Patient Reidentification from Chest Radiographs: An Interpretable Deep Metric Learning Approach and Its Applications.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Deep Learning and Machine Learning

Background:

• Patient reidentification in medical imaging is crucial for data privacy and longitudinal studies.
• Explainable AI (XAI) is increasingly important for understanding model decisions in clinical settings.
• Detecting subtle changes in longitudinal imaging can indicate early-stage disease.

Purpose of the Study:

• To develop an explainable deep learning model for patient reidentification in chest radiograph datasets.
• To assess if changes in model-identified patient features over time can serve as markers for emerging radiologic abnormalities.

Main Methods:

• A deep learning model was trained on over 1 million chest radiographs from multiple hospitals.
• The model was validated on public datasets (ChestX-ray14, CheXpert, MIMIC-CXR).
• A generative adversarial network (GAN) was used for visual explanations of model features.

Main Results:

• The model achieved high performance in patient reidentification (F1 score up to 0.996) and database retrieval (Precision at 1 up to 0.976).
• Key identified features included patient sex, age, and weight.
• The model showed potential for abnormality prediction (AUC 0.73) compared to age prediction error (AUC 0.58).

Conclusions:

• The deep learning model's features for reidentification are human-interpretable.
• Changes in these features over time may indicate emerging radiologic abnormalities.
• This approach offers a novel method for monitoring patient health through longitudinal chest radiograph analysis.