Search

In this work I tried to use different binary classifiers to analyse audio recordings of people coughing and classify them as affected or non affected by covid-19, as has been done during the last year by different research groups all over the world. I took the raw audio data from the CoSwara repository, from which after some manipulation I extracted the mel frequency cepstral coefficients and other features, and I used the logistic regression, the support vector machine and the multilayer perceptron as classifiers to perform my supervised learning analysis. In order to overcome the problem of the imbalance of the dataset (in the training set Positive/Negative ratio was around 1/12) I used the syntetic minority oversampling technique (SMOTE). I performed the hyperparameters tuning through cross validation on the training set and then I built the final model for logistic regression and multilayer perceptron, which I assessed using a test set composed of new observations. In both stages of the work, the models have been assessed using the area under the ROC curve., Incoming