Intelligent reflux and suction detection system for ventricular assist devices: in silico study

Left ventricular assist devices (LVADs) serve both as a bridge to transplantation and as destination therapy for the treatment of congestive heart failure (CHF). However, the inability of the existing control strategies to automatically adapt LVAD flow according to hemodynamic changes can significantly impact patient survival and quality of life. Physiological control strategies for LVAD show promising results, with reflux and suction detection (RSD) increasing device safety.

This study presents in silico results of an RSD system based on measurements of inlet and outlet pressures in continuous-flow LVAD. Two strategies were used to investigate control feasibility, safety, and adjustments to nonlinear variations and comprehensively assess the system’s state considering a structured algorithm (SA) and ensembles of AI models (eAIm): K-nearest neighbors (KNN), support vector machine (SVM), and artificial neural network (ANN).

The SA submodule achieved an accuracy of 99.66% in suction detection but showed limitations in reflux events, with 80.04% accuracy and an F1-Score of 70.4%. The KNN and SVM models demonstrated performance exceeding 96% for both events, exhibiting more excellent stability than the SA submodule. The ANN excelled with low variability and an RMSE of 0.07 in R1, though its suction accuracy (96.7%) was slightly lower than for reflux (99.48%). The KNN was the most effective model, achieving 99.66% accuracy in suction and 98.44% in reflux. The SVM also produced competitive results but with variability across evaluations. The eAIm model showed satisfactory precision (97.78% for suction and 97.14% for reflux), with variations depending on the scenario. The eAIm is recommended for optimization in precision-critical situations.

These strategies are designed to fulfill the proposal’s feasibility, flexibility, and safety requirements. They address the challenges of achieving consistent reproduction using an SA and the capability to handle nonlinear patterns. Additionally, they encompass the comprehensive assessment of the system's state through AI models and ensembles.

This approach is unprecedented in identifying new waveform patterns to classify suction and reflux events and represents the first application of AI for integrating both adverse events, with improved robustness of results achieved through combining multiple AI models. The RSD system demonstrated excellent results, suggesting viability for application to PC-LVADs. The elevated results from the in silico study indicate only potential promise, with anticipated accuracy reduction in practical applications. Effective pattern detection in computational simulations highlights the need to assess the precision of the developed systems in in vitro and clinical studies.