Aims: Inter-subject variability hinders the achievement of the desired efficacy levels atrial fibrillation (AF) treatment. Personalized in-silico models may help with the prediction of the proper treatment for each patient, although reproducing the patient-specific ionic variability can be very time-consuming task, making the clinical use of models difficult.
Methods: A population of 2100 models with 8 currents randomly sampled from -75% to 150% of a baseline chronic AF model was used. The N=922 ionic profiles able to propagate at pacing simulations were simulated in a square tissue (5 x 5 cm, 56776 nodes) after functional re-entry induction. The ability to maintain the re-entry after 6 seconds was evaluated and cycle length (CL) and conduction velocity (CV) measured (a). Machine Learning (ML) was evaluated to predict models that would maintain the re-entry and, if so, what the CL and CV would be, based on the specific ionic profiles.
Results: Preliminary results show that CV show important correlations mainly with gNa (0.83); CL is a multiparametric biomarker depending on gNa (-0.5) and gK1 (-0.61) among other currents (b). When classifying the ionic profiles that could maintain the re-entry (c), random forest ML model showed a sensibility (97%) of and a specificity (60%), on the test sample (20%). Prediction of CV and CL (d), with SVM-model resulted on the test sample in a R2 of 0.92 and 0.9 for CV and CL respectively. The simulation of the 992 ionic profiles took approximately 6 days while the prediction with ML is practically instantaneous.
Conclusion: ML can accelerate the implementation of personalized ionic profiles by rapidly and in mass estimating CV, CL and re-entrant condition for a wide range of currents and selecting those more representative for a patient.