Healthcare

Current large language models will not fix health care. Here’s what could

STAT, 08/25/23. Advancements in large language models (LLMs) like ChatGPT and GPT-4 have generated substantial excitement, with many seeing them as potential assistants or replacements for time-intensive tasks in various fields, including patient-physician communication in healthcare. However, these models may not have the ability to truly transform healthcare in their current state. While their success can be attributed to their massive scale and training on diverse online sources, healthcare data is not as readily available and leads to blind spots and accuracy issues. A more promising approach would involve the development of a healthcare-specific foundation model, trained on electronic health records and other digital health data, allowing for more accurate predictions and applications unique to health. To achieve this, a considerable amount of data and methods like federated learning for data sharing must be utilized. Additionally, careful consideration of model training is necessary to ensure relevance and effectiveness in healthcare tasks. Despite the challenges, a shared foundation model in healthcare could address regulatory oversight and create a more level playing field, ultimately fulfilling the promise of AI in healthcare. READ THE ARTICLE