SK Telecom Co. and clinical-AI vendor Holmes said they started a federated-learning pilot across four Seoul university hospitals, updating brain MRI triage models on site while exchanging only encrypted gradient updates rather than pooling identifiable imaging in a central SK cloud. The trial targets emergency departments that see overnight stroke rule-outs, where radiologist staffing thins and minutes to flag large-vessel occlusions matter for thrombectomy referrals.
How federated training runs
Each hospital keeps DICOM studies inside its approved picture-archiving perimeter. Holmes containers on hospital GPUs train local epochs on de-identified segments radiologists labeled for the study, then send gradient summaries through SK Telecom’s A.X edge security layer to an aggregator SK hosts in a Seoul availability zone without raw pixel data. SK said bandwidth use stays below nightly backup windows by compressing updates and scheduling rounds during low MRI utilization hours.
Holmes previously deployed detection assistants at individual hospitals; federated mode aims to improve sensitivity on Korean skull anatomy and coil protocols without violating Personal Information Protection Commission cross-institution transfer rules. Ethics boards at participating campuses signed harmonized protocols after months of legal review.
Clinical governance
Radiology chiefs retain veto power on model deployments; federated rounds that degrade holdout metrics trigger automatic rollback. Nurses and residents will not see AI scores in production workflows until a six-month validation ends—current access is research-only overlays reviewed in multidisciplinary conferences.
The Ministry of Health and Welfare’s digital health team attended kickoff briefings, asking whether federated stacks can scale to provincial hospitals with weaker GPU estates. SK proposed a tiered plan where small sites contribute labels but not training compute until edge appliances ship next year.
Why SK Telecom is in the loop
SK markets A.X as a sovereign AI cloud for regulated industries, competing with Naver Cloud and NHN after telecom revenue growth slowed. Healthcare federated learning showcases network security skills—VPN hardening, anomaly detection on gradient traffic—beyond selling handsets. Holmes gains telecom-grade operations for hospitals that already buy SK mobile contracts for staff devices.
Competitors including Lunit and Vuno train centrally on curated datasets with explicit patient consent; federated paths cost more engineering hours but answer hospital lawyers nervous about megabyte-scale leaks during transit.
Privacy and security risks
Researchers warned gradient inversion attacks remain theoretically possible if models are too large relative to cohort size; SK and Holmes capped participant counts per round and add differential privacy noise calibrated with Korea Advanced Institute of Science and Technology advisors. PIPC staff asked for incident playbooks if an aggregator node is compromised; SK promised quarterly penetration tests shared under NDA with hospital CISOs.
Patients receive opt-out paths on consent forms explaining AI research uses without promising individual clinical benefit during the pilot.
Metrics and timeline
Primary endpoints include change in false-negative rates on large-vessel occlusion flags against a retrospective gold standard, and wall-clock time to alert on-call attendings during simulated night shifts. Secondary metrics cover GPU power draw and nurse satisfaction surveys—practical barriers if federated jobs slow PACS performance.
If metrics clear, SK and Holmes will pitch provincial expansion to Busan and Daegu stroke centers, though funding may require National Research Foundation grants tied to aging-society initiatives.
What doctors watch
Radiology residents asked whether federated models trained mostly at tertiary centers will bias against community hospitals with older scanners; Holmes plans domain-adaptation layers per site. For now, Seoul’s four-campus trial is a compliance-conscious bet that Korean hospitals can share learning without sharing pixels—a test federated evangelists call overdue in a country with universal imaging volumes but strict privacy politics.
Funding and equipment
SK Telecom is co-funding GPU upgrades at two sites where older PACS servers could not host Holmes containers without slowing clinical reads. Hospital foundations matched grants tied to regional innovation clusters, spreading capital costs beyond SK’s marketing budget.
Insurance reviewers asked whether AI-assisted triage affects malpractice premiums; underwriters told InfoHandle they await peer-reviewed outcomes before adjusting rates, keeping clinicians focused on validation rather than billing codes during the pilot.