Explainable Diagnosis Prediction through Neuro-Symbolic Integration.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning
• Neuro-symbolic AI

Background:

• Accurate diagnosis prediction is vital for patient outcomes.
• Traditional AI models often lack the interpretability required for clinical settings.
• Explainable AI is crucial for trust and adoption in healthcare.

Purpose of the Study:

• To develop explainable AI models for diagnosis prediction using neuro-symbolic methods.
• To integrate domain knowledge with machine learning via Logical Neural Networks (LNNs).
• To enhance interpretability in AI diagnostic tools without sacrificing performance.

Main Methods:

• Implementation of LNN-based models, specifically Mmulti-pathway and Mcomprehensive.
• Integration of domain-specific logical rules with learnable weights and thresholds.
• Comparative analysis against traditional models like Logistic Regression, SVM, and Random Forest.

Main Results:

• LNN models achieved superior performance in diabetes prediction compared to traditional methods.
• Achieved high accuracy (up to 80.52%) and AUROC scores (up to 0.8457).
• Learned weights and thresholds provided direct, interpretable insights into feature importance.

Conclusions:

• Neuro-symbolic approaches, particularly LNNs, effectively bridge the gap between accuracy and explainability in healthcare AI.
• The developed models offer transparent and adaptable diagnostic tools for precision medicine.
• Findings support the advancement of equitable healthcare solutions through interpretable AI.