88 | Chapter 5 of all parameter estimations was reported. In case the estimations from different observations fully overlap, MI=100%. In case of no overlap at all, MI=0%. Uncertainty Quantification of Virtual Patient datasets To test the trueness of the estimation, in silico generated virtual patients were generated. To ensure these virtual patients to be representable for real AC patients, nine virtual patients were created based on the nine real patient datasets. For each real patient, the simulation with maximum likelihood was selected. The output of this simulation was used as virtual patient dataset, which was used as input of the modelling framework. Trueness of the virtual estimations was tested by comparing the estimated distribution with the known true parameter values. For each parameter, the highest density interval (HDI) for which the true value is in the interval was calculated. The HDI was defined as the area of the distribution for which the posterior holds p(θ z) >p(θtrue|z). The distribution was approximated with a histogram with bin width defined by the Freedman-Diaconis rule.22 The HDI for each parameter should be near 0% meaning the true value is near the maximum a posteriori. Application in longitudinal datasets Two subjects with a baseline and two follow-up echocardiograms were selected (Table 2). For all six datasets, clinical data was extracted and the datasets were estimated independently of each other, similarly as described above. The two longitudinal sets of estimated tissue properties were investigated. Due to the retrospective nature of this study, LV EDV was only available at baseline. We assumed that it did not change during follow-up. Table 2. Patient characteristics of the two subjects at baseline and follow-up used in the likelihood function Subject 1 Subject 2 Time after baseline (yr) 0 4.5 9.1 0 5.2 7.3 LV EDV (mL) 112 150 LV EF (%) 61 61 61 59 64 57 RVD (mm) 43 43 42 45 38 40 Code implementation The CircAdapt model was written in C++. All other code was written in Python. Each individual dataset was solved sequentially and independently. The source code of the CircAdapt model has been made available before.3 All other source code is publicly available on Zenodo (https:// doi.org/10.5281/zenodo.5084657). Datasets were estimated in parallel with Python 3.9.4 on a AMD Ryzen Threadripper 3970X.
RkJQdWJsaXNoZXIy MTk4NDMw