Human-AI Systems in Medicine: Outskilling Versus Newskilling

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Artificial intelligence (AI) is increasingly adopted in biomedical fields.
• The precise nature of AI as a 'tool' requires definition, particularly in medical diagnostics.
• Convolutional neural networks (CNNs) learn statistical patterns from images, enabling diverse applications.

Purpose of the Study:

• To differentiate between two primary modes of human-AI interaction in medical diagnostics: outskilling and newskilling.
• To analyze the implications of each interaction mode for clinical practice and scientific discovery.
• To inform the design and application of AI tools in healthcare.

Main Methods:

• Conceptual analysis of AI tool utilization in medical diagnostics.
• Case study of AI-assisted polyp detection (computer-aided diagnostics, CADx) as an example of outskilling.
• Case study of AI-derived "retinal age gaps" from fundus images as an example of newskilling.

Main Results:

• Outskilling, exemplified by CADx for polyp detection, aims to augment human performance but risks deskilling without clear outcome benefits.
• "Newskilling" involves AI discovering novel variables (e.g., "retinal age gaps") beyond human inference capabilities, fostering new scientific insights.
• The distinction highlights potential risks of AI-induced deskilling versus the benefits of AI-driven discovery.

Conclusions:

• The design of AI applications should differentiate between augmenting existing tasks and enabling new discoveries.
• Judicious discernment regarding how and when to employ AI tools is essential for maximizing benefits and mitigating risks in medicine.
• Future AI development must consider the balance between enhancing current capabilities and fostering novel scientific and clinical insights.