Development of a two-stage depression symptom detection model: application of neural networks to twitter data

Abstract

This study aims to help in the area of depression screening in the Philippine setting, focusing on the detection of depression symptoms through language use and behavior in social media to help improve the accuracy of symptom tracking. A two-stage detection model is proposed, wherein the first stage deals with the detection if depression symptoms exist and the second stage focuses on the detection of depression symptom category or type for English and Filipino language. A baseline data set with 14 depression categories consisting of 86,163 tweets was used as input to various machine learning algorithms together with Twitter user behaviors, linguistic features, and psychological behaviors. The two-stage detection models used Bidirectional Long-Short Term Memory type of Artificial Neural Network with dropout nodes. The first stage, with a binary output classifier, can detect tweets with “Depression Symptom” or “No Symptom” categories with an accuracy of 0.91 and F1-score of 0.90. The second stage classifier has 6 depression symptom categories, namely “Mind and Sleep,” “Appetite,” “Substance use,” “Suicidal tendencies,” “Pain,” and “Emotion” symptoms that has an accuracy of 0.83 and F1-score of 0.81. The two-stage algorithm can be used to complement mental health support provided by clinicians and in public health interventions to serve as high-level assessment tool. Limitations on misclassifications, negation, and data imbalance and biases can be addressed in future studies.