LENS: localized updates facilitate patient-level unlearning in clinical large language models

Large language models fine-tuned on clinical records can reproduce patient information that a patient may ask to have removed. A deletion request applies only to the patient who made it, so outputs for other patients in the same fine-tuning data should remain unchanged. Unlearning studies typically check general medical knowledge but rarely examine these non-deleted patients. Restricting updates to a subset of model components has been reported to reduce side effects on untargeted knowledge, but its effect on non-deleted patients remains unclear. We propose layer-wise evidence-based neural suppression (LENS), which confines the unlearning update to a single four-layer window of the fine-tuning adapter. We fine-tuned MedGemma-4B-IT on MIMIC-III notes with synthetic patient names. After fine-tuning, ROUGE-L on deleted and non-deleted patients was 0.98–0.99, and a reference model trained without the deleted patients scored 0.33 on them. Existing methods that brought the deleted patients to this level (0.29–0.39) lowered the non-deleted patients to 0.47–0.69. LENS brought the deleted patients to 0.34 and kept the non-deleted patients at 0.90. External medical benchmark accuracy changed by less than 0.02 for every method and did not distinguish these outcomes. These results suggest that unlearning should also be evaluated on the non-deleted patients.

Authors

Institutions

Publication Details

Journal
Scientific Reports
Published
2026-11-21
DOI
https://doi.org/10.1038/s41598-026-80817-2
Primary Topic
Privacy-Preserving Technologies in Data
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

LENS: localized updates facilitate patient-level unlearning in clinical large language models

Tae Joon Jun, Minkyoung Kim, Jiye Han, Young-Hak Kim et al.
Scientific Reports
Privacy-Preserving Technologies in Data
article

LENS: localized updates facilitate patient-level unlearning in clinical large language models

Tae Joon Jun, Minkyoung Kim, Jiye Han, Young-Hak Kim, Sanghyun Park
article en

Abstract

Large language models fine-tuned on clinical records can reproduce patient information that a patient may ask to have removed. A deletion request applies only to the patient who made it, so outputs for other patients in the same fine-tuning data should remain unchanged. Unlearning studies typically check general medical knowledge but rarely examine these non-deleted patients. Restricting updates to a subset of model components has been reported to reduce side effects on untargeted knowledge, but its effect on non-deleted patients remains unclear. We propose layer-wise evidence-based neural suppression (LENS), which confines the unlearning update to a single four-layer window of the fine-tuning adapter. We fine-tuned MedGemma-4B-IT on MIMIC-III notes with synthetic patient names. After fine-tuning, ROUGE-L on deleted and non-deleted patients was 0.98–0.99, and a reference model trained without the deleted patients scored 0.33 on them. Existing methods that brought the deleted patients to this level (0.29–0.39) lowered the non-deleted patients to 0.47–0.69. LENS brought the deleted patients to 0.34 and kept the non-deleted patients at 0.90. External medical benchmark accuracy changed by less than 0.02 for every method and did not distinguish these outcomes. These results suggest that unlearning should also be evaluated on the non-deleted patients.

Scientific Reports
Yonsei University (KR), Asan Medical Center (KR), University of Ulsan (KR)
Privacy-Preserving Technologies in Data
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon