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

arXiv:2603.23678 · cs.CL, cs.AI · Submitted 2026-03-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "PLACID: Privacy-preserving Large language models for Acronym Clinical Inference and Disambiguation".

Jane: The gist The study pioneers an evaluation of small-parameter models deployed entirely on-device to ensure privacy preservation for clinical acronym disambiguation.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper called PLACID: Privacy-preserving Large language models for Acronym Clinical Inference and Disambiguation. It sounds pretty specific. Jane, what’s the main idea behind that title?

Jane: Well, it basically tackles the problem of using large language models in healthcare without exposing private patient data to external servers. The core idea is moving away from cloud-dependent solutions to something local and secure >

Lu: Exactly. It focuses on making sure clinical information stays private by running these models entirely on a device, which means no Protected Health Information goes anywhere outside that device >

Meng: That's important because if you're dealing with patient records, HIPAA rules mean you can't just send that data over the internet to a big model provider. So, local processing is a necessity for deployment >

Lalam: I agree. It’s about keeping the sensitive data private while still getting clinical help with those confusing acronyms in medical notes >

The paper's summary: Tom: Okay, so what does the PLACID study actually propose? It seems like they are trying to bridge that gap we talked about. Jane, can you explain the pipeline they introduce in simple terms?

Jane: They introduce this privacy-preserving cascaded pipeline. Think of it as a chain where you start with a general-purpose local model to find the acronyms first, and then you pass those detected acronyms to a domain-specific biomedical model for the actual expansion >

Lu: It’s a two-step process. First, the general model detects what an acronym is in the text. Then, it takes that identified acronym and feeds it into a specialized medical model to figure out what it actually means in that clinical context >

Meng: And they are using small parameter models, specifically models between two billion and ten billion parameters, running entirely on-device. That’s key for deployment because those things don't require constant internet access or massive server power >

Lalam: So the main summary is that they use a local-first deployment with this cascaded pipeline to handle acronym disambiguation without sending patient data to external APIs >

The paper's improvements: Tom: Now, the PLACID authors point out some specific improvements. They compared their method against general instruction-following models, and they found a significant gap there. What did they find about the general models?

Jane: They found that while those general instruction-following models can get pretty good at detecting acronyms—around zero point nine eight eight accuracy—their ability to actually expand those acronyms in a medical sense was much weaker, around zero point six five five >

Lu: The paper states that the expansion quality is governed more by the domain-specific language models than by how well they can detect the acronym initially >

Meng: That means just finding the word isn't enough; you need deep medical knowledge to actually give you a good answer, which is what this cascaded approach aims to fix >

Lalam: They found that when they used medically fine-tuned models like MedGemma and BioMistral in their cascaded mode, the expansion accuracy jumped up to zero point eight one >

Conclusion: Tom: So, to wrap up this PLACID paper, the main implication is that routing detected acronyms from a general model to a specialized medical model significantly boosts accuracy for clinical use cases >

Jane: It proves that you don't have to rely on massive cloud models for every task if you design a smart local pipeline like this one >

Lu: The authors suggest the next step is moving toward a fully fine-tuned, single-pass model instead of this cascaded setup to maximize both computational efficiency and clinical fidelity >

Meng: From my point of view, I think that transition to a single-pass model is what we need for actual deployment because it's more computationally efficient and easier to run reliably at scale >

Lalam: And they also plan to integrate Retrieval-Augmented Generation or RAG into the pipeline later on to bridge that domain knowledge gap even further >

Manjushree B. Aithal, Alexander Kotz, James Mitchell

Department of Biomedical Informatics, University of Colorado Anschutz

cs.CL, cs.AI

Submitted: 2026-03-24

Updated: 2026-10-13

Comments: 10 pages, 2 figures, Under review AMIA Symposium

Code: https://github.com/ml-explore/mlx

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 76/100

The gist: The gist The study pioneers an evaluation of small-parameter models deployed entirely on-device to ensure privacy preservation for clinical acronym disambiguation.

Key concepts

On-device LLMs
These are small language models (ranging from 2B to 10B parameters) designed to run directly on a user's local hardware, like a smartphone or desktop. The key benefit is that all data processing happens locally, ensuring sensitive patient information never leaves the device, which is crucial for healthcare privacy.
Privacy-Preserving Cascaded Pipeline
This is a two-stage system where one general model first scans text to find acronyms (detection), and then those identified acronyms are passed to a more specialized medical model (expansion). This method keeps the sensitive information local while using the strengths of both types of models sequentially for better results.
Acronym Disambiguation
This is the process of correctly figuring out what an ambiguous abbreviation in a clinical note actually means. For example, determining if 'CHF' stands for Congestive Heart Failure or something else. Misinterpreting these can lead to dangerous medication errors, so accurate disambiguation is vital for patient safety.
Domain-Specific Models
These are smaller language models that have been specifically trained on medical data, such as PubMed abstracts. They are superior for expanding acronyms because their specialized training allows them to generate contextually relevant and clinically accurate explanations, unlike general models.

Terminology

Summary

The gist The study pioneers an evaluation of small-parameter models deployed entirely on-device to ensure privacy preservation for clinical acronym disambiguation.

