SimpliMed-RAHF: A Human-Feedback-Inspired Reinforcement Learning Framework for Medical Text Simplification With the SimpleDC Dataset

Abstract

Barriers to communication and understanding across health domains remain a critical issue, particularly when complex medical terminology is involved. Advances in large language models offer opportunities to improve accessibility through automatic text simplification. However, challenges persist regarding readability, accuracy, and semantic fidelity—limitations that traditional ATS methods have struggled to overcome. This paper presents SimpliMed, a multi-phase medical text-simplification framework built using the Simple Digestive Cancer corpus. The study evaluates baseline model performance using sequencing and causal architectures, applies supervised fine-tuning to BART-base and T5-small, and experiments with various prompting strategies. Building on these insights, we design a hybrid construction pipeline inspired by Reinforcement Learning from AI Feedback (RAHF). The hybrid system employs an 8-bit quantized BART-Large-CNN model, enhanced by dictionary-based preprocessing and candidate re-ranking guided by a heuristic reward function that balances semantic fidelity (BERTScore), readability (FKGL), and structural overlap (ROUGE-L). Automated evaluations and re-ranking experiments demonstrate the effectiveness of this hybrid approach, achieving strong performance (SARI: 61.63, BERT-F1: 0.95, ROUGE-L: 0.67) while maintaining high readability (FKGL: 6.47).Additional confirmation of this is provided by 31 participants, with 45.2% of subjects clearly preferring the hybrid output over the baseline and 41.9% finding the hybrid output to be equally suited to the human reference. In addition, four consulting physicians confirmed the clinical suitability of the simplified texts. Overall, SimpliMed—and particularly its efficient hybrid re-ranking methodology—shows meaningful progress toward producing accurate, readable, and clinically reliable medical text simplifications.