Development of suction detection algorithms for a left ventricular assist device from patient data
Martin Maw, Christoph Gross, Thomas Schlöglhofer, Kamen Dimitrov, Daniel Zimpfer, Francesco Moscato, Heinrich Schima
Abstract
Left Ventricular Assist Devices are an established treatment option for end-stage heart failure. However, the therapy is still burdened with a high incidence of adverse events. Excessive unloading of the ventricle, also called “suction” has previously been identified as potentially related to worse outcome. In this study, suction was automatically detected without additional sensors in a database of 38 patients who were implanted with the Medtronic HVAD device, containing 500 snapshots with 6258 individual cardiac cycles. A set of 4 classifiers based on 87 features in a wide range of complexities was developed and evaluated, with particular focus on the interpatient variations. The supervised-learning algorithms were trained on the pooled annotation of 6 experts. Analysis was performed on two scales: per-beat -and per-snapshot analysis. A single feature classifier could perform on a similar level to more complex algorithms on a per-snapshot basis (Test sensitivity: 100% specificity: 95.5%). An adaptively boosted tree ensemble classifier managed to achieve higher accuracy on a per-beat basis, but showed signs of overfitting with a reduction in performance from 100% (Interquartile Range (IQR) 0%) in the training dataset to a median sensitivity of 92.5% (IQR 3%) and a median specificity of 100% (IQR 5%) in the testing dataset. The proposed algorithms provide an essential part in assessing the correct level of unloading for the patient, and may be used in different use cases, either as a diagnostic marker, or as a component of an automatic physiological controller.