--- doc_id: ASTRACBC-OCULAR-EXTRACT title: AstraCBC Method, Limits, Assumptions, and Gap Extraction version: v0.1-draft status: draft owner: Tambua Health Engineering audience: Systems, Clinical, Regulatory, ML device_name: AstraCBC updated_at: 2026-02-15 ---

AstraCBC Extract for Ocular SaMD

1) Scope

This document extracts reusable architecture and governance primitives from:

• `../astracbc-whitepaper/ASTRACBC_PHONE_ONLY_CBC_WHITEPAPER.md`

Priority source sections:

• Sections 4-9 (system model)
• Sections 11-16 (validation, governance, risk)
• Section 19 (prohibited overclaims)

2) Extraction Table

| AstraCBC source area | Reusable for ocular program | Required adaptation | Blocking gap | |---|---|---|---| | App scope boundary | Strict smartphone-only/no hardware | Replace CBC framing with ocular-biomarker claims | Claim granularity not finalized | | Capture + QC pipeline | Guided capture and gate states | Ocular region-specific QC metrics and thresholds | No quantified thresholds | | Feature extraction and inference | Ensemble + arbitration pattern | Ocular feature taxonomy and calibration strategy | No dataset/model card baseline | | Safety outputs | Confidence + abstain + escalation | Ocular condition-specific escalation policy | No endpoint-specific care pathways | | Validation and failure modes | Multi-layer validation logic | Ocular trial protocols and SAP definitions | No paired-label protocol | | Misclassification governance | Conservative abstention precedence | Condition-specific override and triage guardrails | No production ruleset | | Risk register direction | Strong foundation for ISO 14971 | Full hazard-to-control traceability matrix | No complete RMF | | Prohibited claims | Directly reusable | Add ocular wording constraints | Claim-map evidence missing |

3) Methods Extracted (System Core)

3.1 Pipeline pattern

Extracted baseline processing sequence:

`capture -> QC -> feature extraction -> inference -> rules+ML arbitration -> output or abstain`

3.2 Safety model

• Output only when QC passes.
• Abstention preferred over uncertain positive labels.
• Every high-risk label must produce explicit confirmatory-care guidance.

3.3 Change governance

• Versioned model releases.
• Drift surveillance.
• Rollback capability.
• Controlled updates with documentation.

4) Limits Extracted (Must Be Carried Forward)

• Smartphone-only architecture does not create direct specimen chemistry channels.
• Device heterogeneity and capture context are dominant uncertainty drivers.
• Confounding can cause false confidence without strict QC and subgroup surveillance.
• Any direct counting or definitive diagnostic claim is prohibited unless independently evidenced.

5) Assumptions Extracted

1. Guided capture can reduce user variance enough for deployable signal quality.
2. QC gates can enforce safe abstention.
3. Ensemble models plus rule arbitration can constrain unsafe outputs.
4. Longitudinal trends can improve confidence over repeat sessions.
5. Operational safety depends on post-market drift controls.

6) Identified Documentation Gaps

6.1 Data gaps

• No finalized ocular dataset dictionary
• No site/device holdout protocol details
• No subgroup minimum sample definitions

6.2 Quantitative gaps

• No baseline metrics (AUC, sensitivity/specificity, MAE/MAPE, calibration)
• No QC pass/reacquire/abstain rate targets by device tier
• No robustness stress-test acceptance thresholds

6.3 Artifact gaps

• Missing complete CSV claim matrix for ocular claims
• Missing trial protocol and SAP aligned to intended claims
• Missing model cards and release criteria pack

7) Required Ocular Adaptation Plan

1. Define ocular-specific acquisition protocols (conjunctiva, sclera, pupil).
2. Freeze QC metric definitions and numeric pass criteria.
3. Build claim-evidence table and lock wording with legal/regulatory review.
4. Create paired-reference trial design with holdout site and holdout device-family splits.
5. Implement section-level acceptance criteria for all dossier chapters.

8) Extraction Acceptance Checklist

• [ ] Each reused method is mapped to an ocular equivalent.
• [ ] Each limitation appears in risk controls and claim language.
• [ ] All assumptions are testable through protocol or monitoring.
• [ ] Every gap has an owner and due date.
• [ ] Prohibited overclaims are copied into external-content guardrails.