Recognition of Sensitive Personal Data in Doctor-Patient Speech

This study addresses the automatic recognition and de-identification of Sensitive Health Information (SHI) from multilingual and spoken clinical data, a critical task for privacy-preserving healthcare analytics. We propose a comprehensive framework integrating speech recognition and Named Entity Recognition (NER), leveraging cool-whisper and parakeet-tdt-0.6b-v2 for transcription and Llama-3.2-3B-Instruct with Low-Rank Adaptation of Large Language Models (LoRA) for entity tagging. The system incorporates in-context learning, system prompts, and post-processing strategies to enhance accuracy across 23 SHI categories. Experimental results demonstrate notable improvements in macro-average F1, increasing from 0.5195 to 0.5615 (+0.042), particularly for ambiguous or variable entities, while highlighting persistent challenges for rare or sparse classes (Our code and data for this paper are made available at: https://github.com/WoZhenDeShenMeDouBuZhidao/aicup-2025 .).