Uncertainty, Evidence, and the Integration of Machine Learning into Medical Practice.

Area of Science:

• Medical diagnostics
• Machine learning
• Artificial intelligence in healthcare

Background:

• Recent advances in machine learning (ML) for medical applications suggest imminent automation of diagnostics.
• However, significant epistemic challenges must be addressed before widespread clinical adoption.

Purpose of the Study:

• To discuss sources of uncertainty clinicians face when evaluating algorithmic evidence for diagnostic judgments.
• To examine limitations of current ML algorithms, particularly deep learning, in the context of medical diagnostics.

Main Methods:

• Analysis of theoretical foundations of deep learning.
• Examination of algorithmic decision opacity.
• Investigation of ML model vulnerabilities and data quality concerns.

Main Results:

• Deep learning's theoretical underpinnings are not fully understood.
• Algorithmic decisions can be opaque, and models are vulnerable.
• Concerns exist regarding the quality of medical data used for training.

Conclusions:

• Identifying and addressing uncertainty is key for trustworthy AI in diagnostics.
• Desiderata for an uncertainty amelioration strategy are discussed.
• Ensuring ML integration is medically beneficial requires a robust approach to uncertainty.