Predictive models play an increasingly important role in clinical decision making and risk assessments. Despite the progress in this area, these models typically do not leverage medical knowledge in medical literature and knowledge bases. At the same time, Evidence-Based Medicine (EBM) involves searching the literature for verification to support clinical decisions, however this evidence is not used as part of predictive models. Endowing models with such knowledge and evidence is important to enhance predictive accuracy, robustness, and human interpretability – for diagnosis, prognosis, and treatment.
The researcher is developing a new generation of predictive models that have access to both well-established medical knowledge, and cutting-edge evidence in the literature. These novel systems automatically find informative signals in the raw text of academic literature as well as in structured scientific knowledge bases, using the retrieved signals to enhance predictions. The researcher has built a prototype for patients hospitalized in the intensive care unit (ICU), in which a system automatically searches for patient-specific literature in PubMed papers to find text that helps predict future patient outcomes, such as in-hospital mortality or prolonged length of stay. A neural language model is leveraged for retrieval of scientific knowledge. The model combines papers’ textual representations with patient data in a shared embedding space.
In preliminary experiments, promising results indicate the potential of this AI application. Utilizing AI retrieval methods will provide several improvements, including accuracy in patient outcome predictions, unlike the current standard models that do not have this tool.
There is much potential in bridging the gap between AI-based clinical models and the Evidence Based Medicine (EBM) paradigm in which medical decisions are based on explicit evidence from the literature. Incorporating evidence identification and inference directly into retrieval and predictive models is a promising avenue for advancing model performance as well as interpretability. Finally, the new approach can be applied across biomedical predictive tasks (e.g., in drug discovery), where any predictive model can be automatically enhanced with systems that mine the literature and external knowledge bases for useful signals.
Code available here: https://github.com/allenai/BEEP