Methodology and Real-World Applications of Dynamic Uncertain Causality Graph for Clinical Diagnosis with Explainability and Invariance

Zhang Zhang 0002, Qin Zhang 0007, Yang Jiao, Lin Lu, Lin Ma, Aihua Liu, Xiao Liu, Juan Zhao, Yajun Xue, Bing Wei, Mingxia Zhang, Ru Gao, Hong Zhao, Jie Lu, Fan Li, Yang Zhang, Yiming Wang, Lei Zhang, Fengwei Tian, Jie Hu, Xin Gou

TL;DR

This paper introduces the Dynamic Uncertain Causality Graph (DUCG), a novel AI approach for clinical diagnosis that prioritizes explainability and invariance across diverse scenarios. DUCG moves beyond black-box models by leveraging expert knowledge to build causal graphs, achieving high diagnostic accuracy and enabling clinicians to understand why a diagnosis is made, fostering trust and improving diagnostic skills.

What Problem Does This Paper Address?

The field of AI in medicine has largely focused on image analysis, often resulting in "black-box" models that lack transparency. This makes it difficult for clinicians to trust or understand the AI's reasoning, especially for complex diagnoses requiring the integration of various data types (symptoms, lab results, patient history). Furthermore, many AI models struggle with generalization across different clinical settings (e.g., large hospitals vs. village clinics) and have difficulty accurately diagnosing rare diseases due to data imbalances. The paper addresses these limitations by proposing a causality-driven, explainable, and invariant AI system that can be trusted and adopted by medical professionals.

Methods

The core of this work is the Dynamic Uncertain Causality Graph (DUCG). Unlike data-driven deep learning models, DUCG is built upon expert knowledge and causal relationships. Key aspects of the methodology include:

  • Causality-Driven Approach: DUCG models represent diseases and their relationships with symptoms, risk factors, and other variables as a graph, explicitly encoding causal links.
  • Uncertainty Handling: The "Uncertain" in DUCG refers to the probabilistic nature of these causal relationships, allowing for nuanced reasoning.
  • Explainability: The graph structure itself provides a visual and logical explanation for diagnoses. Clinicians can trace the reasoning path.
  • Invariance/Generalization: DUCG aims to be invariant across different application scenarios, meaning the underlying causal knowledge remains consistent even with varying data dimensions or clinical settings. This is achieved by building modules for individual diseases and then synthesizing them into larger models.
  • Modular Construction: DUCG models are constructed by first creating transparent and explainable modules for single diseases under a chief complaint. These modules are then synthesized into a comprehensive DUCG model for that chief complaint.
  • Expert-Driven Parameterization: While parameters can be learned from data, the primary approach involves domain experts providing the causal relationships and their intensities.
  • Recommendation Algorithm: A novel algorithm is presented to guide clinicians in selecting the most informative next medical check to perform, balancing diagnostic value, disease danger, and cost.

Datasets

The paper does not explicitly mention specific datasets used for training DUCG models. Instead, it emphasizes that DUCG construction relies on collaboration between clinical experts and technicians, with parameters often provided by domain experts. The verification and real-world application sections describe the use of discharged case records from third-party hospitals and real-world diagnosis cases from various medical units in China.

Key Insights and Ideas

  • Expert Knowledge as a Foundation: DUCG demonstrates that leveraging established medical knowledge and causal relationships can be a powerful alternative to purely data-driven approaches, especially for complex and rare conditions.
  • Explainability Builds Trust: The inherent transparency of DUCG allows clinicians to understand the diagnostic process, fostering trust and enabling them to validate or correct the AI's reasoning. This is crucial for AI adoption in critical fields like healthcare.
  • Invariance is Key for Real-World Deployment: The ability of DUCG to maintain diagnostic performance across different clinical settings (hospitals, clinics) without requiring scenario-specific retraining is a significant advantage for widespread adoption.
  • Uncommon Diseases are Not an Obstacle: By focusing on causal relationships rather than just data frequency, DUCG can theoretically diagnose uncommon diseases as effectively as common ones, addressing a major limitation of many current AI models.
  • Dynamic Updates and Learning: While not data-driven in its core construction, DUCG allows for updates and corrections based on expert feedback and real-world performance, creating a continuously improving system.

Key Results

  • High Diagnostic Precision: 46 DUCG models covering 54 chief complaints and over 1,000 diseases were verified by third-party hospitals. The overall diagnostic precision was no less than 95%, with precision for every disease, including uncommon ones, no less than 80%.
  • Extensive Real-World Application: Over one million real diagnosis cases have been performed in China using DUCG models.
  • Low Error Rate: Out of over one million cases, only 17 incorrect diagnoses were identified. These were traced, corrected, and no similar errors have been reported since.
  • Improved Clinician Skills: Clinicians who frequently used DUCG showed significant improvement in their diagnostic abilities.
  • Cost-Effectiveness: DUCG requires minimal hardware (less than $10,000 for a server) and is computationally efficient (within 1 second per diagnosis).

Limitations and Future Work

  • Limitations:
    • DUCG cannot directly recognize medical images or sounds.
  • The accuracy of DUCG relies on the quality of expert knowledge and the completeness of the causal graph.
  • Prospective studies for verification are still needed, although real-world feedback has been extensive.
  • The current application is primarily in Chinese, with a need for English and other language versions.
  • Future Work:
    • Conduct more prospective studies to further validate diagnostic precisions.
  • Expand DUCG applications to more areas and perform more third-party verifications.
  • Collaborate with data-driven AI to integrate information from medical images and sounds.
  • Develop DUCG models for rare disease diagnoses and traditional Chinese medicine.
  • Create English and other language versions of DUCG.

Summary updated

BibTeX
@article{zhang2024methodology,
  author = {Zhang Zhang and Qin Zhang and Yang Jiao and Lin Lu and Lin Ma and Aihua Liu and Xiao Liu and Juan Zhao and Yajun Xue and Bing Wei and Mingxia Zhang and Ru Gao and Hong Zhao and Jie Lu and Fan Li and Yang Zhang and Yiming Wang and Lei Zhang and Fengwei Tian and Jie Hu and Xin Gou},
  title = {{Methodology and Real-World Applications of Dynamic Uncertain Causality Graph for Clinical Diagnosis with Explainability and Invariance}},
  journal = {Artif. Intell. Rev.},
  year = {2024},
  doi = {10.1007/s10462-024-10763-w},
}
Figure