Use of Client-Side Machine Learning Models for Privacy-Preserving Healthcare Predictions - A Deployment Case Study.

Area of Science:

• Computational Science
• Medical Informatics
• Machine Learning

Background:

• Traditional server-centric machine learning (ML) and deep learning (DL) models in healthcare involve transmitting sensitive patient data to external servers, posing significant privacy risks.
• Existing frameworks like Flask facilitate server-side processing, highlighting the need for more secure alternatives.

Purpose of the Study:

• To demonstrate a privacy-preserving approach for executing healthcare prediction models entirely within a web browser.
• To eliminate privacy concerns associated with data transmission in traditional healthcare ML/DL applications.

Main Methods:

• Leveraging browser-based ML/DL technologies including TensorFlow.js and ONNX Runtime Web.
• Implementing three strategies based on model complexity: direct JavaScript for simple models, ONNX Runtime Web for medium-complexity models (e.g., Random Forest), and TensorFlow.js for complex deep learning models (e.g., Optimized Convolutional Neural Networks).
• Ensuring all computations are performed client-side on the user's device.

Main Results:

• Client-side deployment of healthcare prediction models is feasible and effective.
• Original performance metrics of the models are preserved during client-side execution.
• Substantial privacy benefits are achieved by keeping patient data localized.

Conclusions:

• Patient data is guaranteed to remain on the user's device, mitigating data transmission risks.
• This approach is particularly advantageous for healthcare settings prioritizing data confidentiality.
• The client-side execution model also supports offline functionality for healthcare applications.