Why Accurate Pathology AI Is Not Enough

New framework highlights the evidence needed to translate AI performance into better diagnostic decisions and patient outcomes

Category	Detail
Condition	Clinical evaluation and implementation of artificial intelligence in digital pathology
Key Mechanisms	Errors or variation in tissue preparation, image acquisition, model performance, laboratory workflow, clinical decision-making, and patient outcomes can affect downstream diagnostic decisions.
Target Population	Pathology laboratories and pathologists evaluating or using digital pathology AI
Care Setting	Pathology laboratories considering AI adoption or monitoring AI in routine diagnostic workflows

High diagnostic accuracy alone does not establish clinical value or demonstrate improved patient outcomes.
Tissue fixation, staining, section thickness, scanners, and laboratory procedures can affect image quality and AI performance.
The authors recommend external validation in representative patient populations, local testing, and ongoing performance monitoring.
In research involving 28 pathology experts, incorrect AI recommendations overturned previously correct assessments in approximately 7% of consultations.
The proposed six-stage framework is conceptual and has not been validated in prospective studies.

Evaluate whether AI supports appropriate diagnostic decisions, not only whether it achieves high technical accuracy.

Assess compatibility with existing laboratory systems and workflows before adoption.
Test performance across different scanners and patient populations.
Evaluate how pathologists respond to AI-generated recommendations.

Monitor AI performance after implementation.
Reassess performance when laboratory procedures, software, or patient populations change.

Differences in tissue preparation, image acquisition, scanners, and laboratory procedures may affect AI performance across settings.
Inappropriate reliance on AI recommendations may introduce diagnostic errors.

The editorial discusses representative patient populations as necessary for external validation; it does not specify a particular patient group.

Not applicable; the article addresses evaluation and implementation of pathology AI rather than patient treatment.

Use external validation with representative patient populations and conduct local testing.
Assess the full diagnostic pathway, from tissue preparation and image acquisition through workflow, clinical decisions, and patient outcomes.
Seek evidence that AI integrates safely into laboratory workflows and supports appropriate diagnostic decisions.
Continue performance monitoring after implementation, especially when relevant laboratory conditions change.