Generative artificial intelligence for outcome prediction in critical care: the future is now?
Jessica D. Workum, Christian Jung, Michael BeilPurpose of review
Artificial intelligence (AI) is expected to transform critical care in fundamental ways. The last generation of AI, i.e. generative AI, provides new capabilities to process and generate sequences of data points, such past clinical events, to forecast outcome trajectories. This technology has evolved rapidly in the past year and shown the potential to improve outcome prediction.
Recent findings
The number of publications in this field is still limited indicating major challenges. There are two approaches to use generative AI to predict outcomes – text- and event-based. Whereas large language models (LLM) were deployed to process clinical notes and produce text, generative trajectory models (GTM) analyze sequences of discrete events, such as diagnoses or test results, to predict outcome events. LLM have to extract information from text which may vary substantially in content, granularity and style and, thus, struggle with reliable forecasts. GTM appear to perform significantly better, notably when trained at scale with diverse datasets from various healthcare systems.
Summary
Generative AI may shift prognostication in critical care from static risk scoring to models that sequentially simulate “what happens next,” but the current evidence remains preliminary and does not yet justify incorporation into bedside clinical decision-making.