Stress classification using machine learning methods based on biomedical time series from a wearable device

U ovom radu korišteni su algoritmi strojnog učenja za klasifikaciju biomedicinskih signala u neurološka stanja stresa. Korišten je javno dostupan skup podataka Non-EEG Dataset for Assessment of Neurological Status [1]. Opisan je način na koji su podaci prikupljani i kako su transformirani u željeni oblik. Objašnjene su ideje i standardne metode strojnog učenja Proučavan je manji skup algoritama strojnog učenja i to preciznije, klasifikacije: stabla odluke, naivni Bayesov klasifikator i neuronske mreže. Za učenje modela korištena je platforma MATLAB. Tamo su podaci uvedeni, transformirani te su iz njih izlučene značajke. Biomedicinski signali koji su se pokazali najbitnijima za klasifikaciju signala bili su temperatura i akceleracija u z osi te njihove značajke poput minimuma i standardne devijacije (izvedenih iz temperature) te vrhunca amplitude na spektru snage i Akaikeovog kriterija informacije (izvedenih iz akceleracije u z osi). Modeli su naučeni, provjereni te konačno testirani nad podacima i njihova uspješnost je uspoređena. Točnost od 87,5% na testnom skupu dala je unaprijedna neuronska mreža s tri skrivena sloja, što je ujedno bio i najbolji rezultat.In this paper, machine learning algorithms were used for the classification of biomedical signals into neurological conditions of stress. The publicly available dataset used was the Non-EEG Dataset for Assessment of Neurological Status [1]. The method of data collection and its transformation into the desired format was described. The ideas and standard methods of machine learning were explained. A small set of machine learning algorithms was studied, specifically decision trees, naive Bayes classifier, and neural networks. The MATLAB platform was used for training the models. The data was imported, transformed, and features were extracted from it. The biomedical signals that proved to be crucial for classification were temperature and acceleration on the z-axis and their features such as minimum value and standard deviation (derived from temperature) and peak amplitude on the power spectrum and Akaike's information criterion (derived from acceleration on the z-axis). The models were generated, verified, and finally tested on the data, and their performance was compared. The highest accuracy on the test of 87.5% was achieved by a feedforward neural network with three hidden layers