Comparative Analysis of Machine Learning Algorithms for Anxiety and Depression Risk Prediction

The risk prediction of anxiety and depression can be considered one of the spheres of application of machine learning as the measurement of symptoms of anxiety and depression usually relies on self-reporting, interviews, or chance encounters. In this paper, comparative and literature review of classical machine learning algorithms and deep learning methods will be conducted to evaluate the use of machine learning approaches in risk prediction of anxiety and depression. These approaches include logistic regression, support vector machines, decision trees, random forest, gradient boosting, neural networks, convolutional neural networks, recurrent neural networks, transformers, and multimodal approaches. Unlike the traditional approach, which implies that the accuracy estimates obtained in the literature can be compared, in this review, several other aspects will be explored. The findings concerning risk prediction using electronic medical record information, social media textual data, speech information, and smartphone sensing information as well as other behavioral data indicate that tree-based algorithms and deep learning algorithms can be considered relevant for certain settings while having different performances in different studies depending on different factors. It means that the paper claims that algorithms should be chosen based on their applicability to the specific problem of risk prediction and deployment environment rather than just their accuracies. The framework for practical evaluation of the algorithms is developed, which takes into account such criteria as discrimination, calibration, robustness, fairness, interpretability, privacy, and clinical relevance.