AI-Enabled Cervical Screening: HIHI Pitch

Collaboration with HSE

Python AI Solutions Ltd

2025-10-01

Proposal

AI-assisted cervical screening trained on HSE data to power a global, cloud-deployed diagnostic-decision-support platform.

The Healthcare Need (Ireland → Global)

  • Ireland: high-quality, QA-rich screening data since 2018
  • Global: screening coverage ~33% worldwide; >600k new cases/year; 90% in LMICs (World Health Organization 2020). TAM: 100s of millions of women.
  • Opportunity: privacy preserving training on Irish data → exportable AI to US and LMIC (smartphone enabled).

Our Intervention (What It Does)

  • AI decision support across cytology, colposcopy, and HPV data (multimodal).
  • Human-in-the-loop by design; configurable to jurisdiction and workflow.
  • Privacy enhancing technologies (PETs): train without exposing patient data; models exported, not raw data.

Collaboration Focus (Tier 2)

  • Foreground Tier 2: expert collaboration to interpret data and improve training and labeling fidelity.
  • Tier 1 via PETs remains feasible if needed (minimal IT effort; privacy-preserving access, secure enclave).

Clinical Workflow & Use Cases

  • In-scope steps: HPV triage, cytology review, colposcopy support, QA, recall orchestration.
  • Automation vs human: prioritize cytology and colposcopy assistance; always clinician-overridable.
  • HSE focus: consume data and expertise to produce an exportable support system.
  • Colposcopy AI expected to exceed human expertise and cytology AI expected to match.

Data & PETs (Privacy by Design)

  • Data types: cytology/histology Whole Slide Imaging (WSI), HPV results, reports, demographics (as available).
  • Annotation: as detailed as possible (cell/region/slide, Bethesda system) with expert input.
  • PET toolkit: secure enclave training, de-identification, audit; supports different sharing constraints (Kaissis et al. 2020).
  • Continual learning and feedback…

Integration & IT

  • Interfaces: FHIR/HL7 for LIS/LIMS; DICOM for imaging; REST APIs for services.
  • Deployments: HSE-defined (on-prem/cloud); role-based access, audit, encryption.
  • Vendor-agnostic and standards-first; demo annotation tool already on Kubernetes (HL7 International 2019; DICOM Standards Committee 2024).

Evidence & Metrics

  • Primary metrics: sensitivity/specificity, NPV at fixed sensitivity, time saved.
  • Targets: meet or exceed state-of-the-art; use continual learning to improve on discordant cases.
  • Validation: retrospective testing first.

Regulatory & Governance

  • Frame as research/clinical evaluation initially; not procurement, no funding required from HIHI sites.
  • GDPR-ready: DPIA, DPO engagement, clear controller/processor roles, PETs to minimize risk.
  • Safety: monitoring for drift, bias checks, audit logs, human override, fail-safes.
  • AI system will be developed to meet HPRA/FDA requirements (auditable and explainable).

Pilot Plan (3 months)

  • Weeks 1–3: data schema discovery, secure enclave requirements, SBOM/manifests.
  • Weeks 4–6: enclave setup + approvals; telemetry and compliance protocols.
  • Weeks 7–11: training runs; iterate on architecture from aggregated metrics.
  • Week 12: model extraction, reporting, next-step proposal.

Team & Differentiators

  • Team: clinical advisor, imaging/ML engineers, product, and ops (detail in appendix).
  • Differentiators:
    • Multimodal (cytology, colposcopy, HPV) platform.
    • PET-first training with exportable models.
    • Vendor-agnostic, standards-based integration.

Summary

  • Utilising high volume high quality patient data set to train cloud based AI diagnostic system using privacy enhanced technology.
  • Bringing high quality diagnostics to remote locations in developing world for early cervical cancer screening using Iphone based imaging.
  • Low cost pay per use system for high impact increase in global decentralised cancer screening

The Ask (from HIHI/HSE)

  • Access pathway: data schema, enclave approval, controlled training runs.
  • Clinician time: limited Tier-2 consults (2–4h/participant; total 40–60h across pilot).
  • IT contacts: LIS/LIMS integration points; environment/hosting decision.

