Translating AI to Clinical Practice: Overcoming Data Shift with Explainability.

Area of Science:

• Medical Artificial Intelligence
• Clinical Translation
• Data Science

Background:

• Generalizability of artificial intelligence (AI) models to real-world clinical data is crucial for practical application.
• Data shift, a mismatch between training and deployment data distributions, is a primary obstacle to AI generalizability.
• Current medical AI often suffers from limited training datasets, leading to performance degradation in real environments.

Purpose of the Study:

• To highlight the role of explainable AI (XAI) in addressing data shift for reliable clinical AI.
• To emphasize the importance of detecting and mitigating data shift in medical AI development.
• To demonstrate how XAI can aid in the clinical translation of AI models.

Main Methods:

• Utilizing explainability techniques during AI training (premodel, in-model, post hoc) to identify susceptibility to data shift.
• Analyzing how performance-based assessments can be insufficient without diverse, external test sets.
• Leveraging XAI as a tool to detect and mitigate failures caused by data shift in the absence of external validation data.

Main Results:

• Explainability techniques can reveal model overfitting to training data biases, which are often masked by standard testing procedures.
• Data shift, if undetected, significantly impacts AI performance in clinical deployment.
• XAI provides critical insights into model behavior beyond standard performance metrics.

Conclusions:

• Explainable AI is essential for detecting and mitigating data shift, thereby enhancing the reliability of AI in clinical practice.
• Without external validation data, explainability methods are vital for ensuring AI models generalize to real-world clinical scenarios.
• XAI facilitates the successful clinical translation of AI by addressing the challenge of data shift.