RESUMO:
Digital phenotypes are behavioral patterns inferred from digital data, which have significantly contributed to eHealth research. Machine learning has proven effective in tracking these patterns; however, psychophysiological variables are still underutilized in identifying digital phenotypes. The objective of this study was to test the machine learning algorithms XGBoost, AdaBoost, RFC, and GaussianNB in classifying digital phenotypes of severe anxiety. Data from 360 volunteers, aged between 18 and 68 years of both sexes, were collected through self-report scales via the Neuropesquisa app, which simultaneously captured psychophysiological data, such as response time, accelerometer, and gyroscope readings. Scores from the DAAS-21 scale were used to identify individuals with severe anxiety. The results indicated that all four algorithms detected the digital phenotype with an accuracy greater than 75%. The RFC and GaussianNB classifiers achieved the highest accuracies, 81.48% and 80.56%, respectively, with f1 scores of 44.4% and 60.38%, and AUCs of 79% and 83% on the ROC curves. We conclude that the GaussianNB algorithm was the most effective in predicting the digital phenotype of severe anxiety.