Responsible Use of Machine Learning Classifiers in Clinical Practice.

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Medicine
• Medical Ethics

Background:

• Machine learning (ML) models are increasingly integrated into clinical practice for diagnostics and treatment recommendations.
• Current evaluations often focus narrowly on model accuracy and explanation types, neglecting broader responsible use considerations.

Purpose of the Study:

• To propose a framework for the responsible clinical use of machine learning models.
• To define the standard of care for physicians interacting with ML diagnostic and treatment classifiers.
• To explore the epistemic status of ML models and its impact on physician deference.

Main Methods:

• Literature review integrating human factors (automation bias) and social epistemology (higher-order evidence).
• Conceptual analysis of model explanations, historical accuracy, and epistemic status.
• Ethical assessment of physician responsibility in engaging with ML tools.

Main Results:

• Responsible ML use necessitates evaluating not just accuracy but also how explanations are presented and historical performance.
• The 'epistemic status' of a model, encompassing its explanations and accuracy history, should guide physician deference.
• Physician culpability hinges on understanding appropriate engagement with ML classifiers.

Conclusions:

• A multi-disciplinary approach is crucial for establishing standards for ML in healthcare.
• Physicians must critically assess the epistemic status of ML tools to ensure patient safety and ethical practice.
• Future work should delineate clear guidelines for physician responsibility when using ML-driven clinical decision support systems.