Page 16 - Artificial Intelligence for Computational Modeling of the Heart

List of figures xv

performance and over-fitting was achieved with polynomials of
degree 3 or 4. 167
Fig. 5.4 Measured and computed ECG traces for one representative cases
◦
(estimation errors of 1.6 ms for QRS duration and 0.5 for
electrical axis). 169
Fig. 5.5 Framework overview: self-taught artificial model personalization
agent. 169
Fig. 5.6 Probabilistic on-line personalization phase. 173
Fig. 5.7 Absolute errors for all patients after initialization with fixed
parameter values (blue, dark gray in print version), after
data-driven initialization for increasing amount of training data
(white), and after full personalization (green, light gray in print
version). Data-driven initialization yielded significantly reduced
2
errors if sufficient training data were available (> 10 ) compared
to initialization with fixed values. Full personalization further
reduced the errors significantly. Red (mid gray in print version) bar
and box edges indicate median absolute error, and 25 and 75
percentiles, respectively. 175
Fig. 5.8 EP results: personalization success rate (blue, dark gray in print
version) and average number of iterations (red, mid gray in print
version). Left: performance for increasing number of training data.
Each dot represents results from one experiment (cross-validated
personalization of all 75 datasets), solid lines are low-pass filtered
means. Right: Performance of both reference methods. Each
shade represents 10% of the results, sorted by performance. 177
Fig. 5.9 Goodness of fit (volume and pressure curves) after personalization
of an example patient based on the different WBC setups.
Additional objectives per setup are highlighted in bold. With
increasing number of parameters and objectives, the proposed
method manages to improve the fit between model and data. 179
Fig. 5.10 WBC personalization results (top: success rate, bottom: average
number of forward model runs until convergence) for the different
setups. Left: RL-based method performance over increasing
number of training data (cross-validated personalization of all 48
datasets). Right: Performance of reference method. Each shade
represents 10% of the results, sorted by performance; darkest
shade: best 10%. 179
Fig. 6.1 Illustration of the virtual CRT modeling pipeline from medical
images and pre-operative, non-invasive measurements to the
heart model. 186
Fig. 6.2 Comparison of QRSd measurements and predictions per
stimulation protocol for (A) case 3 and (B) case 7. 187
Fig. 6.3 Illustration of electrical wave propagation for case 3 (upper) and
case 7 (lower). 188
Fig. 6.4 Quantitative analysis of predictive performance. (A) Mean QRSd
per stimulation protocol; (B) Measured vs. predicted QRSd. 189
Fig. 6.5 Workflow for pre-processing the patient-specific anatomical
models. 193
Fig. 6.6 Transforming 3D surface points to cylindrical coordinates, with