MAP

AI image diagnostics can now detect lesions with accuracy equal to or greater than radiologists. However, the rationale behind an AI's "positive" determination often remains a black box.

When a patient files a lawsuit, "because the AI said so" won't hold up. When regulators request an audit, "we have no logs" is unacceptable.

MAP is designed to cryptographically record every AI decision, making "when, what data was reviewed, which model, and why that diagnosis was reached" verifiable in a tamper-evident format.

Why MAP is Needed

Current state and challenges of healthcare AI

Lack of Accountability

"The AI decided" is not a defense in medical malpractice lawsuits. Recording the decision rationale is essential.

Model Drift

Performance changes with retraining. Unable to track which model version made the diagnosis.

Complex Regulatory Compliance

The MDR requires post-market surveillance (Arts. 83–86) and the EU AI Act provides for post-market monitoring of high-risk AI systems (Art. 72; the high-risk rules apply to Annex I systems from 2 Aug 2028); an FDA Predetermined Change Control Plan (PCCP) is optional and pre-specifies planned model changes and how they are validated. MAP (v0.1.2 Working Draft) sets out, provisionally, how records of clinical AI decisions can be made tamper-evident and how a third party can check, for each anchored batch, that no record within the batch's declared scope was omitted after anchoring.

Illustrative Challenge Examples

1 AI Diagnostic Oversight Lawsuit

Patient AI diagnosis: "No abnormality" Cancer discovered 6 months later Lawsuit

Questions Raised:

• "Did the AI actually review the image?"
• "What was the model version?"
• "What was the confidence score at decision time?"

Result: Insufficient evidence, settled out of court

2 Clinical Trial Data Fraud

Third-party testing agency Data falsification FDA Warning Letter Approval revoked

Issues Identified:

• "Unable to prove when data was modified"
• "Audit trail manipulable by administrator"

MAP Scope

Target systems and recording events (concept overview — the v0.1.2 Working Draft covers decision support, imaging, triage, ambient documentation and coverage decisions, and defers surgical-robot control loops)

Target Systems

🩺

AI Diagnostic Support

Image diagnostic AI, ECG analysis AI

Required

🔬

Pathology AI

Digital pathology, cytology AI

Required

💊

Medication Recommendation AI

Drug interaction check, dosage optimization

Required

🏥

Triage AI

Emergency priority determination, ICU alerts

Required

🤖

Surgical Support Robots

Decision support components

Recommended

📊

Clinical Trial AI

Patient screening, endpoint evaluation

Recommended

Recording Events (Causal Chain)

Patient Data Input

Image/Test values

Patient history

Consent info

AI Analysis

Feature extraction

Model inference

Explainability factors

Diagnosis Generated

Classification result

Confidence score

Differential diagnosis

Treatment Plan

Recommended treatment

Risk assessment

Alternative options

Patient Outcome

Actual result

Follow-up

Prognosis data

MAP Architecture

Data flow and event structure

Data Flow

Clinical System

Patient Data

AI Model (SaMD)

Clinical Decision

MAP Recorder (sidecar)

← Records all decision events

Hash Chain

Cryptographically linked

Digital Signature

Ed25519 signed

Model Hash

Version identification

Audit Storage

Encrypted storage (designed for HIPAA/GDPR requirements)

Illustrative Event (early concept — not the MAP v0.1.2 schema)

{
  "event_id": "019234ab-7c8d-7def-8123-456789abcdef",
  "timestamp_ns": 1734567890123456789,
  "event_type": "DIAGNOSIS_GENERATED",
  "facility_id": "HOSPITAL_XXXXX",
  "provenance": {
    "actor": {
      "type": "AI_MODEL",
      "identifier": "chest_xray_classifier_v2.1.3",
      "model_hash": "sha256:abc123...",
      "training_date": "2024-06-15",
      "fda_clearance": "K123456"
    },
    "input": {
      "patient_id_hash": "hmac-sha256:keyed_commitment...",
      "image_study_uid": "1.2.840.113619.2.55...",
      "input_hash": "sha256:def456...",
      "acquisition_timestamp": 1734567800000000000
    },
    "context": {
      "clinical_indication": "CHEST_PAIN_EVALUATION",
      "referring_physician_id": "NPI_XXXXXXXXXX",
      "prior_studies_reviewed": 3,
      "patient_age_range": "60-69",
      "active_protocol": "EMERGENCY_TRIAGE"
    },
    "action": {
      "diagnosis": "PNEUMONIA_SUSPECTED",
      "confidence": 0.87,
      "explainability": {
        "method": "GRADCAM",
        "attention_regions": ["RIGHT_LOWER_LOBE"],
        "contributing_factors": [
          {"factor": "consolidation_pattern", "weight": 0.45},
          {"factor": "air_bronchograms", "weight": 0.32}
        ]
      },
      "differential_diagnoses": [
        {"diagnosis": "ATELECTASIS", "confidence": 0.08},
        {"diagnosis": "LUNG_CANCER", "confidence": 0.03}
      ],
      "recommended_action": "CONFIRM_WITH_CT"
    }
  },
  "prev_hash": "sha256:789xyz...",
  "signature": "ed25519:..."
}

