Data drift in medical machine learning: implications and potential remedies.

Area of Science:

• Medical Artificial Intelligence
• Machine Learning Operations (MLOps)
• Data Science in Healthcare

Background:

• Data drift occurs when training data differs from real-world operational data in medical machine learning (ML) systems.
• Causes include variations in data sampling, clinical practices, patient populations, disease patterns, and data acquisition over time.
• Medical imaging applications are particularly susceptible to diverse data drift scenarios.

Purpose of the Study:

• To review ML data drift terminology, types, and causes in medical applications, focusing on medical imaging.
• To synthesize existing literature on the impact of data drift on medical ML system performance.
• To discuss methods for monitoring and mitigating data drift, including pre- and post-deployment techniques.

Main Methods:

• Literature review of data drift concepts and their impact on medical ML.
• Categorization of data drift types and their specific causes in clinical settings.
• Analysis of drift detection and model retraining strategies.

Main Results:

• Data drift is a primary driver of performance degradation in medical ML systems.
• Existing research overwhelmingly confirms the negative effects of data drift.
• Effective drift monitoring and mitigation are crucial for maintaining model efficacy.

Conclusions:

• Data drift poses a significant challenge to the reliable deployment of ML in healthcare.
• Further research is essential for developing ML models that can proactively identify and address data drift.
• Enhanced strategies for early drift detection and mitigation are needed to prevent performance decay in medical ML applications.