WORKFLOW CLINICAL CARE Reduce administrative burden Support healthcare professionals • Automate repetitive tasks • Draft discharge summaries • Summarize admissions for handover • Provide medical decision support • Implement a reasoning engine OPERATIONS RESEARCH Optimize logistics and capacity Change how research is organized • Predict no-shows • Forecast patient flow • Plan outpatient clinics • EHDS • Federated learning 3
governance must work together from the start 01 DATA INFRASTRUCTURE SECURITY IS A DESIGN REQUIREMENT Connected, near-real-time data Bring multiple sources together securely. NIS2 • NEN 7510 • NEN 7513 02 AI INFRASTRUCTURE Production-ready AI infrastructure Develop, train and deploy AI solutions. AI Act • MDR 03 LANDING ZONES A shared platform for implementation Deliver solutions securely, robustly and at scale. 4
and scale the UMC Utrecht platforms Platform as a product Self-service by default Iterate with users and prioritize their needs. Enable independent use without waiting for our team. Built-in security & privacy Build in controls so compliance is easy and enforced. Paved roads Glue is value Standardize capabilities to improve speed and consistency. Connect UMCU services to improve DX and scale across the region. 5
services create control without slowing teams down GOVERNANCE INGESTION ACCESS ENABLEMENT One auditable control plane Metadata-driven ETL at scale Ownership without bottlenecks Platform as a service Govern data use consistently across the organization. Reusable pipelines ingest data from many sources. Data stewards and owners manage access decisions Teams build and run on one shared central foundation. 7
with secure support across the MLOps lifecycle CURRENT STATE WITH THE AI PLATFORM Fragmented infrastructure slows progress One secure path from experiment to practice • Teams rely on separate tools and environments • Laptops, local servers and research-only solutions create shadow IT • AI Act and MDR compliance is difficult • Clinical implementation is slow A scalable, compliant foundation supports AI teams and researchers at every stage. 8
security, monitoring and data services stay consistent ISOLATION ARCHITECTURE Team environments Isolated by team and by plane. TRAINING PLANE Build and register Experiment, train and register AI models. SERVING PLANE Deploy and monitor Deploy, infer and monitor production models. FOUNDATION · Databricks · Azure Container Apps · shared Azure services 9
outpatient clinics PUBLIC CONVERSATION 6% of appointments are missed despite reminders PROJECT ORIGIN UMC Utrecht began developing its own model in early 2023, inspired by Erasmus MC’s earlier success. OPEN BY DEFAULT The code has been public on GitHub from the start. “Tens of thousands miss appointments; AI helps hospitals predict who.” – AD (2024-08) 10
that calling patients ordered by an AI model can significantly reduce no-shows 01 CALLING WORKFLOW RANK Prioritize appointments Estimate no-show risk from age, travel distance, prior no-shows and other appointment features. 02 CALL Contact selected patients The UMC Utrecht calling team calls patients daily to remind them of their appointment 03 RECORD Log the outcome Staff record call status and result in the workflow. PRIVACY GUARDRAILS No socioeconomic or medical data · No profiling 11
a broader opportunity across teams, hospitals and the region 01 02 03 COLLABORATE SCALE THE AI PLATFORM SCALE THE DATA PLATFORM Collaboration internally and across hospitals Bring more products onto the platform Expand from UMC Utrecht to the region • Collaboration across platform teams • Share knowledge and platforms across hospitals • Onboard more AI products • Deploy capabilities for research teams • Onboarding all data consumers • Scale internally • Extend into the region 12
mark3t1ng
tmiket
mayatellez
sugarenia
lindahogenes
tammyeverts
inesmontani
lauravandoore
marcduiker
geshan
colly
caitiem20
mkilby
