Hybrid Explainable AI Medical Diagnosis System: Integrating Rule-Based Reasoning, Naive Bayes, and Transformer

This project aims to resolve the conflict between accuracy and interpretability in the field of medical AI. It builds an accurate and transparent intelligent diagnosis system by integrating rule-based reasoning, Naive Bayes, and Transformer models. The project is from GitHub, authored by manaskirtisinghal, and was published on June 12, 2026. The core idea is to combine the interpretability of traditional rules, the probabilistic reasoning ability of statistical models, and the advantage of Transformer in processing unstructured text to provide reliable assistance for medical diagnosis.

Medical diagnosis is a complex decision-making process. Traditional AI systems face three major challenges: 1. Accuracy bottleneck: A single algorithm can hardly cover the complexity of medical data; rule-based systems are ineffective for ambiguous/rare cases; pure machine learning models fail in data-sparse scenarios. 2. Lack of interpretability: The black-box nature of deep neural networks makes it difficult for doctors to trust them. 3. Knowledge integration difficulty: It is hard to effectively integrate structured medical knowledge (e.g., clinical guidelines) and unstructured experience (e.g., case records).

The project adopts a three-layer hybrid architecture:

1. Rule-based reasoning layer: Encodes explicit logic based on medical experts' knowledge (e.g., consider infection if body temperature exceeds 38.5°C for 3 consecutive days), providing transparent basis and safety boundaries.
2. Naive Bayes layer: Calculates statistical correlations between symptoms and diseases, with advantages of fast training, probabilistic output, and tolerance for missing data.
3. Transformer layer: Processes unstructured medical record text, captures long-range dependencies and contextual semantics, and mines implicit diagnostic patterns.

The system may adopt multiple fusion strategies:

• Weighted voting: Weighted fusion of predictions from the three methods, where the rule-based system may have veto power.
• Cascade architecture: Rule-based preliminary screening → Naive Bayes ranking → Transformer fine-grained discrimination.
• Feature fusion: Concatenate rule features, Bayesian probability features, and Transformer hidden layer representations for joint decision-making.
• Uncertainty guidance: Dynamically adjust method weights to reduce the impact of methods with high uncertainty.

The system is designed with multi-dimensional interpretability mechanisms:

• Rule traceability: Record triggered rules and display diagnostic basis.
• Attention visualization: Show symptom descriptions that the Transformer focuses on.
• Probability decomposition: Decompose Naive Bayes probabilities into contributions from each symptom.
• Contrastive explanation: Generate analysis of how symptom changes affect diagnostic results.

Application scenarios of the hybrid system include:

• Auxiliary diagnosis: Provide second opinions for doctors, assisting in the analysis of rare and complex cases.
• Medical education: Help students understand the diagnostic reasoning process.
• Quality control: Identify risks of missed diagnosis/misdiagnosis.
• Telemedicine: Provide preliminary diagnostic suggestions in primary care and remote scenarios to alleviate resource inequality.

Project limitations: High cost of rule maintenance (requires continuous updates by experts), data privacy issues, and complex regulatory compliance. Future directions: Introduce multi-modal data fusion (imaging, genomics), optimize fusion strategies with reinforcement learning, and implement privacy-preserving collaborative training via federated learning.