Abstract
We present a novel machine learning model to accurately predict the blood-analog viscosity during support of a pathological circulation with a rotary ventricular assist device (VAD). The aim is the continuous monitoring of the hematocrit (HCT) of VAD patients with the benefit of a more reliable pump flow estimation and a possible early detection of adverse events, such as bleeding or pump thrombosis. A large dataset was generated with a blood pump connected to a hybrid mock circulation by varying the pump speed, the physiological requirements of the modeled circulation, and the viscosity of the blood-analog. The inlet pressure and the intrinsic signals of the pump were considered as inputs for the model. Gaussian process yielded models with the best performance, which were then combined using a variant of stacked generalization to derive the final model. The final model was evaluated with unseen testing data from the dataset created. For these data, the model yielded a mean absolute deviation of 1.81% from the true HCT, while it proved to correctly predict the direction of the HCT change. It showed to be independent of the set speed and of the condition of the simulated cardiovascular circulation. The accuracy of the prediction model allows an improvement of the quality of flow estimators and the detection of adverse events at an early stage. The evaluation of this approach with blood is suggested for further validation. Its clinical application could provide the clinicians with reliable and important hemodynamic information of the patient and, thus, enhance patient monitoring and supervision.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have