AI-aided general clinical diagnoses verified by third-parties with dynamic uncertain causality graph extended to also include classification

Zhan Zhang, Yang Jiao, Mingxia Zhang, Bing Wei, Xiao Liu, Juan Zhao, Fengwei Tian, Jie Hu, Qin ZhangArtif. Intell. Rev 2022 · DOI

TL;DR

This paper introduces an extended Dynamic Uncertain Causality Graph (DUCG) model that incorporates classification relationships, addressing limitations of traditional machine learning in clinical diagnosis. The enhanced DUCG achieves high diagnostic precisions (96.5-100%) across six chief complaints, demonstrating strong generalization ability and interpretability, crucial for real-world medical applications.

What Problem Does This Paper Address?

Artificial intelligence (AI) holds immense promise for aiding clinicians in diagnosis, but current machine learning (ML) models often struggle with generalization and interpretability. This means models trained on specific datasets may perform poorly when applied to different real-world scenarios, and their decision-making processes can be opaque ("black box"). Furthermore, ensuring accurate diagnosis of rare diseases, which are often underrepresented in training data, remains a challenge. The authors aim to overcome these limitations by developing a more robust and interpretable AI diagnostic system.

Methods

The core of this work is the extension of the Dynamic Uncertain Causality Graph (DUCG) model. DUCG is designed to explicitly represent causal relationships with uncertainties, allowing for probabilistic reasoning. The key innovation here is the introduction of "C-type variables" to handle non-causal classification relationships. These variables, along with a unit matrix representation, allow for a more accurate and intuitive modeling of medical knowledge.

The authors construct six DUCG knowledge bases, each corresponding to a chief complaint (e.g., arthralgia, dyspnea). These knowledge bases are built by experts, synthesizing subgraphs that represent diseases and their causal links to symptoms. The inference process, while constructed with C-type variables for interpretability, is performed using an equivalent DUCG model without them, ensuring computational efficiency.

Datasets

The paper does not explicitly mention training datasets. However, it details the verification process using case records from the hospital information systems (HIS) of third-party hospitals. These case records served as the basis for evaluating the diagnostic precisions and generalization abilities of the developed DUCG models.

Key Insights and Ideas

  • C-type variables bridge the gap: The introduction of C-type variables is crucial for accurately representing non-causal relationships, which are common in medicine (e.g., a test result being an indicator rather than a direct consequence of a disease). This significantly improves the model's ability to capture complex medical knowledge.
  • Invariance is key to generalization: The authors argue that domain knowledge, as captured by DUCG, possesses invariance, which is fundamentally different from the data-driven nature of ML models. This invariance is presented as the reason for DUCG's strong generalization ability across different application scenarios.
  • Interpretability is paramount: The DUCG framework prioritizes interpretability, allowing clinicians to understand the reasoning behind a diagnosis. This is achieved through the graphical representation of knowledge bases and the ability to trace causal pathways.
  • Third-party verification is essential: The rigorous verification by independent hospitals underscores the practical applicability and reliability of the DUCG system.

Key Results

The DUCG-aided general clinical diagnoses demonstrated impressive results:

  • High Diagnostic Precisions: Total diagnostic precisions for the six chief complaints ranged from 96.5% to 100%.
  • Robustness for Rare Diseases: The precision for every individual disease was no less than 80%, indicating the model's ability to handle less common conditions.
  • Strong Generalization: The close agreement in verification results between two independent groups of third-party hospitals strongly suggests that the DUCG system can be applied in various real-world scenarios without a significant drop in precision.

Limitations and Future Work

  • Limited scope of chief complaints: The study focused on six specific chief complaints. Future work could expand this to a wider range of medical conditions.
  • AI-aided medical image and sound recognition: The current DUCG model does not handle AI-aided medical image examination or medical sound recognition. The authors suggest that these areas could be addressed by ML models, implying a potential for hybrid systems.
  • Outpatient data: The verification process primarily used discharged patient case records, which are generally of higher quality. The authors acknowledge that verifying diagnostic precision with less information from outpatient records is an open issue for future research.
  • Next medical checks: In real applications, the AI-aided system could be enhanced to recommend the next medical checks needed to gather further information for more accurate diagnoses. This is mentioned as a topic for a future paper.

Summary updated

BibTeX
@article{zhang2022ai,
  author = {Zhan Zhang and Yang Jiao and Mingxia Zhang and Bing Wei and Xiao Liu and Juan Zhao and Fengwei Tian and Jie Hu and Qin Zhang},
  title = {{AI-aided general clinical diagnoses verified by third-parties with dynamic uncertain causality graph extended to also include classification}},
  journal = {Artif. Intell. Rev.},
  year = {2022},
  doi = {10.1007/s10462-021-10109-w},
}
Figure