Bayesian logical neural networks for human-centered applications in medicine.

Area of Science:

• Biomedical informatics
• Artificial intelligence in medicine
• Data science

Background:

• Medicine faces inherent uncertainty due to imperfect data, especially in Electronic Health Records.
• Noisy and unstructured data in healthcare impede accurate interpretation and AI model performance.
• Epistemic uncertainty is prevalent across biomedical research, affecting clinical decision support.

Purpose of the Study:

• To develop a novel AI methodology for modeling uncertainty in medical data.
• To enhance the reliability of AI-driven clinical recommendations.
• To create a user-centered system that informs physicians about recommendation uncertainty.

Main Methods:

• Combined structural explainable models (Logic Neural Networks) with Bayesian Networks.
• Utilized logical gates within neural networks to replace conventional deep learning.
• Developed adaptive models to handle data variability and inherent uncertainty in medical procedures.

Main Results:

• The developed model assists physicians by providing accurate recommendations.
• The system explicitly informs clinicians when a recommendation, such as a therapy, is uncertain.
• Demonstrated the model's capability to adapt to input data uncertainty for personalized recommendations.

Conclusions:

• The novel methodology enhances clinical decision-making by addressing data uncertainty.
• The AI system acts as a user-centered tool, promoting critical evaluation of recommendations.
• Tested on heart insufficiency patient data, this approach offers a foundation for future medical recommender systems.