Rapid personalisation of cardiovascular models using invasively measured right ventricular pressure

Patient digital twins offer the promise of providing personalised healthcare via the continual updating of computational models that are adjusted to deliver refined predictions and insights into a patient's physiological state. However, clinical information available for personalisation is diverse and constrained by practical clinical limitations, raising important questions about how to optimally leverage the available data. In this work, we pursue two main objectives: (1) to develop a real-time personalisation and calibration framework for cardiovascular digital twins, and (2) to determine how different representations of the same underlying clinical data affect the identifiability and recovery of haemodynamic parameters during calibration. Using a lumped-parameter cardiovascular model, we evaluate the relative utility of discrete clinical indices, pressure waveform data, and principal components. We found that combining temporally rich principal components of the data with clinical indices provided the most information for calibration. However, the model exhibits some sloppiness causing identifiability issues, and demonstrates that the choice of data representation directly influences which haemodynamic parameters can be estimated. To achieve real-time personalisation, we use linear emulators that enable essentially instantaneous Bayesian calibration of model parameters, providing an efficient approach that minimises computational burden. Our methodology and findings are demonstrated using invasively recorded right ventricular (RV) pressure in patients with pulmonary arterial hypertension (PAH), demonstrating that linear emulation coupled with discrete clinical indices can provide real-time model personalisation, and is a stepping stone towards the development of digital twins of the cardiovascular system.