Privacy-Preserving Deep Learning for the Detection of Protected Health Information in Real-World Data: Comparative...

Area of Science:

• Medical Informatics
• Machine Learning
• Data Privacy

Background:

• Collaborative privacy-preserving training enables machine learning on local, private datasets while maintaining confidentiality.
• Ensuring data security is crucial for integrating sensitive information into AI models.

Purpose of the Study:

• To evaluate a state-of-the-art neural network for detecting protected health information (PHI) in texts.
• To assess the performance of privacy-preserving collaborative training for PHI detection.

Main Methods:

• Utilized distributed selective stochastic gradient descent for privacy-preserving training.
• Trained five neural networks on separate, real-world clinical datasets comprising 1304 patient records.
• Employed a privacy-protecting protocol throughout the training process.

Main Results:

• The collaboratively trained networks achieved a mean F1 score of 0.955 for PHI detection.
• Centralized training without privacy considerations reached a slightly higher F1 score of 0.962.
• The performance difference between methods was minimal, highlighting the effectiveness of privacy-preserving approaches.

Conclusions:

• Privacy-preserving collaborative training successfully secures the detection of protected health information in clinical data.
• Deep learning is feasible on distributed and confidential clinical datasets, ensuring robust data protection.
• This approach validates the use of advanced AI techniques in sensitive healthcare settings.