Using Foundation Models to Prescribe Patients Proper Antibiotics

Simon Austin Lee, Helio Halperin Ortiz, Yanai Halperin Ortiz, Trevor Brokowski, Jeffrey N Chiang

Abstract: The rise of antibiotic-resistant bacteria presents a significant global health threat by reducing the effectiveness of essential treatments. This study evaluates the potential of clinical decision support systems powered by biomedical language foundation models to enhance antibiotic stewardship using electronic health records (EHRs). We test several state-of-the-art models, focusing on predicting whether each of eight different antibiotics will be effective for an individual patient. Additionally, we emphasize interpretability, aiming to understand how the models make decisions, where they excel, and where they fall short. Unlike previous research, which primarily benchmarks accuracy metrics, we provide insights into both the successes and limitations of these models, offering clinical and non-clinical experts a clearer understanding of their current state and reliability. These findings highlight the potential of AI systems to combat this global health threat, as well as the need for further improvements to address the limitations of existing models. We hope this work offers valuable guidance for improving AI-driven decision support systems and leveraging these advanced models for other clinical applications.