Large language models (LLMs) are increasingly deployed in clinical documentation workflows to alleviate physician burnout and improve efficiency. As encounter transcripts often exceed 2,000-5,000 tokens, prompt compression techniques like LLMLingua have emerged, promising 50-80% token reduction to manage computational costs. However, these generic methods optimize for maximum compression without domain awareness, creating systematic risks in healthcare settings where information loss can directly impact patient safety. This paper presents a critical analysis establishing that clinical text possesses three properties that resist aggressive compression: (1) high information density (predominantly medically relevant content), (2) semantic fragility (single-token changes invert clinical meaning), and (3) liability context (documentation errors cascade to patient harm). We demonstrate through analysis and failure mode examination that generic compression creates dangerous error patterns-negation inversions ("denies chest pain" → "chest pain"), dosage omissions ("metformin 500mg" → "metformin"), and laterality loss ("left knee" → "knee")-that standard NLP metrics like ROUGE fail to detect. We propose the Clinical BERT Safety Gate, a safety-constrained framework with five architectural principles requiring conservative compression limited to demonstrably safe filler removal, domain-aware span protection, and clinical fidelity evaluation. This work challenges the field's efficiency-first paradigm and establishes compression safety as a first-class architectural requirement for clinical NLP systems. Our framework provides actionable guidance for researchers, practitioners, and healthcare AI vendors deploying LLMs in high-stakes clinical applications.

Prompt Compression, Clinical NLP, Patient Safety, LLMLingua

Pan, H., Zhao, X., Liu, Z., Chen, W., Chen, W., & Wong, K. F. (2023). LLMLingua: Compressing prompts for accelerated inference of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 12759-12774).

Pan, H., Zhao, X., Zhang, S., Wang, Z., Liu, Z., Chen, W., & Wong, K. F. (2024). LLMLingua-2: Data distillation for efficient and faithful task-agnostic prompt compression. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (pp. 4907-4928).

Jiang, H., Wu, Q., Lin, C. Y., Yang, Y., & Qiu, L. (2023). LongLLMLingua: Accelerating and enhancing LLMs in long context scenarios via prompt compression. arXiv preprint arXiv:2310.06839.

Li, Y., Wei, Y., Zhang, Y., Zhang, H., & Chen, W. (2023). Compressing context to enhance inference efficiency of large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (pp. 6342-6353).

Chapman, W. W., Bridewell, W., Hanbury, P., Cooper, G. F., & Buchanan, B. G. (2001). A simple algorithm for identifying negated findings and diseases in discharge summaries. Journal of Biomedical Informatics, 34(5), 301-310.

Alsentzer, E., Murphy, J., Boag, W., Weng, W. H., Jindi, D., Naumann, T., & McDermott, M. (2019). Publicly available clinical BERT embeddings. In Proceedings of the 2nd Clinical Natural Language Processing Workshop (pp. 72-78).

Weng, W. H., Wagholikar, K. B., McCray, A. T., Szolovits, P., & Chueh, H. C. (2017). Medical subdomain classification of clinical notes using a machine learning-based natural language processing approach. BMC Medical Informatics and Decision Making, 17(1), 1-13.

van Aken, B., Trajanovska, I., Siu, A., Mayrdorfer, M., Budde, K., & Loeser, A. (2021). Assertion detection in clinical notes: Medical language models to the rescue? Proceedings of the Second Workshop on NLP for Medical Conversations, Association for Computational Linguistics. https://doi.org/10.18653/v1/2021.nlpmc-1.5

Uzuner, Ö., Solti, I., & Cadag, E. (2010). Extracting medication information from clinical text. Journal of the American Medical Informatics Association, 17(5), 514-518.

Sun, W., Rumshisky, A., & Uzuner, O. (2013). Evaluating temporal relations in clinical text: 2012 i2b2 Challenge. Journal of the American Medical Informatics Association, 20(5), 806-813.

Singhal, K., Azizi, S., Tu, T., Mahdavi, S. S., Wei, J., Chung, H. W., ... & Natarajan, V. (2023). Large language models encode clinical knowledge. Nature, 620(7972), 172-180.

Stapleton, B. F. (2015). Evaluating Patient Safety Risks in Pediatric Primary Care Settings. NEJM Journal Watch, NA38047.

Unruh, L., Lugo, N. R., White, S. V., & Byers, J. F. (2005). Managed Care and Patient Safety. The Health Care Manager, 24(3). https://doi.org/10.1097/00126450-200507000-00009

Pereira, E. R., Rocha, R. G., Monteiro, N. da C. A., Oliveira, A. B. de., & Paes, G. O. (2020). Risco de infecção associado ao cuidado no atendimento pré-hospitalar: impactos para a segurança do paciente. Research, Society and Development, 9(8). https://doi.org/10.33448/rsd-v9i8.5846

World Health Organization. (2013). Exploring patient participation in reducing health-care-related safety risks. WHO.

Ratwani, R. M., Bates, D. W., & Classen, D. C. (2024). Patient safety and artificial intelligence in clinical care. JAMA Health Forum, 5(2), e235514. https://doi.org/10.1001/jamahealthforum.2023.5514

Jana, W., Karch, Y., Hummel, T., Jaeger, A., Pfeifer, G.-M., & Austrian Andrologic Working Group. (2023). Lack of patient compliance compromises safety of vasectomy. The Journal of Sexual Medicine, 20(Suppl 1), qdad062.152. https://doi.org/10.1093/jsxmed/qdad062.152

Pivovarov, R., & Elhadad, N. (2015). Automated methods for the summarization of electronic health records. Journal of the American Medical Informatics Association, 22(5), 938-947.