PLACID: Privacy-preserving Large language models for Acronym Clinical Inference and Disambiguation

Manjushree B. Aithal; Ph. D.; Alexander Kotz; James Mitchell; Ph. D

cs.CLMar 24, 2026

PLACID: Privacy-preserving Large language models for Acronym Clinical Inference and Disambiguation

Authors: Manjushree B. Aithal, Ph. D., Alexander Kotz, James Mitchell, Ph. D

Organizations: Department of Biomedical Informatics, University of Colorado Anschutz, Aurora, CO

Abstract

Large Language Models (LLMs) offer transformative solutions across many domains, but healthcare integration is hindered by strict data privacy constraints. Clinical narratives are dense with ambiguous acronyms, misinterpretation these abbreviations can precipitate severe outcomes like life-threatening medication errors. While cloud-dependent LLMs excel at Acronym Disambiguation, transmitting Protected Health Information to external servers violates privacy frameworks. To bridge this gap, this study pioneers the evaluation of small-parameter models deployed entirely on-device to ensure privacy preservation. We introduce a privacy-preserving cascaded pipeline leveraging general-purpose local models to detect clinical acronyms, routing them to domain-specific biomedical models for context-relevant expansions. Results reveal that while general instruction-following models achieve high detection accuracy (~0.988), their expansion capabilities plummet (~0.655). Our cascaded approach utilizes domain-specific medical models to increase expansion accuracy to (~0.81). This novel work demonstrates that privacy-preserving, on-device (2B-10B) models deliver high-fidelity clinical acronym disambiguation support.

Figures & tables

Figure 1: PLACID multi-stage process pipelines for clinical acronym disambiguation using local small models. As illustrated, inference can be deployed in either a single-pass (simultaneous acronym detection and expansion via a single model) or cascaded (general model for detection followed by a medically fine-tuned model for expansion) configuration, depending on local computational constraints.

Dataset	Mode	Total Instances	Average Tokens	Unique Acronym	Unique Expansion	Overshadowed Instances	Overshadowed Ratio (%)
Biomedical	Single-pass	2,026	25.26	671	748	126	6.22
Biomedical	Cascaded	6,299	30.60	1,103	1,323	613	9.73
General	Single-pass	1,000	56.86	56	57	3	0.3

Table 1: Distribution of acronyms, expansions uniqueness, average tokens, and overshadow ratio (detailed definitions provided in Section 0.0.3 ) for parsed Biomedical and General GLADIS 13 data in our study.

			Expansion Accuracy
Model	Text	Det. Acc.	BLEU	METEOR	ROUGE-L
Foundation	Raw	0.785 ± 0.369	0.654	0.655	0.680
Foundation	Clean	0.785 ± 0.369	0.654	0.655	0.680
Gemma2:2b	Raw	0.811 ± 0.392	0.699	0.704	0.729
Gemma2:2b	Clean	0.811 ± 0.392	0.699	0.704	0.727
Llama3.2:3b	Raw	0.687 ± 0.464	0.612	0.620	0.643

Table 2: Zero-shot local small model’s performance accuracy for single-pass mode (detection and expansion) for general-domain 1k input data over 5 iterations. Note, blue text highlights the best scores.

			Expansion Accuracy
Model	Text	Det. Acc.	BLEU	METEOR	ROGUE-L
Foundation	Raw	0.988 ± 0.072	0.528	0.541	0.596
Foundation	Clean	0.988 ± 0.072	0.541	0.564	0.598
Gemma2:2b	Raw	0.930 ± 0.255	0.535	0.546	0.608
Gemma2:2b	Clean	0.930 ± 0.255	0.550	0.573	0.608
Llama3.2:3b	Raw	0.945 ± 0.229	0.518	0.535	0.588

Table 3: Zero-shot local small model’s performance accuracy for single-pass mode (detection and expansion) for biomedical-domain

\sim

2k input data over 5 iterations. Note, blue text highlights the best scores.

Figure 2: Average expansion accuracy comparison of all models in cascaded-mode over 5 iterations.

Clinical NLP

Clinical natural language processing (NLP) develops methods to extract, interpret, and generate information from healthcare text and conversations.

See all 142 papers

Selective Token-Level Cryptographic Redaction for Privacy-Preserving Clinical Deployment of Large Language Models
Jun 2, 2026Farhan Sheth, Ziyuan Yang, Yongying Lan +1Privacy-Preserving Language Models Medical LLMs
MedVisAI Lab, Singapore · Lee Kong Chian School of Medicine, Nanyang Technological University, Singapore · Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, China
Clinically Grounded Privacy Evaluation of Medical LMs
Jun 8, 2026Sasha Ronaghi, Sana Tonekaboni, Lena Stempfle +6Data Leakage Memorization in Language Models
1Stanford University · 2Massachusetts Institute of Technology · 3American Board of Family Medicine
Future Querying: Can LLMs Serve as Implicit Medical World Models?
Aug 24, 2026Siri Willems, James Butterworth, Lore Goetschalckx +4Medical LLMs Clinical Prediction
imec, AI-labs⋆⋆, Paris, France

Jun 2, 2026J/K move · Enter open · S save

Selective Token-Level Cryptographic Redaction for Privacy-Preserving Clinical Deployment of Large Language Models

Farhan Sheth, Ziyuan Yang, Yongying Lan, Si Yong Yeo

MedVisAI Lab, Singapore · Lee Kong Chian School of Medicine, Nanyang Technological University, Singapore · Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, China

Privacy-Preserving Language Models Medical LLMs

Open paper AskPDF arXiv

While large language models (LLMs) are increasingly used for clinical applications, many existing pipelines require sending raw sensitive health information to remote servers for processing, which heightens the risk of privacy leakage. A natural approach to mitigate this risk is to encrypt the data before transmission. However, straightforward solutions such as encrypting the entire dataset introduce prohibitive computational, alignment, and communication overheads, rendering large-scale practical deployment infeasible. To preserve privacy while maintaining usability, we present Healthcare Encryption & Redaction via Adaptive Linguistic Decomposition (HERALD), a token-level cryptographic redaction framework designed to achieve this balance by encrypting only sensitive tokens while preserving the surrounding context for downstream model utility. HERALD combines medical named-entity recognizer (NER) with part-of-speech (POS) driven policies to select candidate tokens, performs targeted lemmatization to stabilize surface forms, and substitutes each protected token with a deterministic ciphertext wrapped in explicit delimiters. Notably, HERALD is model-agnostic and operates entirely on the client side, ensuring that sensitive content remains encrypted throughout storage, transmission, and processing without requiring changes to downstream models. We evaluated HERALD on both classification and medical question answering (MQA) tasks on public datasets. Across different tasks, experiments illustrate that fully secured baselines suffer significant utility loss, whereas HERALD consistently recovers performance close to plaintext. Overall, HERALD provides a novel utilization pipeline.

Abstract

Clinical NLP

Explore similar work

Selective Token-Level Cryptographic Redaction for Privacy-Preserving Clinical Deployment of Large Language Models

Clinically Grounded Privacy Evaluation of Medical LMs

Future Querying: Can LLMs Serve as Implicit Medical World Models?