Emboli detection using a wrapper-based feature selection algorithm with multiple classifiers

Traditionally, analyzing spectral recordings and Doppler shift sounds for detecting emboli is done by experts visually and aurally. These techniques for detecting emboli are both subjective and expensive. In the proposed study, an emboli detection system, which makes binary classification to decide whether a signal is emboli or not, is developed using Q-factor tuned Directional Dual-Tree Rational-Dilation Complex Wavelet Transform for feature extraction. A Doppler ultrasound signal dataset including 400 samples - 200 from embolic, 100 from speckles, and 100 from artifacts - is used. Besides feeding dataset to Support Vector Machines, Multilayer Perceptron, Logistic Regression algorithms for classification; the ensemble voting approach was also applied to obtain higher performance. The experiments including the feature selection and classification algorithms are conducted with an unbiased two-step cross-validation procedure. Firstly, grid-search was used for finding optimum hyper-parameters and features on the training and validation sets. Next, the optimal model was applied to the test set. Lastly, a wrapper-based feature selection algorithm called Boruta was applied to the dataset to overcome insufficient number of samples problem. This problem is very common in biomedical studies due to the difficulties occurring in their creation such as the high-dependency to well-trained human power to acquire meaningful data, and expensiveness in terms of money and time. The results showed that ensemble learning has higher performance than single classifiers when a limited number of training samples is available. Besides, the results point out that close prediction performance was obtained with fewer samples when the Boruta algorithm was applied.

Automated summary

Traditionally, emboli detection is done visually and aurally by experts, but a new study proposes an automated emboli detection system using complex wavelet transform for feature extraction. The research shows that ensemble learning has higher performance than single classifiers when training samples are limited, and using the Boruta algorithm can achieve close prediction performance with fewer samples.