Screening Patient Misidentification Errors Using a Deep Learning Model of Chest Radiography: A Seven Reader Study.

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Patient misidentification in medical imaging poses significant risks.
• Accurate patient identification from paired chest radiographs (CXRs) is crucial for diagnostic integrity.
• Existing methods for patient identification may be prone to human error.

Purpose of the Study:

• To evaluate the performance of deep learning (DL) models in identifying patients from paired CXRs.
• To compare the diagnostic accuracy of DL models against human radiology experts.
• To assess the non-inferiority of DL-based patient identification compared to experienced radiologists.

Main Methods:

• Developed and validated deep learning models using a large dataset of 240,004 CXRs.
• Utilized multiple validation datasets (internal, CheXpert, Chest ImaGenome) representing diverse populations.
• Compared model performance against junior residents, senior residents, and board-certified radiologists in a reader study.

Main Results:

• The SimChest deep learning model demonstrated superior patient identification performance across all datasets (AUC range: 0.933-0.999).
• Radiologists achieved a mean accuracy of 0.900, with performance increasing with experience (0.874 to 0.935).
• SimChest achieved non-inferior performance (P=0.015) compared to human experts, with a mean accuracy of 0.904.

Conclusions:

• Deep learning models, specifically SimChest, can effectively identify patients from paired CXRs.
• DL models offer non-inferior performance to human experts in patient identification from CXRs.
• This technology holds promise for enhancing patient safety by screening for misidentification.