Figure 1

Design of experiments workflow to develop ML models. (a) 65% CoA geometry. (b) The five flow waveforms that were used as features representing a range between rest and exercise. (c) Design of experiments workflow. Step 1) Run 50 simulations in the patient-specific CoA (combination of ten viscosity values and five flow rates). Each simulation can be defined by a viscosity-flow rate pairing. Step 2) In the patient-specific CoA, train and test a neural network to predict ΔP from the simulations. Select 40 simulations for training and ten for the test set. Step 3) If the correlation coefficient R >= 0.98 in the test set, reduce the training set size by one, add that simulation to the test set, and return to Step 2. Otherwise, record the minimum simulation set defined by the viscosity-flow rate pairings in the training set, and return the correlation coefficient for analysis. Step 4) Test the robustness of the ML model by using the minimal simulation set defined by the viscosity-flow rate pairings to run simulations on two different CoA geometries. Also, run nine more simulations in the new geometries to use as a test set. Step 5) In the two new geometries, train and test neural networks to predict ΔP and compute the correlation coefficients in the test set. (d) ML model results for the 50 simulation set in the 65% CoA. (e) ML model results for one of the new 65% CoA geometries using nine simulations in the training set and nine simulations in the test set.