1st place · Rakathon 2026

KOMPACS KOMparátor PACS Prior scans for comparison before CT, MRI, PET and tumor boards

KOMPACS finds them in the NIS, in report text and with the insurer, checks them in PACS and requests the missing ones.

available missing to decide
The problem

An old scan is missing. The case waits, a nurse hangs on the phone.

  • 1
    Before a CT or tumor board, the comparison scans are not in PACS.
  • 2
    Where they were done is only a sentence in a report.
  • 3
    The physician waits, a nurse calls, the patient waits.
6of 10

cases in the demo data lack a scan the physician needs.

Radiologists love two things in life: darkness and complete patient data. Sadly, today's healthcare mostly gives them only the darkness.

AKAdam Karas · medical student, KOMPACS
The solution

The check runs by itself, the moment an exam is ordered.

Exactly the workflow from the FN Motol and Na Homolce brief: 1 trigger → 2 what is relevant → 3 check PACS → 4 resolution

NIS record198
Mention in a report (AI)74
Health insurer14
Manual+
→
AI verdict

Relevant or not. With a reason and a guideline citation.

→
Coordinator

Decides the uncertain ones. Approves the request in one click.

→
PACS

Images arrive. The status turns “available”.

AI that cites its source

For every scan: a reason, the source sentence and a verified guideline citation. When the AI is unsure, a human decides.

AI verdict with guideline citation

Patients often forget. The insurer never forgets what it paid for.

AKAdam Karas · medical student, KOMPACS

VZP data: target 2028-01-01 (consulted with a VZP representative). Simulated in the prototype.

Searching prior exams…

KOMPACS case detail
Data from 40 demo cases

Where the comparison gets lost.

24 / 40 cases lack a scan in PACS that the physician needs
scan missing (24) all available (16)
286 → 127 → 34 from candidates to missing scans
Candidates
286
Relevant
127
Missing in PACS
34
NIS 198 reports (AI) 74 insurer 14

All 34 missing scans are named only in report text. A NIS or PACS search does not find them.

A non-obvious approach

The obvious solution would not find the missing scans.

Obvious approach
KOMPACS
Where to search
Query NIS and PACS. Finds 0 of 34 missing priors.
AI reads report text: “CT at Thomayer Hospital 12/2025”. Finds all 34.
What the patient forgot
Ask the patient. They often do not remember where or when.
Ask the insurer. It knows exams that no single hospital knows (14 in the demo).
How to request scans
An AI voice bot calls hospitals. It reads sensitive data to whoever picks up.
A precise e-mail or FHIR request. Verified contact. A human approves.
Why a scan is relevant
A black-box AI, or fixed rules. The physician cannot see why.
A reason, the source sentence and a verified guideline citation. A human decides the uncertain ones.
Deployment
A new system. The hospital changes its processes, integration is slow and costly.
Connects what the hospital has: HL7 v2, FHIR, a PACS feed. The hospital changes nothing.
Security and integration

No AI robots on the phone. A precise e-mail in one click.

Picture our AI reaching the hospital porter and reciting a birth number and a cancer diagnosis. Funny, yes. But your data protection officer would have a heart attack.

AKAdam Karas · medical student, KOMPACS
Toradiologie@thomayer.example
SubjectImage request: chest and abdomen CT 12/2025

Hello, for a CT planned on 23 Oct we kindly ask for the images and report of the chest and abdomen CT from 12/2025 via ePACS. Patient: PAC-DEMO-001. Thank you.

SendAI draft · approved by the coordinator

InterSystems IRIS for Health

HL7 v2 in, FHIR R4 inside, vector search in IRIS.

Data stays in the EU

Report text goes only to an EU Data Zone model. Every AI verdict has an audit trail.

A human decides

AI suggests. Nothing leaves without the coordinator. Prototype, not a medical device.

HL7 v2 ORM^O01FHIR R4IRIS Vector SearchFHIR SQL BuilderMCPEU Data Zone
Built this weekend

A working system, not slides.

  • A real HL7 order → IRIS → FHIR in ~1 s
  • AI judges the whole case and cites the guideline page
  • An unverified citation is dropped
  • 89,461 report sentences as vectors in IRIS
  • Case summary as a versioned FHIR document
  • MCP server for any AI agent

The prototype simulates orders, data from other hospitals and the insurer, and e-mail sending.

For physicians

The whole history on one timeline. A red flag when a critical scan is missing.

Interactive timeline

Scans, reports, treatment, labs. Semantic search.

2D anatomical map

Lights up only areas with proven imaging. Missing ones in red.

Impact

Fewer duplicates. Shorter waits. Less radiation for cancer patients.

34

missing scans found in 40 cases

60 %

of cases would start without a comparison

10 h → 2 h

work for 40 cases

estimate
~68

phone calls saved

estimate
Scalability

Not tied to one hospital.

The hospital changes nothing

It connects what it has: HL7 v2, FHIR, a PACS feed.

One stack per hospital

On-prem or EU cloud. Supported by a Czech IT integrator.

Works today

Image exchange under EHDS comes only in 2031.

What next

For us, this is not just a weekend hackathon.

Pilot at FN Motol

Real orders, clinical sign-off of the AI.

Na Homolce Hospital

Second site, connection to ePACS.

VZP data

Target 2028-01-01.

Team 404 Tumor Not Found

Get in touch.

info@kompacs.cz

VB

Vojtěch Bešťák

ML / data, frontend

PhD student · ČVUT FEL

LinkedIn
AK

Adam Karas

product owner

Medical student · 1. LF UK

LinkedIn
PK

Pavel Kos

product, business

LinkedIn
VN

Vojtěch Novotný

backend, integration

MSc student · ČVUT FIT

LinkedIn

KOMPACS. Compact data, so you can treat in peace.