Machine Learning for Handheld Vascular Studies

There are three main research tasks for this project: 1) the identification of discriminant features of Doppler audio for patient classification, 2) the selection and training of classification algorithms, and 3) CWD audio data enrichment using physics-based models. The investigators will determine which discriminant features are optimal for patient classification from ultrasound Doppler audio. To this end, the investigators will employ signal features in the frequency domain such as bandwidth, peak frequency, mean power, mean frequency, and time harmonic distortion, among others. Furthermore, the investigators will investigate whether time domain features are necessary for accurate sound classification. Other studies have shown that specific features of audio waveforms can classify the data. The investigators will employ some of the most effective machine-learning algorithms for classification such as SVM, logistic regression, and Naïve Bayes, among others. The investigators will start with a binary classification problem in which individuals will be classified as healthy or unhealthy. Then, the investigators will move in complexity to multi-class classification problems in which individuals will be categorized into different groups according to defined abnormal arterial conditions. Data enrichment using physics-based models employing physiologically accurate finite element models of fluid flow in arteries to generate synthetic sound signals corresponding to various arterial conditions. Physics-based simulations would allow the investigators to produce a wealth of training data that can span many known arterial conditions. This capability can augment the classification accuracy and generalization of our algorithms, as clinical data may not be exhaustive enough to incorporate all the known arterial conditions. The investigators will study the performance of the trained algorithms on patient data. To this end, the investigators will partition the data into training and testing samples. The training samples will be used for training of the algorithms, while the testing set will be used to assess generalization capability. The investigators will compute misclassification rates for each algorithm as a metric for performance.