Pilot proposal · Health Service Executive · Ireland

One integrated AI platform for the
acute hospital crisis Ireland is in.

The Med AI Platform is 19 interoperable services covering ED triage, sepsis, deterioration, bed management, oncology risk, waiting-list intelligence and discharge orchestration. Engineered end to end. Locally instrumented. Ready for a HIHI-framed clinical pilot. Aligned to the HSE "AI for Care 2026 to 2030" Year-1 priorities.

The ask: a 12-week pilot at one Model 4 site, zero cost to HSE, falsification clause included.
25,290+
on trolleys, Jan to Feb 2025 (INMO)
750k
on hospital waiting lists (HSE)
64%
waiting longer than target (HSE)
€263M
2026 Digital Health budget (DoH)
Independent submission by Harishankar Somasundaram, MSc Artificial Intelligence (Ireland). No commercial intent today. Offered for evaluation under HIHI, as a research collaboration, or via the HSE AI and Automation Centre of Excellence. Models trained on MIMIC-IV; Irish-data validation is an explicit Phase 1 deliverable.
01 · The crisis

The numbers HSE wakes up to.

These are not vendor projections. They are figures published by INMO, the HSE, the Department of Health and the NTPF in the last 12 months. Every one of them maps to a specific platform module.

Single-day trolley record
501
August 2025. Daily political headlines. Direct INMO reporting.
→ ED Flow · Trolley Watch · Bed Management
Extra ED attendees per day
700+
vs prior year, across the system. Front-door pressure is structural, not seasonal.
→ ED Triage · ED Flow Optimizer
6-hour PET compliance
<70%
at most acute sites. Ireland's national ED target is being missed nationwide.
→ ED Flow Optimizer
Sepsis mortality penalty
4 to 8%
per hour of delayed treatment. Every shift saved is lives recovered.
→ Sepsis ICU · Deterioration (NEWS2)
Why the situation persists

Existing AI deployments in Ireland are point products. Aidoc for radiology at the Mater. eAltra for pre-chemo at St James's. Oneview for patient engagement at CHI. There is currently no integrated AI platform covering ED triage, deterioration, bed management, oncology risk and waiting-list intelligence in a single system. That gap is the one this platform fills.

02 · Strategic alignment

Mapped to the HSE strategies that are already funded.

This is a working platform deliberately engineered against Ireland's three governing health-tech strategies and the National Cancer Strategy 2017 to 2026.

Sláintecare
10-year reform programme
  • Demand & capacity management
  • Reduction of waiting times
  • Care closer to home (virtual ward roadmap)
  • Integrated electronic records (FHIR)
Digital for Care 2024 to 2030
Foundational digital strategy
  • HL7 FHIR R4 interoperability (service app_19)
  • Data governance and DSAR (app_17)
  • Explainable AI for human oversight (app_18)
AI for Care 2026 to 2030
First national AI strategy
  • AI scribe (built, app_10)
  • Demand and Capacity (built, app_03/08/14/16/21)
  • Support function automation (built, app_15)
  • Radiology AI (complementary to Aidoc)
  • Contact centre (out of scope v1)
→ 4 of 5 Year-1 priorities already addressed.
03 · The platform

19 interoperable services. One unified clinical dashboard.

FastAPI microservices, MongoDB durable event log, Kafka (Redpanda) streaming, React + TypeScript dashboard. Every prediction carries a SHAP rationale; every clinical action publishes a Kafka domain event auditable from app_17 (GDPR).

