Diagrammatization and Abduction to Improve AI Interpretability With Domain-Aligned Explanations for Medical Diagnosis

Research Background and Problem

• Problem and Challenges: Existing explainable artificial intelligence (XAI) methods often rely on simple visualization tools (e.g., bar charts, line graphs, heatmaps), which are difficult to align with the user habits or knowledge systems of complex domains, especially in high-risk areas like medical diagnosis. Users are forced to perform additional reasoning to interpret results, leading to an explanation gap.
• Importance: Decision-making in medical diagnosis has a significant impact on patient health, requiring higher standards for model interpretability. Addressing this issue can help increase expert trust in AI, improve diagnostic accuracy, and reduce misdiagnoses.
• Research Motivation and Related Work: The authors emphasize that explanations should align with domain knowledge and reasoning processes, proposing a framework that supports domain-specific knowledge structures and reasoning methods through diagrammatic representation. This study focuses on cardiac diagnosis, utilizing heart sound auscultation data for diagnosis, and introduces the DiagramNet model.

Solution

• Method Overview: The authors propose a domain-aligned XAI design framework that integrates diagrammatic representation and selective abductive reasoning, applied to the cardiac diagnosis scenario. The proposed DiagramNet model is a modular pre-explainable model that visualizes heart murmur data to explain diagnostic results and employs abductive reasoning to evaluate multiple hypotheses for optimal diagnosis.
• Innovations:
  • Introduced a framework combining diagrammatic representation (using domain-specific diagrams) with abductive reasoning to enhance explanation credibility.
  • Utilized heart sound data and modeled heart murmur shapes as diagrammatic explanations.
  • Provided a formalized abductive reasoning model to clarify relationships between multiple hypotheses and observational evidence.
• Implementation Steps:
  1. Data Preparation: Extract time-series data from heart sound recordings and perform sliding window segmentation to generate training instances.
  2. Diagrammatic Modeling: Use piecewise linear functions to formalize heart murmur shapes under domain constraints.
  3. Model Architecture Design: Develop DiagramNet to implement observation, hypothesis generation, evidence evaluation, and optimal explanation parsing through a modular approach.
  4. Explanation Generation: Generate diagrammatic explanations, comparative explanations, counterfactual explanations, and instance-based explanations.

Research Outcomes

• Specific Results:
  • DiagramNet achieved pre-explainability, directly providing diagrammatic explanations aligned with domain knowledge in cardiac diagnosis.
  • The model not only outperformed other baseline models in predictive performance but also delivered more faithful explanations.
• Comparative Advantages:
  • Compared to existing heatmap-based explanation methods, DiagramNet's explanations are more closely aligned with clinical practice, enhancing user trust.
  • Diagrammatic constraints reduce the potential for spurious explanations, offering highly verifiable explanations.
• Experimental or Evaluation Results:
  • Model Evaluation: DiagramNet demonstrated outstanding performance in prediction accuracy and explanation fidelity. Diagnostic accuracy improved from 86.0% in baseline models to 95.7%.
  • User Study: Interviews and tests with medical students revealed that domain-aligned diagrammatic explanations significantly increased user trust compared to traditional heatmap explanations (trust ratio for diagrammatic explanations was 86%, while heatmaps were only 38%).
• Limitations and Future Directions:
  • Limitations: The current study focuses solely on the cardiac auscultation domain. While medical students were used as experimental subjects, the findings do not fully encompass diverse clinical scenarios or broader expert experiences. Additionally, adaptability to complex diagnostic scenarios has not been tested.
  • Future Directions: The approach could be extended to other high-risk domains, such as skin cancer detection or fairness explanations in bank loan approvals. Further research could explore combining linguistic explanations with diagrammatic explanations to enhance multimodal interpretability.

This study demonstrates how domain-aligned methods can enhance the credibility and practicality of AI explanations, offering a valuable solution for high-risk domains.