Demo & Next Steps

  • Current demo: https://cervical-screening.pythonaisolutions.com (UX/algorithm improving over coming days).
  • Slides hosted via GitHub Pages; PDF export available.
  • Follow-up: Appendix contains deeper methods, governance, and KPIs.

References

DICOM Standards Committee. 2024. DICOM Standard 2024d. https://www.dicomstandard.org/current/.
HL7 International. 2019. FHIR Release 4.0.1. https://hl7.org/fhir/R4/.
Kaissis, Georgios A., Marcus R. Makowski, Daniel R"uckert, and Rickmer F. Braren. 2020. “Secure, Privacy-Preserving and Federated Machine Learning in Medical Imaging.” Nature Machine Intelligence 2: 305–11. https://doi.org/10.1038/s42256-020-0186-1.
World Health Organization. 2020. Global Strategy to Accelerate the Elimination of Cervical Cancer as a Public Health Problem. https://www.who.int/publications/i/item/9789240014107.

Appendix — Q&A: Data Flows

  • Diagram: sources → enclave → training → model export → deployment.
  • Controls: de-identification, access control, audit, encryption.

flowchart LR
  A["HSE Data Sources<br/>HSE: LIS/LIMS, WSI, HPV, reports"] -->|de-identify/ingest| B(("Secure Enclave"))
  B --> C["Approved Training Jobs<br/>(whitelisted containers, SBOM)"]
  C --> D["Aggregated Metrics/Logs<br/>(no raw data egress)"]
  C --> E["Trained Models<br/>(weights only)"]
  E --> F["Export to Deployment<br/>(Ireland / External)"]
  D --> G["Governance & Audit<br/>(DPIA, DPO, REC)"]

  style B fill:#eef,stroke:#66f,stroke-width:2px
  style E fill:#efe,stroke:#6c6,stroke-width:2px
  style G fill:#ffe,stroke:#cc6,stroke-width:2px

Appendix — KPIs & Baselines

  • Clinical: sensitivity/specificity, NPV at fixed sensitivity.
  • Operational: turnaround time, throughput, rescreen rates.
  • Success thresholds: to be co-defined with clinical leads.

Appendix — Interoperability

  • FHIR resources: Patient, Observation, Specimen, DiagnosticReport.
  • HL7 v2 messaging for LIS; DICOM for WSI.
  • Testing: vendor-agnostic validation; K8s demo environment.

flowchart TB
  subgraph HSE
    LIS["LIS/LIMS"]
    WSI["WSI Scanner/PACS"]
    IDP["Identity / SSO"]
  end
  subgraph Platform
    API["REST API"]
    FHIR["FHIR R4 Endpoints"]
    DICOM["DICOMweb"]
  end

  LIS -->|HL7 v2| API
  LIS -->|"FHIR R4<br/>Observation/Specimen"| FHIR
  WSI -->|DICOMweb WADO/QIDO| DICOM
  IDP -->|OIDC/SAML| API

Appendix — Methods (High Level)

  • Model families: WSI foundation, MIL, detection; LLMs for report drafting.
  • Training: PET-enabled; continual learning strategy across sites.
  • Explainability: saliency, exemplars; uncertainty-aware triage.

Appendix — Governance & Risk

  • Approvals: DPO/DPIA/REC as required by HSE.
  • Risks: drift, bias, change management, support load.
  • Mitigations: monitoring, RBAC, logs, safe rollback.

Appendix — IP & Commercial

  • IP: trade secrets for algorithms; potential patents for interfaces.
  • Commercial path: SaaS, tiered pricing; vendor-agnostic partnerships.
  • Post-pilot: HIHI report → adoption path (procurement outside scope).