Machine learning classifiers

We trained three machine learning classification models in order to evaluate the predictive power of certain physiological features for the presence or absence of vocalizations. Five custom features were defined and computed from the ECG and BP in order to predict vocalizations: heart rate, respiratory rate, respiratory amplitude, blood pressure and Mayer wave power. We used the standard Scipy and Numpy libraries in Python 3.8.5 for all computations. We tested the performance of the three classifiers based on the mean value of each feature using a 2-s time window prior to call onset. The scikit-learn toolbox (Pedregosa et al., 2011) in Python was used to train three different classifiers: logistic regression, neural networks (multi-layer perceptrons with two hidden layers), and SVM. These three algorithms are the most commonly used classifiers when analyzing real world data (e.g., they are commonly used by Spotify, Evernote or Booking.com; https://scikit-learn.org/stable/testimonials/testimonials.html). The first is a linear classifier, whereas the other two are standard nonlinear classification approaches in machine learning. For neural networks, we tested several different multi-layer perceptron architectures and the one with the two hidden layers (of size 10 each) and ReLu (restricted linear unit) output nonlinearities proved to yield the highest predictive accuracy among the others. To determine the learning rate of neural networks we tested various methodologies and decided on the constant learning rate of 0.001. For SVM, we used RBF (radial basis function) kernels and an L2-regularizer parameter with value C = 3.0. The regularization parameter makes the SVM less susceptible to outliers and improves its generalization. The kernel coefficient for the RBF kernel was scaled inversely proportional to the number of features and the feature matrix covariance.