SMALL LANGUAGE MODELS FOR EARLY DETECTION OF ADRD - OPEN-SOURCE LANGUAGE MODELS IN CLINICAL SPEECH CLASSIFICATION

Q: Why consider small language models for ADRD detection instead of large language models?

LLMs are limited by cost, privacy risks, and HIPAA concerns. SLMs can run entirely on local hardware, reducing latency and dependence on external services while supporting privacy-preserving use.

Q: How was the study evaluated and what data was used?

The study tested four SLMs on 398 transcripts from the Pitt Corpus, using the Cookie Theft picture description task for dementia patients and control participants. Models were trained and tested on a subset of the corpus.

Q: Which model performed best and what were the key results?

GPT-2 achieved the highest reported performance at 78.3% accuracy, with inference times under three seconds. The balance of accuracy and efficiency supports locally hosted SLMs for privacy-preserving dementia detection.

Sep 28, 2026

Uploaded by mieayamfan

Large language models (LLMs) in biomedical and clinical research face constraints from cost, privacy risks, and HIPAA requirements. Small Language Models (SLMs) with fewer than 2 billion parameters can run locally on on-premise hardware, lowering latency and removing reliance on external services. This study evaluates four SLMs on 398 clinical speech transcripts using a Cookie Theft picture description task to classify ADRD versus controls, reporting accuracy comparable to LLMs under HIPAA-compliant low-latency conditions.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Methods

Model evaluation and dataset

Results

Classification accuracy and inference time

Implications and future work

Privacy-preserving deployment and multimodal real-time analysis

Innovation in Aging, 2025, Vol. 9, No. S2 1295 Abstract citation ID: igaf122.4317 SMALL LANGUAGE MODELS FOR EARLY DETECTION OF ADRD: OPEN-SOURCE LANGUAGE MODELS IN CLINICAL SPEECH CLASSIFICATION Venkatanand ram Addepalli1, Praveen Rao2, Erich Kummerfeld3, Andrew Kiselica4, and Knoo Lee5, 1. Institute of Data Science and informatics, University of Missouri, Columbia, Missouri, United States, 2. University of Missouri, Columbia, Missouri, United States, 3. University of Minnesota-Twin Cities, Twin Cities, Minnesota, United States, 4. University of Georgia, Athens, Georgia, United States, 5. University of MissouriColumbia, Columbia, Missouri, United States Large language models (LLMs) have shown promise in biomedical and clinical applications, including aging research, patient care, and translational geriatrics. However, their use is limited by costs, privacy risks, and HIPAA concerns. These challenges point to an alternative: Small Language Models (SLM; < 2 billion parameters) that run entirely on local hardware, reducing latency and dependence on external services. Recent open-source releases (e.g., DeepSeek R1) highlight the growing capability and efficiency of open-source models. These advances draw attention to SLMs, which are inherently efficient and easier to deploy in local clinical settings. This underscores the need to evaluate smaller, open-source language models on similar tasks. In this study, we evaluated four SLMs (DeepSeek Coder, GPT-2, Microsoft Phi-3, SmolLM2) on 398 transcripts from the Pitt Corpus, comprising dementia patients and control participants evaluated with the Cookie Theft picture description task. To our knowledge, this is the first on-premise deployment of an SLM directly on linguistic features of transcripts on those with ADRD, demonstrating comparable dementia classification accuracy to LLMs under HIPAA-compliant, low-latency conditions. Models were trained and tested on a subset of the Pitt corpus. The best-performing model (GPT-2) reached 78.3% accuracy with inference times under three seconds. This balance of accuracy with efficiency demonstrates the feasibility of locally-hosted SLMs as privacy-preserving tools for dementia detection. Our findings suggest that SLMs may offer practical pathways for deploying AI in aging research and clinical care. Future work will assess real-time speech analysis with integration of multimodal data for comprehensive geriatric assessment.

FAQs

Why consider small language models for ADRD detection instead of large language models?

LLMs are limited by cost, privacy risks, and HIPAA concerns. SLMs can run entirely on local hardware, reducing latency and dependence on external services while supporting privacy-preserving use.

How was the study evaluated and what data was used?

The study tested four SLMs on 398 transcripts from the Pitt Corpus, using the Cookie Theft picture description task for dementia patients and control participants. Models were trained and tested on a subset of the corpus.

Which model performed best and what were the key results?

GPT-2 achieved the highest reported performance at 78.3% accuracy, with inference times under three seconds. The balance of accuracy and efficiency supports locally hosted SLMs for privacy-preserving dementia detection.

Sep 28, 2026

Uploaded by mieayamfan