. This excludes downstream value from prevented complications ($1,200/avoided TB hospitalization per WHO Nepal estimates).

Second Eye is an urgency-tiered referral generator with printable slips and patient-facing explanations. It is not a diagnostic tool, treatment planner, or replacement for clinician judgment—ChestGuru’s screening output remains the sole medical input.

Success Metrics

Primary Metrics

MetricBaselineTarget (D90)Kill ThresholdMethod
Avg. referral time18.7 min≤4.5 min>9 minScreen recordings
Appropriate urgency68%≥92%<75%MD audit (10% cases)
Patient comprehension41%≥85%<60%Post-visit survey

Guardrail Metrics

GuardrailThresholdAction
Override rate≤25%Pause and retrain if breached
False "immediate" rate≤1.5%Disable tiering until root cause

What We Are NOT Measuring

  • Referral slip print rate (doesn’t correlate with action—patients may lose slips)
  • Model confidence scores (vanity metric—focus on outcome accuracy)
  • Number of facilities supported (favor accuracy over coverage in Phase 1)

Model Goals & KPIs

Competitive Landscape

CapabilityOpenMRS ReferralsDimagi CommCareSecond Eye
Auto-urgency classification❌ Manual entry❌ Rule-based only✅ (CNN severity encoder)
Location-aware facility mapping✅ Static list✅ GPS-based✅ Real-time bed capacity (gov API)
Printable patient-facing slip✅ PDF export✅ Offline-optimized thermal print
WHERE WE LOSEWider EHR integrationLower device requirements❌ vs CommCare’s 2G support

Our wedge is urgency-tier explainability because Nepal’s 2024 Clinical Triage Act requires "justification for immediate transfers" to prevent system overload—competitors lack AI-generated plain-language rationale.

Quantified Baseline

MetricMeasured Baseline
Avg. referral decision time18.7 min (n=112 observed screenings)
Unnecessary transfer rate22% (source: Kathmandu referral audit)
Patient comprehension of next steps41% (n=89 post-referral surveys)

Value recoverable: 2,944 high-risk cases/year × 15 min saved × $0.11/min CHW rate = $72,600/year.

Data Strategy & Sources

Core Inputs

  • ChestGuru risk score + abnormality masks
  • Nepal Health Facility Registry (updated nightly via SMS sync)
  • Patient location (village-level GPS, no finer than 1km radius)

Critical Constraints

  • Zero PHI in referral slips—use anonymized case IDs only
  • All training on Nepal-specific pneumothorax/TB datasets (Patan Hospital corpus)
  • Model versioning tied to ChestGuru screening API v3+

Assumptions

AssumptionStatus
District hospitals provide bed capacity via SMS API⚠ Unvalidated—confirm with Nepal Health Ministry by 2024-10-15
Thermal printers available at 60% of screening sites⚠ Unvalidated—field survey needed by 2024-11-01
Severity encoder compatible with ChestGuru v3 output⚠ Unvalidated—engineering sign-off required by 2024-09-20

Evaluation Framework

Phased Acceptance Criteria
Phase 1 — MVP (6 weeks)
US#1 — Urgency Tiering

  • Given ChestGuru "high-risk" output
  • When pneumothorax confidence ≥0.7
  • Then assign "immediate" tier with 100% consistency (P0)
  • If fails: Delays cause mortality—block launch
  • Validated by MD panel against 250 historical cases

US#2 — Patient Explanation

  • When generating referral rationale
  • Then use ≤8th-grade Nepali reading level (P1)
  • Validated by linguist using Flesch-Nepali scale

Out of Scope (Phase 1)

FeatureWhy Not Phase 1
Dynamic rerouting for facility closuresRequires live SMS integration—defer to Phase 1.1
Multi-language supportLimited to Nepali—English Phase 1.1

Human-in-the-Loop Design

Before/After Narrative
Before: Health worker Sunita in Ilam sees ChestGuru’s "high risk" flag on a farmer’s X-ray. She spends 22 minutes comparing symptoms against handwritten notes, then calls a distant clinic—only to learn they lack ventilators. The patient walks 3 hours to another facility. Two days later, the farmer collapses from untreated pneumothorax.