Regulatory Mapping

Correspondence with international regulations

International Regulatory Correspondence

Regulation	Jurisdiction	Requirements	MAP Support
EU AI Act Art. 6(1), Annex I Section A	EU	AI that is a safety component of, or is itself, a device requiring notified-body conformity assessment under MDR/IVDR is high-risk; logging (Art. 12) applies from 2 Aug 2028	Design target (draft)
FDA draft guidance: AI-Enabled Device Software Functions — Lifecycle Management (Jan 2025)	USA	Lifecycle-management and performance-monitoring recommendations (draft; non-binding)	Design target (draft)
FDA PCCP	USA	Predetermined Change Control Plan, model change tracking	Design target (draft)
MDR 2017/745	EU	Medical device traceability	Complementary
21 CFR Part 11	USA	Electronic records/signatures, audit trail	Design target (draft)
HIPAA	USA	PHI protection, access logs	Crypto-Shredding
GDPR	EU	Right to be forgotten, data minimization	Crypto-Shredding
PMD Act — IDATEN change-plan confirmation procedure	Japan	Post-market change plans for AI-enabled SaMD	Planned

FDA 21 CFR Part 11 — Design Targets (Working Draft; not a compliance claim)

Audit Trail

Recorded events linked in a hash chain

Electronic Signature Linkage

Operator identified with Ed25519 signature

Record Integrity Assurance

Tamper detection via cryptographic hash

Timestamp Accuracy

UUID v7 + NTP/PTP synchronization

System Admin Operation Restrictions

External anchoring makes deletion detectable, even by administrators

ALCOA+ Principles Correspondence

ALCOA+ Principle	Description	MAP design target (draft)
Attributable	Who recorded it	actor.identifier + signature
Legible	Readable	Standard JSON format
Contemporaneous	Recorded at time of action	Real-time log generation
Original	Originality	Verifiable via hash chain
Accurate	Accuracy	Verified with input data hash
Complete	Completeness	Event gaps detected via chain breakage
Consistent	Consistency	Linked via trace_id
Enduring	Durability	Tamper-evident, cryptographically verifiable
Available	Availability	Accessible via standard API

Privacy Protection: Crypto-Shredding

Balancing append-only, tamper-evident audit trails with the right to be forgotten

Traditional Dilemma

"Append-Only Audit Trail"

Cannot delete

"GDPR Right to Erasure"

Erasure on request, subject to exceptions (Art. 17(3))

Appears irreconcilable

MAP Solution: Crypto-Shredding

Patient Data

Encryption

Encrypted Data

Record

MAP

Encryption Key → Securely stored in Key Management System

On Deletion Request:

Destroy encryption key → Data becomes unreadable → Rendered unreadable — whether this counts as erasure is jurisdiction- and case-dependent
Hash chain maintained → Audit trail integrity preserved

Use Cases

Specific application scenarios

1 Medical Malpractice Lawsuit Scenario

Phase	Without MAP	With MAP
At Diagnosis	AI determines "No abnormality"	Same + Decision rationale recorded
Lawsuit Filed	"AI missed it" claim	Same
Discovery	Logs potentially tampered	Cryptographically verifiable evidence
Root Cause	"Model problem or data problem unknown"	"Accuracy issue under specific conditions in model v2.1.3" identified
Liability	Settlement with ambiguous liability	Judgment on records whose integrity and completeness can be verified

2 FDA PCCP Audit Scenario

FDA Auditor

PCCP Compliance Verification Request

Evidence that MAP-style records could supply (illustrative)

Model version history ✓
Registered model-version hashes (training lineage is VAP-SLEEP scope) ✓
Tamper-evident performance verification records ✓
Signed change approval records ✓

Audit Evidence Provided

3 Clinical Trial Data Integrity

CRO (Contract Research Organization)

EDC System

MAP Anchoring

Hash recorded to external blockchain

Tampering detectable — even by system administrators

Verifiable data-integrity evidence for regulatory inspection

MLOps Integration

Model-lifecycle provenance is defined by VAP-SLEEP (SMP, v0.2.0 Working Draft); MAP v0.1.2 records model registration and refers to it

