VBridge: Connecting the Dots Between Features and Data to Explain Healthcare Models

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Human-Computer Interaction

Background:

• Machine learning (ML) models show promise for clinical prediction using Electronic Health Records (EHRs).
• Limited model transparency and interpretability hinder the adoption of ML in clinical practice.
• Existing explainable ML techniques present challenges for direct clinical application.

Purpose of the Study:

• To address the challenges of ML interpretability in clinical settings.
• To develop a visual analytics tool, VBridge, that integrates ML explanations into clinical workflows.
• To improve clinicians' understanding and utilization of ML model predictions.

Main Methods:

• Conducted literature surveys and collaborated with six experienced clinicians to identify key challenges.
• Employed an iterative design process to develop the VBridge tool.
• Incorporated a hierarchical display of contribution-based feature explanations and enriched interactions.
• Evaluated VBridge through two case studies and expert interviews with four clinicians.

Main Results:

• VBridge effectively integrates ML explanations into the clinical decision-making workflow.
• The tool's visual association of model explanations with patient records aids clinician interpretation.
• Clinicians showed improved ability to interpret and use ML predictions with VBridge.
• Identified design implications for future explainable ML tools in healthcare.

Conclusions:

• Visual analytics tools like VBridge can bridge the gap between ML predictions and clinical decision-making.
• Addressing clinician-identified challenges is crucial for developing effective explainable AI in healthcare.
• VBridge demonstrates the potential of visual tools to enhance trust and adoption of ML in clinical practice.