Building a trustworthy AI differential diagnosis application for Crohn's disease and intestinal tuberculosis

Keming Lu, Yuanren Tong, Si Yu, Yucong Lin, Yingyun Yang, Hui Xu, Yue Li, Sheng YuBMC Medical Informatics Decis. Mak 2023 · DOI

TL;DR

This paper introduces a trustworthy AI system for differentiating Crohn's disease (CD) and intestinal tuberculosis (ITB) using endoscopic reports. By combining knowledge distillation, interpretable AI techniques (Integrated Gradients), and a novel debiasing method, the system achieves high accuracy and provides clinically relevant diagnostic features, making it a valuable tool for physicians.

What Problem Does This Paper Address?

Differentiating between Crohn's disease (CD) and intestinal tuberculosis (ITB) based on endoscopic findings is a significant clinical challenge. Both conditions can present with similar symptoms and endoscopic appearances, leading to misdiagnosis and delayed treatment. While AI has shown promise in medical diagnosis, many models operate as "black boxes," lacking interpretability, which hinders clinical adoption. Clinicians need AI tools that are not only accurate but also transparent, allowing them to understand why a diagnosis is made and to trust the system's recommendations. This paper aims to bridge this gap by developing an AI application that is accurate, interpretable, and robust for this specific differential diagnosis task.

Methods

The authors propose a six-step workflow to build a trustworthy AI differential diagnosis system:

  1. Finetune: A pretrained language model (ROBERTa-wwm-ext) is finetuned on Chinese endoscopic reports for classification.
  2. Distill: The large, finetuned model is distilled into a smaller, more efficient TextCNN model. This is done to improve inference speed and reduce computational resources, making the model more practical for deployment.
  3. Interpret: The Integrated Gradients (IG) method is used to obtain local explanations for each sample, identifying features that contribute most to the model's prediction.
  4. Manually Check: Medical doctors review the identified differential diagnosis features. They create a "blacklist" of features that are considered meaningless or potentially misleading (e.g., the disease names themselves).
  5. Debias: A debiasing step is introduced by adding an "attribution penalty" to the loss function during finetuning. This penalty discourages the model from relying on the blacklisted features, forcing it to focus on more clinically relevant ones.
  6. Deploy: The final, robust model is deployed as a web service, allowing doctors to input text descriptions and receive classification results along with visualizations of the model's reasoning.

Datasets

The study utilized a dataset of 1271 electronic health records (EHRs) from patients who underwent colonoscopies at Peking Union Medical College Hospital (PUMCH). These patients were clinically diagnosed with either CD (n=875) or ITB (n=396). The data was collected between January 2008 and November 2018. The dataset was split into an 80% training set and a 20% test set. A "noisy dataset" was also created to evaluate the robustness of the models.

Key Insights and Ideas

  • Knowledge Distillation for Efficiency: Distilling a large PTM into a TextCNN model significantly speeds up inference while retaining high accuracy, making AI diagnostics more practical.
  • Interpretable AI is Crucial for Clinical Trust: The use of Integrated Gradients allows the model to highlight specific phrases and terms that support its diagnosis, enabling clinicians to verify the reasoning and build trust.
  • Debiasing Enhances Robustness and Clinical Relevance: Manually identifying and penalizing "blacklisted" features (like disease names) forces the model to learn more subtle, clinically meaningful diagnostic cues, improving performance on noisy data and aligning with expert knowledge.
  • AI Can Unearth Clinically Relevant Features: The debiased model successfully identified features that were not only statistically significant but also aligned with established clinical guidelines, validating the AI's diagnostic capability.

Key Results

  • The debiased TextCNN model achieved the highest accuracy of 0.83 (CR F1: 0.87, ITB F1: 0.77) on the standard dataset, outperforming baseline models.
  • On the noisy validation set, the debiased model's accuracy was 0.70, significantly higher than models without debiasing.
  • The debiased TextCNN identified 39 diagnostic features in phrase form, with 17 recognized as key diagnostic features by clinical guidelines.
  • The Robust TextCNN model showed the highest recall rate for CD (0.87) and the highest F1 score for ITB (0.77).
  • The study found that while AI can provide strong diagnostic evidence, it may not fully replace the need for other clinical examinations due to the inherent difficulty in differentiating CD and ITB.

Limitations and Future Work

  • Text-Only Input: The current model relies solely on endoscopic report text. The authors acknowledge that incorporating other clinical data (lab tests, pathology reports, etc.) could further improve diagnostic capability.
  • Language and Institutional Variability: Language patterns and reporting styles can vary across institutions, suggesting that the model's portability to different clinical settings requires further testing.
  • Distinguishing Difficult Cases: The paper notes that some patients may remain indistinguishable solely by endoscopy, highlighting the need for continued research into combining AI with other diagnostic modalities.
  • Future Directions: The authors suggest that further clinical and biological research could evaluate how AI-driven feature extraction can push the accuracy limits in differentiating CD and ITB. They also propose that the task-agnostic nature of their method could be applied to other challenging differential diagnosis tasks involving free text.

Summary updated

BibTeX
@article{lu2023building,
  author = {Keming Lu and Yuanren Tong and Si Yu and Yucong Lin and Yingyun Yang and Hui Xu and Yue Li and Sheng Yu},
  title = {{Building a trustworthy AI differential diagnosis application for Crohn's disease and intestinal tuberculosis}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2023},
  doi = {10.1186/s12911-023-02257-6},
}
Figure