Depressive Emotion Detection and Behavior Analysis of Men Who Have Sex With Men via Social Media

Abstract

Background: A large amount of evidence has indicated an association between depression and HIV risk among men who have sex with men (MSM), but traditional questionnaire-based methods are limited in timely monitoring depressive emotions with large sample sizes. With the development of social media and machine learning techniques, MSM depression can be well monitored in an online and easy-to-use manner. Thereby, we adopt a machine learning algorithm for MSM depressive emotion detection and behavior analysis with online social networking data. Methods: A large-scale MSM data set including 664,335 users and over 12 million posts was collected from the most popular MSM-oriented geosocial networking mobile application named Blued. Also, a non-MSM Benchmark data set from Twitter was used. After data preprocessing and feature extraction of these two data sets, a machine learning algorithm named XGBoost was adopted for detecting depressive emotions. Results: The algorithm shows good performance in the Blued and Twitter data sets. And three extracted features significantly affecting the depressive emotion detection were found, including depressive words, LDA topic words, and post-time distribution. On the one hand, the MSM with depressive emotions published posts with more depressive words, negative words and positive words than the MSM without depressive emotions. On the other hand, in comparison with the non-MSM with depressive emotions, the MSM with depressive emotions showed more significant depressive symptoms, such as insomnia, depressive mood, and suicidal thoughts. Conclusions: The online MSM depressive emotion detection using machine learning can provide a proper and easy-to-use way in real-world applications, which help identify highrisk individuals at the early stage of depression for further diagnosis.

Machine Learning Applications in Mental Health and Substance Use Research Among the LGBTQ2S+ Population: Scoping Review

“…Most of the studies were carried out in the United States (7/11, 63%) [ 36 , 38 , 39 , 41 - 43 , 45 ]. Among the target population categories that were not mutually exclusive, sexual minority men (gay, men who have sex with men, bisexual) were the most commonly studied (5/11, 45%) subgroups [ 37 , 40 , 42 - 44 ], followed by transgender (3/11, 27%) [ 39 , 45 , 46 ] and LGBTQ+ (3/11, 27%) [ 36 , 38 , 41 ] people at large, whereas sexual minority women (lesbian and bisexual) (2/11, 18%) [ 43 , 45 ] were the least commonly represented populations. None of the studies included two-spirit persons as their target population ( Table 1 ).…”

“…The most commonly used class of ML methods was supervised (9/11, 82%) [ 37 - 39 , 41 - 46 ], followed by unsupervised (3/11, 27%) [ 36 , 37 , 40 ] and deep learning (1/11, 9%; Table 1 ) [ 41 ]. The most frequently used ML algorithms were latent Dirichlet allocation (3/11, 27%) and boosting (3/11, 27%), followed by random forest, support vector machines, penalized regression (ie, least absolute shrinkage and selection operator, elastic net regularized regression, and ridge regression), classification tree, logistic regression, naive Bayes, multilayered perceptron, and convolutional neural network ( Table 1 ).…”

“…Approximately two-thirds (7/11, 64%) of the studies [ 37 , 38 , 42 - 46 ] discussed their methods of feature selection, among which the median number of features used was 19. Most of the studies used cross-validation methods (7/11, 64%) [ 37 - 39 , 41 , 44 - 46 ], especially 10-fold cross-validation.…”

“…Approximately two-thirds (7/11, 64%) of the studies [ 37 , 38 , 42 - 46 ] discussed their methods of feature selection, among which the median number of features used was 19. Most of the studies used cross-validation methods (7/11, 64%) [ 37 - 39 , 41 , 44 - 46 ], especially 10-fold cross-validation. Furthermore, 18% (2/11) of the articles used the hold out method [ 39 , 41 ], 18% (2/11) used external validation [ 37 , 41 ], and 36% (4/11) articles [ 36 , 40 , 42 , 43 ] did not report how they validated their method.…”

Psychiatric Advance Directives and Artificial Intelligence: A Conceptual Framework for Theoretical and Ethical Principles

“…The emotions, the energy level, or the presence of symptoms with their perceived intensity (ruminations, hallucinations, and suicidal ideation) can be analyzed. These data can provide useful indicators to identify the increased symptomatology (crisis, manic episode) of many pathologies (bipolar disorder, schizophrenia, major depressive episode, substance abuse) such as logorrhea, increased communicability or reduced social contact, increased behavioral activation, agitation, or psychomotor deceleration (23)(24)(25)(26)(27)(28)(29)(30). For instance, CDSS enhanced with AI could make compulsory admissions more efficient to provide appropriate psychiatric care (4).…”