Healthcare, biotechnology & medical AI Research
Research tests selective forgetting for clinical language models
LENS restricts an unlearning update to part of a model adapter, studying removal of selected patient records while preserving responses about other patients. [1]
- Stage of evidence
- Peer-reviewed accepted article; further editing expected. [1]
What the researchers found
The authors evaluated a localized update on MedGemma fine-tuned with MIMIC-III notes and synthetic names. [1]
Why it matters
General medical benchmarks alone can miss changes affecting patients whose records were retained. [1]
Who can use the findings
Clinical-model developers and privacy evaluators. [1]
Key dates
- LENS patient-unlearning research published [1]
Practical uses
Evaluate retained-patient behavior alongside deletion tests. [1]
Limits of the evidence
An accepted research article and controlled evaluation do not establish complete erasure, regulatory compliance or deployed clinical safety. [1]
What to watch next
Independent replication and broader model testing. [1]
Market relevance
Sector context
-
Clinical-model privacy research Editorial assessment
Selective unlearning is relevant to teams examining how patient-related training information can be removed without damaging unrelated model behavior. [1]
Confirmed developments
-
The paper studies adapter-localized unlearning in a clinical language model using narratives with synthetic patient identifiers, comparing deletion and retained-patient behavior. [1]
What to watch next
-
Editorial watchpoint
Check independent removal tests and effects on retained records; a controlled model experiment does not establish legal erasure compliance or safe clinical deployment. [1]
Sources and verification
Catalog evidence checked How we verify records
Related profiles
Healthcare, biotechnology & medical AI
nference–Cedars-Sinai clinical research platform partnership
nference and Cedars-Sinai
The planned system combines different records and research inputs using nference’s federated software, longitudinal curation and physician adjudication. Intended uses include investigator-led studies and developing and evaluating AI models.
Healthcare, biotechnology & medical AI
RADINA imaging workflow platform
RADIN Health
The company describes referral, scheduling, authorization and follow-up workflows, with exceptions escalated to staff. A Concierge component is presented as a mobile robot, kiosk or tablet for patient access and registration tasks.
Healthcare, biotechnology & medical AI
Microsoft Dragon Copilot
Microsoft
Organizations can tailor flowsheet profiles to supported nursing settings. A simulation preview uses synthetic patients for practice, while nursing analytics expand and export to Microsoft Fabric becomes generally available.