app_01port 8201
ED Triage
XGBoost · 5-class acuity · disposition · LOS
app_02port 8202
Sepsis ICU
LSTM · 4h horizon · SOFA · external validation Q1 2026 (eICU-CRD)
app_03port 8203
Hospital Ops
DES + MARL · 8-department staffing
app_04port 8204
Oncology AI
XGBoost · readmission + mortality · MIMIC-IV trained, NCRI validation pending
app_05port 8205
Patient Journey
Timeline · vitals · labs · meds · cohort
app_06port 8206
Clinical Chat
LLM (Ollama) · multi-module RAG
app_07port 8207
Data Ingestion / Twin
Digital twin · synthetic patient streams
app_08port 8208
Bed Management
Real-time bed board · predictive discharge
app_09port 8209
Waiting List
ML prioritisation · 750k patient context
app_10port 8210
Clinical Scribe
SOAP generation · ICD-10 coding
app_14port 8214
ED Flow Optimizer
PET breach prediction · queue simulation
app_15port 8215
Hospital ERP
Finance · rostering · procurement signals
app_16port 8216
Trolley Watch
INMO-style counter · breach alerts
app_17port 8217
GDPR Compliance
Audit trail · DSAR · DPIA artefacts
app_18port 8218
XAI
SHAP per prediction · decision log
app_19port 8219
FHIR Gateway
HL7 FHIR R4 · NSCR-ready · EHR-agnostic
app_20port 8220
Deterioration
NEWS2 · PEWS · IMEWS · escalations
app_21port 8221
Discharge Lounge
Predicted discharge · bed handoff orchestration
app_22port 8222
Alerts
Coalesced · WebSocket · ack and route hints
Architecture
FastAPI · MongoDB · Kafka (Redpanda) · React 19 · Tailwind 4 · Linkerd mesh ready · Helm charts
ML stack
XGBoost · LSTM (PyTorch) · DES + MADDPG MARL · Transformers · LLM (Ollama or Claude API ready)
Observability
Prometheus · Grafana · Loki · Jaeger · OpenTelemetry across every service
04 · Quantified impact

What becomes possible after pilot validation.

Modelled on a representative ~600-bed Model 4 acute hospital, ~75,000 ED attendances and ~20,000 inpatient discharges per year. Calibrated against published deployments at scale: Qventus reduces excess bed-days 20 to 35% across US health systems; GE Command Centre cut ED wait 35% at Johns Hopkins; Bradford NHS deployed it in the UK. The pilot exists to produce the Irish-context number that replaces these.

Modelled, not measured Every figure in this section is a projection. The pilot exists to verify or falsify it.
€20 to 30M
Annual benefit envelope
Direct savings + deferred capex + consultant-time recovery (mid-case)
~20,000
Bed-days released
~1 day shorter mean LOS × 20k discharges (Qventus benchmark applied to Irish census)
12 to 25
Sepsis lives saved (per ICU)
4-hour earlier detection at sensitivity over 90% × hour-of-delay 4 to 8% mortality
120 to 160
Avoided onco readmissions
15 to 20% reduction × ~800 baseline onco 30-day readmissions
National extrapolation
€250 to 400M
Bed-days released across HSE acute estate (notional throughput value).
Documentation-time
5 to 8%
Of total consultant capacity returned to patient-facing time. AI for Care Y1 target met.
Sepsis lives, nationally
250 to 400
Lives per year, modelled across HSE adult ICU census.
Three-year business case · single Model 4 hospital
Modelled · pilot must validate before HSE commits
Line item Year 1 Year 2 Year 3 Calculation basis
Platform licence + integration €1.8M €1.2M €1.2M Indicative only. Final commercials post-pilot.
HSE-side implementation effort €0.6M €0.3M €0.2M ~6 FTE × Y1 ramp; ~2 FTE BAU.
Total cost €2.4M €1.5M €1.4M
Bed-days released (mid-case) €8M €18M €22M ~1 day mean LOS reduction × 20k discharges × ~€1,000/bed-day (HSE BIU range €800 to €1,200).
Avoided oncology readmissions €0.4M €0.8M €1.0M ~150 avoided × ~€6,500 average admission cost.
Documentation-time recovery €0.8M €1.6M €2.0M ~50 min / consultant / day × ~150 consultants × ~€80/hr (pro-rata).
Net benefit €6.8M €18.9M €23.6M
Payback (mid-case) < 4 months Sensitive to LOS-reduction realisation rate.

Bed-day value is notional throughput, not cash. Cash equivalent depends on whether released capacity admits more elective patients (revenue) or closes winter-pressure escalation beds (cost avoidance). Both are valid HSE business-case framings. Sepsis mortality reduction is intentionally not monetised.

05 · Proposed pilot

One site. One front-door bundle. 12 weeks. HIHI-framed.

The lowest-risk path to a credible HSE-wide reference is a single Model 4 site running the front-door bundle (ED Triage + ED Flow + Trolley Watch + Bed Management) under the Health Innovation Hub Ireland framework, with HSE Demand and Capacity Visualisation as the downstream procurement target.

