Cardiolpm
Home › Educational Guides › Quality Systems
Quality Systems

ISO 13485 & IEC 62304 Quality Lifecycle for Cardiovascular AI Platforms

Building traceable software requirements, SOUP governance, and hazard controls across agile deep learning cardiology teams.

Connecting ISO 13485 QMS to ML Engineering

ISO 13485 provides the overarching Quality Management System (QMS) governing design controls, supplier qualification, document approval, and corrective/preventive actions (CAPA) for medical device companies. In AI cardiology platforms, design history files (DHF) must capture not just application source code, but dataset curation protocols, labeling SOPs, and model training reproducibility.

Integrating lightweight automated checks into CI/CD pipelines allows agile teams to produce immutable verification artifacts without stalling clinical feature delivery.

  • Maintain versioned dataset manifests and data-splitting hashes inside the DHF.
  • Link every software requirement bidirectionally to automated verification tests.
  • Track third-party ML libraries as Software of Unknown Provenance (SOUP).
  • Conduct formal Design Reviews before promoting new inference containers.

IEC 62304 Safety Classification (Class A, B, and C)

IEC 62304 categorizes medical device software based on potential patient harm if a software failure occurs. While non-clinical administrative dashboards may qualify as Class A, diagnostic ECG triage engines—where a missed lethal arrhythmia or false critical suppression could delay emergency care—typically classify as Safety Class B (non-serious injury possible) or Class C (death or serious injury possible).

Class B and C demand detailed software architecture diagrams, unit verification, integration testing, and rigorous anomaly tracking prior to release. Segregating high-criticality inference pipelines from non-clinical reporting portals isolates verification scope cleanly.

  • Separate Class B/C signal inference services from Class A billing or analytics UI.
  • Document architectural interfaces and error-handling contracts between modules.
  • Archive signed CI/CD unit, integration, and regression reports per release tag.

ISO 14971 Hazard Analysis for AI Failure Modes

Standard software risk management extends under ISO 14971 to cover machine-learning-specific hazards: domain shift across new ECG hardware, clipped millivolt saturation, adversarial or corrupted DICOM headers, and clinician over-reliance (automation bias).

Each identified hazard links to concrete technical mitigations—such as input range assertions, quality-gating rejections, and clear uncertainty UI indicators—verified objectively in test suites. Residual risk evaluations must confirm overall clinical benefits outweigh remaining software uncertainties.

SOUP Governance for Deep Learning Dependencies

Cardiology AI stacks rely heavily on open-source frameworks (PyTorch, ONNX Runtime, NumPy, SciPy). Under IEC 62304, each library must be cataloged in a SOUP register documenting exact version pins, hardware/OS constraints, known bug advisories, and regression tests proving upgrades never alter floating-point inference outputs.

Deterministic ONNX inference containers with bit-exact golden waveform checks ensure library security patches never silently shift diagnostic sensitivity.

Integrating automated SBOM generation, container vulnerability scanning, and checksum verification inside release pipelines satisfies both IEC 62304 software configuration management and modern hospital cybersecurity audits.

  • Run golden-dataset regression tests on every SOUP runtime or driver upgrade.
  • Document known anomalies and severity classifications in release notes.
  • Verify deterministic numeric outputs across CPU and GPU inference targets.
  • Conduct annual internal QMS audits covering data-labeling and ML workflows.

Core Medical Software Standards & Engineering Deliverables

StandardPrimary ScopeKey Cardiology AI ArtifactAudit Verification Focus
ISO 13485Enterprise Quality ManagementDesign History File (DHF) & CAPA logsControlled SOPs, training records & change review
IEC 62304Software Development LifecycleSRS, Architecture Spec & SOUP RegisterRequirement-to-test traceability & bug disposition
ISO 14971Medical Device Risk ManagementFMEA / Hazard Matrix & Residual Risk ReportMitigation effectiveness against missed arrhythmias
IEC 62366-1Usability EngineeringSummative human-factors study with cliniciansPreventing ECG misreads & alarm misinterpretation
ISO 27001 / SOC 2Information Security ControlsCloud access logs & encryption architectureTenant isolation & ePHI protection

Traceability matrices link clinical user needs down through software units and residual risk evaluations.

Worked Illustration: Traceable Risk Mitigation for Lead-Gain Scaling Error

During ISO 14971 hazard analysis of an ECG ingestion microservice, engineers identify a potential failure where vendor XML exports encode amplitude in microvolts instead of millivolts (1000x scaling mismatch).

  1. Step 1 (Risk Estimation): Unchecked 1000x amplitude scaling could flatten ST morphology or cause false-negative ischemia screening (Severity: Serious; Probability: Occasional without checks).
  2. Step 2 (Software Requirement): Engineers create requirement SRS-ING-042 mandating strict DICOM/XML unit-of-measure parsing and physiological R-wave amplitude bounds checking (0.1 mV to 6.0 mV).
  3. Step 3 (Implementation & SOUP Check): Input validators reject out-of-range waveforms with an explicit 'Calibration Metadata Invalid' warning rather than running silent model inference.
  4. Step 4 (Verification Evidence): Automated unit and integration tests inject synthetic microvolt-scaled files into CI/CD, archiving signed test logs directly into the DHF.

Worked Illustration Summary: Systematic ISO 14971 to IEC 62304 traceability converts a silent scaling hazard into a verified, test-gated architectural safeguard.

ISO 13485 & IEC 62304 Engineering Readiness Checklist

  • Assign documented IEC 62304 Safety Class (A/B/C) per software subsystem.
  • Maintain a bidirectional traceability matrix from user needs to test results.
  • Pin exact container hashes and SOUP dependency versions for reproducible builds.
  • Evaluate AI-specific hazards including signal clipping, drift, and automation bias.
  • Validate usability (IEC 62366-1) of waveform viewers with practicing clinicians.
  • Execute formal change-impact assessment before updating inference runtimes.

Frequently Asked Questions

Can agile DevOps teams comply with ISO 13485 without slowing to a halt?

Yes. Modern eQMS and Git-based traceability workflows automate requirement linking, code-review approvals, and verification test archiving inside standard pull-request pipelines.

What is SOUP in IEC 62304 and why does it matter for AI?

SOUP stands for Software of Unknown Provenance—third-party libraries not developed under medical QMS controls. Teams must document versions, evaluate vulnerabilities, and verify behavior within their runtime.

How does ISO 14971 treat false negatives versus false positives in ECG AI?

Both are analyzed as clinical hazards: false negatives risk delayed intervention for acute arrhythmias, while excessive false positives risk alarm fatigue or unnecessary invasive procedures.

Why is IEC 62366 usability testing critical for cardiology dashboards?

Even accurate AI can cause harm if cluttered UI layouts obscure lead labels, sweep speed, or preliminary status warnings during fast emergency assessments.

Primary Standards & Clinical References

  • ISO 13485: Medical devices - Quality management systems - Requirements for regulatory purposes.
  • IEC 62304: Medical device software - Software life cycle processes.
  • ISO 14971: Medical devices - Application of risk management to medical devices.

Related Technical Guides