A Roadmap for Automating Lineage Tracing to Aid Automatically Explaining Machine Learning Predictions for Clinical...

Area of Science:

• Clinical informatics
• Machine learning applications
• Data science

Background:

• Machine learning (ML) models offer potential for clinical decision support, improving patient outcomes and reducing costs.
• A significant barrier to ML adoption in healthcare is the 'black box' nature of most models, hindering interpretability.
• Existing automated explanation methods provide rule-style insights but lack detailed data drill-through capabilities.

Purpose of the Study:

• To enhance automated ML model explanation functions by enabling users to drill down into raw data.
• To address the challenge of understanding aggregated feature values within ML model explanations.
• To facilitate better clinical decision-making by providing deeper insights into patient data.

Main Methods:

• Development of an automated lineage tracing approach.
• Integration of drill-through capability into existing automated explaining functions for ML models.
• Focus on rule-style explanations derived from tabular data.

Main Results:

• The automated lineage tracing approach enables users to rapidly access relevant raw data underlying aggregated features.
• This enhances the interpretability of ML model predictions by connecting high-level explanations to specific data points.
• Facilitates a more thorough understanding of patient situations for clinical decision support.

Conclusions:

• Automated lineage tracing significantly improves the utility of ML explanation tools in clinical settings.
• This approach helps overcome the transparency barrier, promoting greater trust and adoption of ML in healthcare.
• Future research should focus on refining and expanding these automated interpretability techniques.