Integrating Symbolic Knowledge and Machine Learning in Healthcare

The intersection of Artificial Intelligence and healthcare has driven advancements, particularly through machine learning, which exploits large datasets to develop predictive models and identify risk factors. Despite its success in clinical medicine, only a few models are FDA-approved due to issues of trustworthi- ness and lack of explainability, hindering adoption in clinical settings. Addressing these issues, symbolic knowledge injection and symbolic knowledge extraction have emerged. The first approach integrates domain-specific expertise encoded as rules into machine learning models, while the second extracts interpretable rules from trained models. In this study, this framework is validated using the Pima Indians diabetes dataset, a benchmark in diabetes research. By incorporating a diagnostic protocol for diabetes into machine learning models, the study demonstrates an improvement in the predictive capabilities of these models. By extracting rules from pure data-driven trained models and integrating them with medical knowledge, we reduce false negatives, while achieving a fully explainable diagnostic system. Finally, a combination of these two methods is explored, reporting higher diabetes detection rates and improved model explainability. Accordingly, this study demonstrates the potential of combining machine-learnt insights with medical guidelines to improve healthcare outcomes.