| Sumario: | Introduction: Locally deployed speech-to-text systems such as Whisper enable privacy-preserving transcription of medical encounters. However, the resulting transcripts are often lengthy, noisy, and insufficiently structured for direct integration into clinical documentation workflows. Large language models (LLMs) can transform such transcripts into concise clinical summaries, yet evaluating their quality and reliability remains challenging, particularly in the absence of gold-standard reference summaries. Materials and Methods: We propose a reference-free evaluation framework for benchmarking LLM-based post-processing of clinical transcripts generated by a locally deployed Whisper system. The framework combines embedding-based transcript-summary semantic similarity, stability analysis across repeated generations, and structured human evaluation. It was applied to 26 German physician-patient conversations. Four LLMs generated four summaries per transcript using standardized prompts and decoding parameters, resulting in 416 summaries. A subset of summaries was additionally assessed by three human raters using six predefined quality criteria. Results: GPT-OSS-120B and MedGemma-27B achieved the highest transcript-summary similarity scores across most embedding models. Although absolute similarity values varied across embedding models, relative model rankings remained largely consistent, indicating robustness of the evaluation framework. Stability analysis showed high consistency across repeated runs, with cosine similarities typically exceeding 0.90, while higher sampling temperatures reduced semantic similarity. Human evaluation showed partial agreement between embedding-based similarity signals and human judgments of summary quality. Conclusion: The proposed framework enables scalable, reference-free evaluation of LLM-based clinical summarization in privacy-sensitive settings. By combining semantic similarity, stability analysis, and human evaluation, it supports systematic model benchmarking, relative model comparison, and optimization without requiring reference summaries. These findings suggest that embedding-based metrics can provide useful signals for selecting and optimizing LLM-based postprocessing models in local speech-to-text pipelines.
|
|---|