(A-272) Enhancing Safety Monitoring in Pharmacoepidemiology through High-Quality, Industry-Specific Translation Technology

Background: Translation is crucial to pharmacoepidemiology safety studies using global data. When leveraging medical records, adverse event reports, or patient reported outcomes data from multiple countries, physician notes must often be translated from the local language to ensure accurate signal detection and risk assessment. Medical concepts may be expressed differently across languages, and accurate translation is critical to maintaining data integrity, mitigating misclassification, and ensuring patient safety globally.

Objectives: To demonstrate how high-quality, industry-specific translation technology can enhance safety monitoring, we highlight case studies showcasing translation technology effectiveness and discuss implications for future research and practice.

Methods: We describe an approach to developing and implementing specialized hybrid “tech-enabled” translation models. This includes integrated Translation Management Systems with proprietary Neural Machine Translation (NMT) engines designed for healthcare use cases such as pharmacoepidemiology. Engines are trained on extensive datasets comprising clinical trial reports, adverse event records, and regulatory documents, ensuring they are robust to the complexities of pharmaceutical safety data. Human linguists review and post-edit translations when necessary, ensuring ‘expert-in-the-loop’ workflows. Large language models (LLMs) are being tested as part of the quality and post editing workflows, showcasing improvements to efficiency and translation quality.

Results: Our NMT output for pharmaceutical documents yields a better Bilingual Evaluation Understudy (BLEU) score compared to generic engines for most commonly used language pairs. BLEURT scores our engines performance in comparison to Google with +8.2% in English (US) to French, +6.1% in English (US) to Chinese (PRC) and +5.2% in English (US) to Portuguese (Brazil). Compared with Deepl, our engine show +10.8% in English (US) to Chinese (PRC), +6.7% in English (US) to French. This higher score indicates better translation quality and fidelity to the original text. Additionally, our Translation Edit Rate (TER) on pharmaceutical documents is 0.33 compared to 0.45 for Google engines. This reduction in TER means that less manual post-editing is required, significantly reducing human review steps and shortening timelines. Continuous training has led to substantial BLEU score improvements the past years, with increases up to 14.2 for Spanish to English, 17 for Italian to English, and 24.9 for German to US English translations. In the last 12 months, our AI enabled workflows were applied to >150 safety programs in >80 languages and the translation of >300 million words across various clinical translation use cases.

Conclusions: Industry-specific translation technology is essential for managing global healthcare data and improving communication among researchers, healthcare professionals, and regulatory bodies. Integrating advanced NMT technology tailored to the pharmaceutical industry is a must to enhance precision and efficiency of safety data analysis, ultimately improving global patient safety. Next development are to involve AI agents for automatic quality and accuracy enhancements which will further reduce TER rates and enhance multilingual safety data analysis.