Use of Ventricular Assist Device Acoustical Signatures to Detect Device Thrombosis
Abstract
Background: Pump thrombosis (PT), the most common cause of failure in ventricular assist devices (VADs), can be mitigated with earlier diagnosis before triggering any clinical events or requiring VAD replacement.
Hypothesis: Acoustical signatures of VADs can provide substantial information regarding PT.
Methods: We included 10 patients (6 axial, 4 centrifugal VADs) presenting with pump power elevations and recurrent abnormal hemolysis markers consistent with suspected PT. VAD sounds were recorded multiple times using a digital stethoscope (EKO Core) at baseline or before developing thrombosis (normal baseline), during thrombosis episodes and after treatment with heparin or tissue plasminogen activator (Fig. (a)). The 2 types of pumps were analyzed separately (42 recordings from axial and 30 recordings from centrifugal) using 3 steps: (1) Extracting features (time-frequency, acoustic and nonlinearity features) from each recording; (2) Feature importance ranking using Extreme Gradient Boosting (XGBoost) to evaluate which features were most relevant for distinguishing among different states; (3) Principal Component Analysis (PCA) for dimensionality reduction and 2D visualization (Fig. (b)).
Results: We found that separation among the different states is achieved using the combination of XGBoost feature importance ranking and PCA (Fig. (c)). The suggested pipeline of analysis can overcome the problems emerging from inter-subject variability. In both pumps, sounds recorded from VADs with PT had different signatures than baseline sounds. Better separation was observed for axial pumps relative to centrifugal pumps, which may be due to the difference in operating speeds (average 9400 vs 2600 rpms, respectively). We observed a slight overlap between the post treatment and baseline episodes, suggesting the possible presence of residual PT that is not identified by hemolysis biomarkers and/or pump powers.
Conclusion: Applying machine learning algorithms on pump acoustical signals provides substantial information regarding thrombosis states, and potentially allows for sensitive detection of PT assisting hemolysis biomarkers and pump powers.
