Interpretable machine learning text classification for clinical computed tomography reports - a case study of...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Radiology Informatics

Background:

• Machine learning (ML) excels at classifying patient outcomes from clinical notes.
• Interpreting the decision-making process of complex ML models remains a significant challenge.

Purpose of the Study:

• To investigate interpretable text representations for machine learning classification models.
• To enhance physician understanding and trust in ML-driven clinical decision support.

Main Methods:

• Developed ML models (XGBoost, SVM, Logistic Regression, Random Forest) to classify temporal bone fractures from 164 CT text reports.
• Employed average word frequency score (WFS) to quantify keyword differences between positive and negative classifications.
• Utilized Local Interpretable Model-Agnostic Explanations (LIME) for word-level contribution analysis.

Main Results:

• Random Forest achieved a high F1-score of 0.93 for temporal bone fracture classification.
• WFS effectively highlighted keyword usage disparities between fracture and non-fracture cases.
• LIME visualization demonstrated keyword contributions, with LIME-based interpretation achieving 0.97 accuracy.

Conclusions:

• Interpretable text explainers can significantly improve physician comprehension of ML predictions.
• Visualizations from methods like LIME foster greater trust in computerized clinical decision-making.
• This approach promotes transparency in AI applications within healthcare settings.