Discovery of an Artificial Intelligence Label Feedback Loop: How the Success of a Clinically Implemented Artificial...

Area of Science:

• Medical Diagnostics
• Artificial Intelligence in Healthcare
• Laboratory Medicine

Background:

• Artificial intelligence (AI) enhances laboratory tests, improving quality and efficiency.
• Clinical AI implementation can lead to unforeseen challenges in model retraining.
• This study identifies an "AI label feedback loop" in an AI-augmented kidney stone composition test.

Purpose of the Study:

• To describe the discovery of an AI label feedback loop.
• To analyze its impact on model retraining and evaluation.
• To propose mitigation strategies for clinical AI systems.

Main Methods:

• Evaluated two versions (V1 and V2) of an AI-augmented kidney stone composition test.
• V2 utilized six times more data than the initial V1 model.
• Performance was assessed on three datasets: AI-influenced labels, pre-AI human labels, and human-corrected labels.

Main Results:

• V2 showed a 10% lower concordance than V1 on AI-influenced data, despite more training data.
• Performance was similar between V1 and V2 on pre-AI, human-labeled data.
• V2 outperformed V1 on human-corrected data, especially for rare kidney stone types, revealing the feedback loop.

Conclusions:

• AI integration in clinical practice can influence test results, complicating AI model development and evaluation.
• Ongoing human annotation and meticulous validation set construction are vital.
• These strategies ensure reliable AI performance assessment and support safe clinical AI evolution.