Accurate Patient Level Reasoning with Medical Language Models

Novel applications of Generative AI enable medical data analysis to go beyond single data points to make deductions about longitudinal, multi-modal patient histories. For example, while NLP can be used to extract tumor characteristics from a pathology report, computer vision can be used to analyze a medical image, or time-series analysis can find anomalies in vital signs or claims data – we are now able to build a unified picture from a patient’s full diverse history, taking all data into account, and using common-sense reasoning to deal with data conflicts and gaps.

This enables applications such as automated creation of patient cohorts, question answering about patient histories, or patient matching applications (to clinical guidelines, to clinical trials, or to research protocols). This session covers a reference architecture for getting this done, using healthcare-specific large language models to deliver state-of-the-art accuracy on reproducible benchmarks. We’ll also cover how to address common challenges such as explaining results, helping clinicians refine their questions, and providing an audit trail of the model’s reasoning.