Ann Biomed Eng. 2025 Nov 17. doi: 10.1007/s10439-025-03890-0. Online ahead of print.
ABSTRACT
PURPOSE: Digital twin (DT) cohorts are collections of models where each member represents an individual real-world asset. DT cohorts can be used for in-silico trials, outlier detection and forecasting, and are used across engineering, industry, and increasingly in personalised medicine. To increase the scalability of DT cohorts, researchers often train emulators to be used as cheap surrogates of computationally expensive mathematical models. Frequently, each cohort member is emulated individually, without reference to other members. We propose that instead, we can treat each DT as a thread in a larger network, and that these threads can be woven together into a digital tapestry using cohort learning methods.
METHODS: We propose two statistical approaches for transferring knowledge between threads. The first method, ‘latent-feature emulators’, utilises a latent representation of individual cohort members to generate a single emulator for the entire cohort. The second method, ‘discrepancy emulators’, learns the discrepancy between a new cohort member and existing members.
RESULTS: In two cardiac DT case studies, we show that these methods can reduce computational costs by more than 50% compared to the standard approach of training individual emulators, even in small cohorts.
CONCLUSIONS: We find that by transferring information between meshes, the cohort methods improve both the computational efficiency and the accuracy of emulators when compared to the standard approach of individually emulating each cohort member. As cohort size increases, the computational savings grow further. We focus on the use of Gaussian process emulators, but the transfer methods are applicable to other surrogate approaches such as neural networks.
PMID:41249625 | DOI:10.1007/s10439-025-03890-0