Problem and Motivation

Large Language Models (LLMs) offer transformative solutions across many domains, but healthcare integration is hindered by strict data privacy constraints, as transmitting Protected Health Information to external servers violates privacy frameworks like HIPAA and GDPR<ref:2603.23678#pg2> Clinical narratives are dense with ambiguous acronyms, misinterpretation these abbreviations can precipitate severe outcomes like life-threatening medication errors<ref:2603.23678#pg2> While cloud-dependent LLMs excel at Acronym Disambiguation, transmitting Protected Health Information to external servers violates privacy frameworks<ref:2603.23678#pg2>

Proposed Solution and Methodology

The study introduces a privacy-preserving cascaded pipeline leveraging generalpurpose local models to detect clinical acronyms, routing them to domain-specific biomedical models for contextrelevant expansions<ref:2603.23678#pg2> This novel work demonstrates that privacy-preserving, on-device (2B-10B) models deliver high-fidelity clinical acronym disambiguation support<ref:2603.23678#pg2> The pipeline supports two distinct operational models: the single-pass model allows one model to independently execute both acronym detection and expansion, offering a high computationally efficient end to end approach requiring only one inference call per input<ref:2603.23678#pg2> The cascaded mode workflow proceeds in four sequential steps, beginning with the processing of a clinical text sample by the prompt template mapping to generate a zero-shot acronym detection prompt<ref:2603.23678#pg2> This prompt is sent to the general-domain small model which processes the input instance and determines the acronyms present in the given text<ref:2603.23678#pg2> Following this step, the detected acronym(s) are sent to the prompt mapping along with the original input text to construct an expansion prompt, reformulating the task from open-ended identification to focused disambiguation and expansion generation<ref:2603.23678#pg2> Finally, this prompt is routed to a domain-specific model which leverages its biomedical pretraining to generate a clinically grounded expansion for the acronym and a self-assessed confidence score in the range of [0, 1] for the expansions<ref:2603.23678#pg2>

Experimental Setup and Data

All experiments are performed on-device (Apple M4-Max GPU using MLX setup) ensuring zero transmission of data to external or cloudbased APIs<ref:2603.23678#pg2> The foundation of the experimental setup relies on the biomedical subset as well as some portion of general domain of the GLADIS benchmark<ref:2603.23678#pg2> The biomedical subset of the GLADIS comprises of ∼12k samples (6,295 training, 3,150 validation, and 3,149 test) that were obtained by adapting the MedMentions47 corpus of PubMed abstracts<ref:2603.23678#pg2> The dataset preparation involved parsing the biomedical validation and test data of the GLADIS dataset to yield a total of 6,299 unique text inputs<ref:2603.23678#pg2> A subset of single acronym per input text was separated for the acronym detection accuracy test, resulting in a final subset of 2,026 single acronym per input text<ref:2603.23678#pg2> Model selection was restricted to small language models within the 2 to 10-billion parameter range to align with privacy-preserving constraints<ref:2603.23678#pg2>

Model Comparison and Results

The performance of off-the-shelf general instruction-following small models showed a dissociation between detection and expansion accuracy in the biomedical setting, as expansion quality in the biomedical setting is governed more by domain specific language models than by the detection reliability<ref:2603.23678#pg2> Specifically, Mistral7B delivered superior expansions despite its lower detection rate<ref:2603.23678#pg2> Conversely, when testing the cascaded mode with medically fine-tuned models like MedGemma and BioMistral, the expansion accuracy was significantly higher, achieving an expansion accuracy of 0.81<ref:2603.23678#pg2> The study observed that domain-specific models substantially outperforming the general-purpose models<ref:2603.23678#pg2> Furthermore, the analysis of MedGemma inference logs illustrated that MedGemma avoids mapping acronyms to its most statistically frequent expansions but instead computed contextual relevance that was not observed in general-models<ref:2603.23678#pg2>

Conclusion and Future Work

The findings validate the proposed hybrid pipelines of pairing general models for rapid acronym detection with specialized models for disambiguation and expansion generation, which is an effective and secure solution for clinical decision support<ref:2603.23678#pg2> This work establishes a baseline for privacy-preserving acronym disambiguation with future optimization, demonstrating that transitioning from cascaded pipeline to a fully fine-tuned, single-pass model will be the critical next step for maximizing both computational efficient and clinical fidelity in on-device deployments<ref:2603.23678#pg2> Future work plans include fine-tuning smaller models on a hybrid dataset and reintroducing an ensemble output methodology for expansion generation<ref:2603.23678#pg2> Finally, the team intends to integrate Retrieval-Augmented Generation (RAG) into the pipeline to further bridge the domain-knowledge gap<ref:2603.23678#pg2>

Improvements for AI systems

  1. This study demonstrates that the cascaded approach utilizes domain-specific medical models to increase expansion accuracy to (∼0.81) compared to general instruction-following models which plummet [in] their expansion capabilities (∼0.655). This allows the improved AI system to achieve high-fidelity clinical acronym disambiguation by routing detected acronyms from a general model to a medically fine-tuned model for context-relevant expansions.

  2. The improved system can be deployed as a privacy-preserving cascaded pipeline leveraging generalpurpose local models to detect clinical acronyms, routing them to domain-specific biomedical models for context relevant expansions. This enables the AI to maintain zero transmission of data to external or cloudbased APIs while ensuring high accuracy in expansion generation.

  3. The system can leverage a structured evaluation process where fine-tuned models employed a structured evaluation process to weigh competing semantic expansions against the surrounding context, as illustrated in MedGemma inference logs, which allows for auditable rationale supporting expert clinical reasoning.

  4. The system can utilize a dual-mode pipeline offering options between a single-pass model and a cascaded specialization mode, allowing deployment flexibility based on computational constraints, such as choosing the single-pass mode for strict latency constraints.

  5. The improved system can be optimized by transitioning to a fully fine-tuned, single-pass model to maximize both computational efficiency and clinical fidelity in on-device deployments, addressing the observation that general models fundamentally lack the deep, specialized domain knowledge required to expand accurate, contextually relevant clinical narratives.

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.

Sources

Related papers