Weaving the Digital Tapestry: Methods for Emulating Cohorts of Cardiac Digital Twins Using Gaussian Processes.

Area of Science:

• Computational modeling and simulation
• Personalized medicine
• Engineering applications

Background:

• Digital twin (DT) cohorts model individual assets for applications like in-silico trials and personalized medicine.
• Emulators are used as computationally cheaper surrogates for complex DT models.
• Current methods often emulate each DT member individually, limiting scalability.

Purpose of the Study:

• To introduce novel cohort learning methods for DTs to improve scalability and efficiency.
• To propose statistical approaches for knowledge transfer between DT members.
• To enhance the computational performance and accuracy of DT emulators.

Main Methods:

• Developed two statistical approaches: 'latent-feature emulators' and 'discrepancy emulators'.
• 'Latent-feature emulators' create a single emulator for the entire cohort using a latent representation.
• 'Discrepancy emulators' learn differences between new and existing cohort members.
• Applied methods to cardiac DT case studies.

Main Results:

• Achieved over 50% reduction in computational costs compared to individual emulation in cardiac DT case studies.
• Demonstrated improved computational efficiency and emulator accuracy, even with small cohorts.
• Observed increasing computational savings as cohort size grows.

Conclusions:

• Cohort learning methods enhance both computational efficiency and accuracy of DT emulators by transferring information between members.
• These methods offer significant advantages over individually emulating each DT.
• The proposed transfer methods are applicable to various surrogate models, including Gaussian processes and neural networks.