# Machine Learning Prediction of Left Ventricular Assist Device Thrombosis from Acoustic Harmonic Power

## TL;DR

This study uses machine learning to predict thrombosis in HeartWare LVAD patients by analyzing acoustic harmonic power data.

## Contribution

The study demonstrates that machine learning models using harmonic power values can predict LVAD thrombosis with reasonable accuracy.

## Key findings

- Machine learning models using the first two or three harmonic power values showed reasonable thrombosis prediction accuracy.
- Applying PCA to harmonic power variables improved the prediction accuracy of thrombosis outcomes.
- K-nearest neighbor models achieved the best predictive accuracy for the dataset.

## Abstract

Left ventricular assist device (LVAD) thrombosis typically presents late and may have devastating consequences for patients. While LVAD pump thrombosis is uncommon with current pump designs, many patients worldwide remain supported with previous generations of LVADs, including the HeartWare device (HVAD). Researchers have focused on investigating the acoustic signatures of LVADs to enable earlier detection and treatment of this condition. This study explored the use of machine learning algorithms to predict thrombosis from harmonic power values determined from the acoustic signatures of a cohort of HVAD patients (n = 11). The current dataset was too small to develop a predictive model for new data, but exhaustive cross validation indicated that machine learning models using the first two or the first three harmonic power values both resulted in reasonable prediction accuracy of the thrombosis outcome. Furthermore, when principal component analysis (PCA) was applied to the harmonic power variables from these promising models, the use of the resulting PCA variables in machine learning models further increased the thrombosis outcome prediction accuracy. K-nearest neighbor (KNN) models gave the best predictive accuracy for this dataset. Future work with a larger HVAD recording dataset is necessary to develop a truly predictive model of HVAD thrombosis. Such a predictive model would provide clinicians with a marker to detect HVAD thrombosis based directly on pump performance, to be used along with current clinical markers.