Automatic Speech Recognition and Large Language Models for Multilingual Pathology Report Generation: Proof-of-Concept Study

Source abstract

Abstract Background Accurate transcription of pathology gross examination dictation is important for clinical documentation, but multilingual dictation remains challenging in settings where clinicians mix Chinese and English while final pathology reports are written in English. Objective This study aimed to evaluate whether a Whisper-based automatic speech recognition (ASR) pipeline guided by contextual system messages and combined with open-source large language models (LLMs; Qwen2:72b, Llama3.1:70b, Gemma2:27b) could improve multilingual (Chinese-English) pathology dictation transcription accuracy and generate clinically appropriate English gross description reports. Methods We conducted a controlled proof-of-concept study using 125 simulated mixed Chinese-English pathology gross examination audio recordings created by physicians or pathologists. Audio recordings were transcribed using Whisper ASR with and without a contextual system message. The ASR transcripts were then converted into English gross description reports using 3 open-source LLMs: Qwen2:72b, Llama3.1:70b, and Gemma2:27b. Outcomes included character error rate, Bilingual Evaluation Understudy, Recall-Oriented Understudy for Gisting Evaluation (ROUGE)-1, ROUGE-2, ROUGE-L, Metric for Evaluation of Translation with Explicit Ordering, pathologist Win-Tie-Lose rankings, report-level error categories, inference time, and interrater agreement. Results The ASR contextual system message reduced the mean character error rate from 0.344 (SD 0.176; 95% CI 0.313‐0.375) to 0.066 (SD 0.100; 95% CI 0.048‐0.084; P <.001). Qwen2:72b achieved the highest automated metric scores, including a Bilingual Evaluation Understudy of 0.644 (SD 0.307), ROUGE-1 of 0.866 (SD 0.163), ROUGE-2 of 0.771 (SD 0.235), ROUGE-L of 0.842 (SD 0.178), and Metric for Evaluation of Translation with Explicit Ordering of 0.805 (SD 0.214). Pathologist-coded total error rates were 16.8% (21/125) for Qwen2:72b, 45.6% (57/125) for Llama3.1:70b, and 92.8% (116/125) for Gemma2:27b. The exact agreement between the 2 pathologists across full ranking categories was 76.8% (96/125; Cohen κ=0.668), and agreement on the top-ranked model or tied top group was 81.6% (102/125; Cohen κ=0.722). Conclusions In this proof-of-concept evaluation, contextual prompting improved ASR transcription accuracy, and Qwen2:72b generated the most accurate English pathology reports among the evaluated LLMs. However, the study used simulated audio recordings, a local vocabulary prompt, and report-level rather than term-level clinical annotation. LLM-generated reports should therefore be considered draft documentation requiring pathologist verification, and prospective validation in real clinical workflows is needed before clinical deployment.