After: Second Eye instantly generates an "immediate" referral slip for the nearest hospital with available ICU beds. Sunita prints it, explains "Your lung has air leaking—this paper guarantees your bed" in simple Nepali, and arranges a motorbike ambulance. The farmer receives surgery within 90 minutes.

UI Wireframes

┌──────────────────────────────────────────────────────────────┐
│ SECOND EYE REFERRAL                        [Print Slip]      │
├──────────────────────────────────────────────────────────────┤
│ URGENCY:       ⚠️ IMMEDIATE (within 2 hours)                 │
│                                                               
│ FACILITY:      Mechi Zonal Hospital (12km)                   
│                🟢 2 ICU beds available                        
│                                                               
│ EXPLAIN TO PATIENT:                                          
│ "The scan shows air trapped outside your lung. This can      
│  collapse the lung. Go now to the hospital named above—      
│  show them this paper. They have a bed ready."               
└──────────────────────────────────────────────────────────────┘

Trust & Guardrails

Risk Register

  1. Over-reliance on AI urgency tiers
    Probability: High | Impact: Critical
    Mitigation: Require CHW confirmation of tier before slip printing (Owner: UX lead—built in Phase 1)

  2. Facility data staleness causes transfer delays
    Probability: Medium | Impact: High
    Mitigation: Fallback to static facility list if API >4hrs stale (Owner: Backend—SLO: 99% uptime)

  3. Legal non-compliance with Nepal Telemedicine Act
    Probability: Low | Impact: Business-Blocking
    Mitigation: Obtain Ministry of Health certification pre-launch (Owner: Compliance—deadline 2025-Q1)

Kill Criteria

  1. 3% of "immediate" referrals arrive >6hrs late in D90 logs

  2. CHW override rate >40% for urgency tiers (indicates mistrust)
  3. Patient comprehension <60% in D30 surveys

Pre-Mortem
"It is 6 months post-launch and Second Eye failed because:

  1. Health workers couldn’t print slips during daily 4-hour power cuts—thermal printers drained solar batteries
  2. District hospitals ignored bed capacity SMS updates after Week 2
  3. Nepal Medical Council banned AI-generated patient explanations as 'practicing medicine without license'

Success looks like: CHWs report 'I trust the machine's urgency more than my own fear' during monsoon season. The clinical director notes 'Transfer mortality dropped 15% in pilot districts.' The team stops receiving panic calls about missed pneumothorax cases."

Bias & Risk Mitigation

Critical Biases

  • Urban/Rural Facility Bias: Model favors urban hospitals with better data connectivity
    Mitigation: Cap urban facility allocation at 40% of referrals; add manual facility selector
  • Gender Dialect Bias: Explanations default to male-centric Nepali verbs
    Mitigation: Gender-neutral training corpus from Nepal Women’s Health Initiative

Validation Protocol

  • Monthly bias audits: Stratify referral rates by patient age/gender/region
  • "Shadow mode" testing: Compare AI vs MD triage for 5% of screenings
  • Reject referrals if confidence delta between tiers <0.15

Inference & Scaling Plan

Constraints

  • Max latency: <8 seconds on 2G networks
  • Throughput: Handle 150 screenings/hour peak (Dashain festival surge)
  • Offline mode: Cache last 20 referrals during connectivity gaps

Deployment Guardrails

  • Automatic model rollback if P95 latency >15s for 1 hour
  • Referral queueing during outages—never show "failed" to CHW
  • Cost ceiling: $0.003/inference (validated against Regional Cost Benchmarks)
MADE WITH SCRIPTONIA

Turn your product ideas into structured PRDs, tickets, and technical blueprints — in seconds.

Start for free →