Adaptação dermatológica de modelos de linguagem pequenos via parameter-efficient fine-tuning em cenários de baixos recursos
Year: 2026
Type: dissertation
Abstract: Adapting Large Language Models (LLMs) to specialized clinical subdomains remains a significant challenge for small healthcare organizations due to the high computational costs of full-parameter fine-tuning. This work presents an empirical study using 4-bit Quantized Low-Rank Adaptation (QLoRA) to specialize a compact 1.5B parameter model (Qwen2.5-1.5B-Instruct) into a dermatological educational assistant under hardware constraints. Utilizing the Portuguese MedPT dataset, we evaluate three fine-tuned variants on a single GPU: a general clinical model (Qwen All), a directly specialized model (Derma 1), and a two-stage sequential model (Derma 2). Quantitative performance was measured across 1,625 test records using ROUGE and BERTScore metrics. To assess clinical safety and regulatory compliance under the Federal Council of Medicine (CFM) guidelines, a blind qualitative review of 100 cases was conducted by a non-specialist human evaluator and two LLM judges using a five-dimensional Likert scale and a selection preference to identify the best overall response. While automated metrics improved after fine-tuning, they proved insufficient to assess true clinical safety or regulatory adherence. Qualitative results revealed that fine-tuned models consistently outperformed the baseline, which was completely excluded from the best-answer selection in 73.0% of cases due to frequent hallucinations despite its high answer completeness. The findings demonstrate that parameter-efficient fine-tuning provides a viable path for small language models to acquire specialized clinical orientation capabilities within resourceconstrained environments. However, due to persistent vulnerabilities such as occasional factual errors in complex tasks, these models cannot autonomously replace healthcare professionals and must serve strictly as auxiliary educational tools under active human supervision.
Author: Mews, Matheus Oliveira
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7203504784