Neuromuscular disorders, which affect millions of people in Europe alone, lead to (progressive) muscle weakness or sensory deficits that gravely affect life expectancy and quality of life. To diagnose the disorders, needle electromyography (nEMG) data must be assessed audio-visually by experts, which is subjective and time-consuming. In this project, experts in clinical neurophysiology, data science and instrumentation will develop an artificial-intelligence platform to automatically, objectively and accurately interpret nEMG data. They will validate the method using real nEMG data from around the world, and take first steps towards integrating the platform into existing software for clinical use.

To address emergent global environmental problems, geoscientists envision creating a "Digital Twin of Earth" by fully modeling environmental processes from the data collected by Earth Observation (EO) satellites (e.g., ESA Sentinel, NASA Lansat). Two different modeling paradigms are available for creating such a digital twin. On the one hand, there exists a group of widely-adopted theory-driven physical radiative transfer models for environmental parameter estimation from EO data. These models often lead to inaccurate results due to their ill-posed nature. Data-driven models acquired from machine learning algorithms, on the other hand, can offer higher accuracy in environmental parameter estimation by capturing global spatio-temporal patterns in data. This quality is, however, attained at the cost of physical consistency. Our goal in this proposal is to realize a hybrid framework that benefits from the physical consistency of theory-driven models and the spatio-temporal patterns captured by the data-driven models. Combining these models, we can for the first time open the door to generating highly accurate forecasting models of detailed environmental parameters. Three main challenges need to be addressed in exploring the combining of these models: (1) representing interacting spatio-temporal processes from high-dimensional and noisy EO data; (2) configuring the combined model with its complex parameter space resulting from the combination of theory-driven and data-driven models; and (3) acquiring ground-truth data of environmental parameters for training data-driven models. We aim to address the first challenge by designing robust multi-aspect representations based on neural representation learning that can capture relevant spatial and temporal structures in data. We address the second challenge by designing an automated method for configuring the combined model using the Bayesian optimisation framework. The third challenge is tackled by defining new objective functions that weakly depend on ground-truth data relying on the knowledge acquired from globally consistent spatio-temporal patterns, alongside statistical indicators. Together with a user committee composed of the European Space Agency (ESA), the main organisation in charge of collection of EO data in Europe, and environmental scientists, we demonstrate the applicability of these hybrid models in biophysical parameter estimation.