Benchmarking Interpretability in Healthcare Using Pattern Discovery and Disentanglement.

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Informatics
• Machine Learning for Medicine

Background:

• Healthcare systems increasingly use AI, but their 'black box' nature hinders practitioner trust and understanding.
• Interpreting AI decision-making is crucial for safe and effective clinical integration.

Purpose of the Study:

• To benchmark the Pattern Discovery and Disentanglement (PDD) unsupervised learning algorithm for clinical note analysis.
• To evaluate PDD's ability to provide interpretable outputs and aid healthcare decision-making.
• To compare PDD's performance and interpretability against supervised deep learning models and post-hoc interpretability techniques.

Main Methods:

• Utilized the MIMIC-IV dataset, processing clinical notes and ICD-9 codes using Term Frequency-Inverse Document Frequency and Topic Modeling.
• Applied the PDD algorithm for feature discretization, pattern discovery in a disentangled statistical space, and clinical record clustering.
• Benchmarked post-hoc interpretability methods (Feature Permutation, Gradient SHAP, Integrated Gradients) against PDD's global interpretability.

Main Results:

• PDD demonstrated unsupervised clustering performance comparable to supervised deep learning models.
• The PDD algorithm generated an interpretable knowledge base linking clinical data, patterns, and knowledge.
• Post-hoc interpretability techniques showed limitations in clinical diagnosis compared to PDD's global interpretability.

Conclusions:

• PDD offers a viable solution for enhancing AI interpretability in clinical settings.
• The system's global interpretability aids practitioners in understanding AI decision processes.
• PDD effectively clusters diseases, providing valuable insights to support clinical diagnosis.