Model Lifecycle Recording

Development

MAP Record

Validation

MAP Record

Approval

MAP Record

Deploy

MAP Record

Operation

MAP Record

Retrain

MAP Record

Recording Details:

Training dataset hash

Hyperparameters

Validation results (sensitivity/specificity)

Approver's electronic signature

Deployment environment info

Inference logs

Retraining trigger and differential information

GMLP (Good Machine Learning Practice) Correspondence

GMLP Principle	MAP design target (draft)
Training/Test data independence	Dataset hash enables post-verification
Dataset representativeness	Demographic metadata recorded
Model transparency	Explainability factors (SHAP/LIME/GradCAM) recorded
Continuous performance monitoring	Operational inference results continuously recorded

Technical Specifications Summary

Concept parameters — where they differ from the MAP v0.1.2 Working Draft, the draft prevails

Timestamp Precision

Millisecond (NTP synchronized)

Event Recording Frequency

Per diagnostic event

Hash Algorithm

SHA-256

Signature Algorithm

Ed25519 (Future: Dilithium)

PHI Protection

PHI excluded from events (commitments only); crypto-shredding of protected payloads

Storage Format

Canonical JSON (RFC 8785); FHIR AuditEvent mirroring is an informative sketch

Retention Period

Configurable per regulatory requirements (typically 10-30 years)

Roadmap

MAP development schedule (no dates set for the remaining steps). MAP is currently a v0.1.2 Working Draft; the later milestones have not been reached.

Done — July 2026

MAP Working Draft v0.1 (current: v0.1.2)

Planned — no date set

Technical Validation with Medical Device Manufacturers & Hospitals

Planned — no date set

Information Provision to FDA/PMDA/EMA

Planned — no date set

MAP v1.0 Official Release

Planned — no date set

IHE (Integrating the Healthcare Enterprise) Collaboration

Planned — no date set

Exploring standardization as an HL7 FHIR extension

Relationship with VAP/VSO

Framework hierarchy

VAP (Verifiable AI Provenance)

Cross-domain parent framework for all domains

defines & maintains

VSO (VeritasChain Standards Organization)

Standards organization that develops and maintains VAP

publishes profiles

VCP

Finance

v1.2 Released

DVP

Automotive

Planned

Medical

v0.1.2 Working Draft

EIP

Energy

Planned

PAP

Public

Partial Draft

Get Involved in MAP Development

We welcome participation from medical device manufacturers, hospitals, and regulatory authorities.

View VAP Framework GitHub Join MAP Development

"When AI says 'cancer,' what patients want to know isn't the probability. It's the rationale."

— VeritasChain Standards Organization

"In medicine, trust is not given. It is verified."

MAP Medical AI Profile

Why MAP is Needed

Lack of Accountability

Model Drift

Complex Regulatory Compliance

Illustrative Challenge Examples

1 AI Diagnostic Oversight Lawsuit

2 Clinical Trial Data Fraud

MAP Scope

Target Systems

AI Diagnostic Support

Pathology AI

Medication Recommendation AI

Triage AI

Surgical Support Robots

Clinical Trial AI

Recording Events (Causal Chain)

MAP Architecture

Data Flow

Clinical System

Illustrative Event (early concept — not the MAP v0.1.2 schema)

Regulatory Mapping

International Regulatory Correspondence

FDA 21 CFR Part 11 — Design Targets (Working Draft; not a compliance claim)

Audit Trail

Electronic Signature Linkage

Record Integrity Assurance

Timestamp Accuracy

System Admin Operation Restrictions

ALCOA+ Principles Correspondence

Privacy Protection: Crypto-Shredding

Traditional Dilemma

MAP Solution: Crypto-Shredding

Use Cases

1 Medical Malpractice Lawsuit Scenario

2 FDA PCCP Audit Scenario

Evidence that MAP-style records could supply (illustrative)

3 Clinical Trial Data Integrity

MLOps Integration

Model Lifecycle Recording

Recording Details:

GMLP (Good Machine Learning Practice) Correspondence

Technical Specifications Summary

Timestamp Precision

Event Recording Frequency

Hash Algorithm

Signature Algorithm

PHI Protection

Storage Format

Retention Period

Roadmap

MAP Working Draft v0.1 (current: v0.1.2)

Technical Validation with Medical Device Manufacturers & Hospitals

Information Provision to FDA/PMDA/EMA

MAP v1.0 Official Release

IHE (Integrating the Healthcare Enterprise) Collaboration

Exploring standardization as an HL7 FHIR extension

Relationship with VAP/VSO

VAP (Verifiable AI Provenance)

VSO (VeritasChain Standards Organization)

VCP

DVP

MAP

EIP

PAP

Get Involved in MAP Development