The falsification clause

At pilot kickoff, three success metrics are agreed with HSE clinical sponsors. Candidates: door-to-doctor time (front-door bundle), NEWS2-score-to-escalation latency (deterioration bundle), 30-day oncology readmission stratification (cancer bundle). At week 12, if none of the three move in the right direction by a clinically meaningful margin, the pilot terminates and HSE owes nothing. No licence, no integration cost, no ongoing obligation. The hypothesis is falsifiable. The downside is bounded.

Lead candidate
Mater Misericordiae

Already runs Aidoc. AI-friendly executive. Large urban ED.

Bundle: ED Triage + ED Flow + Trolley Watch
Phase 2
St James's Hospital

Largest hospital. NCCP centre. Runs eAltra pilot.

Bundle: Oncology + Discharge + Bed Mgmt
Phase 3
UH Galway

"Ruadhan" RPA bot already live. HSE West innovation appetite.

Bundle: Hospital Ops + Sepsis + Deterioration
Phase 4
Beaumont Hospital

Neuroscience centre. Persistent ED pressure.

Bundle: ED Triage + Deterioration + FHIR
12-month roadmap
Q1
Localisation + dossier

Manchester Triage System mapping. Irish ESI to MTS retrain. ICD-10-AM alignment. EU AI Act technical files.

Q2
Mater pilot live

Front-door bundle in production. FHIR Gateway resource mapping completed. Irish validation cohort begins.

Q3
St James's + HIHI submission

Oncology + Discharge + Bed Mgmt live. HIHI.AI 2026 national call submission with 2 reference sites.

Q4
UH Galway + national posture

ICU bundle live. Three-site reference deck. HSE Demand and Capacity Platform engagement opens.

Procurement and funding pathways available today
→ Health Innovation Hub Ireland (HIHI) clinical pilots, Q3 to Q4 2026 call
→ HSE Demand and Capacity Visualisation Platform, 2026 to 2027 procurement
→ HSE AI and Automation Centre of Excellence, continuous
→ Virtual Care Programme, €2.7M allocated 2026 (Phase 2 module)
→ Hospital Medicines Management, 12-site go-live 2026 (Phase 2)
→ Enterprise Ireland Innovation Vouchers / HPSU
06 · Why me

Built solo. By a recent MSc AI graduate in Ireland.

I'm Harishankar Somasundaram, an MSc Artificial Intelligence graduate based in Ireland. I built this platform solo, end to end, over the course of my MSc. Not as a commercial product, but as a working artefact for a thesis proposition: Ireland's acute hospital crisis isn't solvable by another point product, and the missing piece is a single integrated AI substrate covering the front door, the ward, the ICU, the cancer pathway and the back door.

To be precise about state. 19 FastAPI services and a unified React dashboard run end to end on a developer workstation today, with a Kafka event spine, durable Mongo event log, Helm charts, and Prometheus / Loki / Jaeger observability. None are clinically deployed yet. Models are MIMIC-IV-trained. External validation on eICU-CRD, Irish ICD-10-AM coding alignment, and Manchester Triage System retraining are explicit Phase 1 deliverables, not assumed.

Not asking HSE to buy anything. Asking for a 12-week pilot at one site under HIHI, with a falsification clause: if the pilot doesn't produce a measurable reduction in one of door-to-doctor time, NEWS2-to-escalation latency, or oncology-readmission stratification accuracy, the pilot ends and HSE owes nothing. Why HSE specifically: this was built in Ireland, against Irish published numbers, for an Irish problem. Sending it to NHS England or a US system would be intellectually dishonest about who it was built for.

Education
MSc Artificial Intelligence (Ireland)
Stack proven
FastAPI · PyTorch · XGBoost · MARL · React · Kafka · MongoDB · Helm · GKE
Available for
HIHI pilot · Research collab · HSE evaluation · Live demo
07 · The ask

A 30-minute demo. One pilot site. Zero up-front cost under HIHI.

Happy to walk HSE clinical and technical teams through the platform, then put a HIHI submission in front of the next national call. The technical work is done. The regulatory work is in flight. The remaining variable is HSE's choice of pilot site.

Document set
Business Requirements Document v1.1 · Architectural Roadmap · Irish Hospital AI Strategy Report · System Design · Technical Documentation
Source positioning
Independent submission. Open to collaboration with HSE, HIHI, HIQA, and academic partners. Not affiliated with any incumbent vendor.
Compliance posture
EU AI Act high-risk dossier in draft. GDPR DPIAs in flight. MDR Class IIa pathway identified. HIQA AI principles mapped.