Enhancing Clinician Trust in AI Diagnostics: A Dynamic Framework for Confidence Calibration and Transparency.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Clinical Decision Support

Background:

• AI-driven Decision Support Systems (DSSs) offer potential improvements in diagnostic accuracy and workflow efficiency.
• However, clinician trust is limited by poor confidence calibration, lack of transparency, and misalignment with clinical workflows, leading to high AI override rates.

Purpose of the Study:

• To develop and validate a dynamic scoring framework to improve clinician trust in AI-generated diagnoses.
• The framework integrates AI confidence scores, semantic similarity, and transparency weighting into the override decision process.

Main Methods:

• A dynamic scoring framework was developed and validated using 6,689 cardiovascular cases from the MIMIC-III dataset.
• Override thresholds were calibrated based on varying transparency and confidence levels, with override rate as the primary metric.
• The framework integrated AI confidence scores, semantic similarity measures, and transparency weighting.

Main Results:

• The framework reduced the overall override rate to 33.29%.
• High-confidence (90-99%) AI predictions had a low override rate (1.7%), while low-confidence (70-79%) predictions had a high override rate (99.3%).
• Reduced transparency was associated with higher override rates (73.9% for minimal vs. 49.3% for moderate transparency), with significant associations found (p < 0.001).

Conclusions:

• Enhanced transparency and confidence calibration in AI diagnostics can significantly decrease override rates.
• These improvements are crucial for promoting clinician acceptance and trust in AI diagnostic tools.
• Future research should prioritize clinical validation to optimize AI's impact on patient safety, accuracy, and efficiency.