Sorano.net

The healthcare landscape is increasingly witnessing the transformative potential of generative artificial intelligence (AI) across various applications, including the automation of clinical documentation and the enhancement of patient communication. Recent advances in large language model (LLM) architectures and transformer based sequence forecasting have opened new perspectives for longitudinal clinical decision support (CDS). However, a significant impediment to achieving equitable and effective healthcare delivery, particularly within the realm of AI applications, lies in the inherent linguistic diversity of global populations. This diversity encompasses not only the multitude of languages spoken worldwide but also the variations within those languages, including dialects and culturally specific expressions (i.e. practice). Generative AI models, often trained predominantly on English language data, frequently encounter difficulties in accurately processing and understanding different languages and their nuances. This limitation can lead to inaccurate translations, misinterpretations of critical medical information, and ultimately, the exacerbation of existing health disparities among different linguistic communities. This project explore the features and capabilities of SNOMED CT in the context of using LLMs and elucidate how they can effectively address the challenges posed by linguistic diversity in the application of generative AI within clinical practice, ultimately fostering more inclusive and accurate healthcare solutions. The presentation will demonstrate how the Foresight Timeline generation tool [i] using synthetic patient data based on various instances of the SNOMED Basic Synthetic Patient Data Generator (BSPG)[ii] performs with various AI models and languages (editions of SNOMED CT).

[i] Foresight: https://foresight.sites.er.kcl.ac.uk/

[ii] BSPG: https://github.com/IHTSDO/health-data-analytics/blob/ui-prototyping/generator/README.md#basic-synthetic-patient-data-generator