[Paper Review] Detecting Individuals with Depressive Disorder fromPersonal Google Search and YouTube History Logs

Boyu Zhang, Anis Zaman+4 more|arXiv (Cornell University)|Oct 28, 2020

Mental Health via Writing20 references4 citations

TL;DR

This study proposes a personalized depression detection framework using longitudinal Google Search and YouTube history logs, modeling user behavior with a multidimensional mutual-exciting point process (Hawkes process) and distributional phrase embeddings to capture semantic and temporal dynamics. The best model achieved an F1 score of 0.77±0.04 and AUC ROC of 0.81±0.02 in classifying individuals with moderate to severe depressive symptoms from 212 participants.

ABSTRACT

Depressive disorder is one of the most prevalent mental illnesses among the global population. However, traditional screening methods require exacting in-person interviews and may fail to provide immediate interventions. In this work, we leverage ubiquitous personal longitudinal Google Search and YouTube engagement logs to detect individuals with depressive disorder. We collected Google Search and YouTube history data and clinical depression evaluation results from $212$ participants ($99$ of them suffered from moderate to severe depressions). We then propose a personalized framework for classifying individuals with and without depression symptoms based on mutual-exciting point process that captures both the temporal and semantic aspects of online activities. Our best model achieved an average F1 score of $0.77 \pm 0.04$ and an AUC ROC of $0.81 \pm 0.02$.

Motivation & Objective

• To develop a non-invasive, personalized method for detecting depressive disorder using private online behavioral data.
• To overcome limitations of public social media data, which suffer from self-censorship and lack of individual-level tracking.
• To leverage longitudinal search and YouTube engagement logs to model temporal and semantic patterns associated with depression.
• To improve upon rule-based temporal features by using a stochastic, semantic-aware model that captures interactivity between online activities.
• To provide a scalable, privacy-compliant system that supports early detection and clinical decision support.

Proposed method

• Collected personal Google Search and YouTube history logs and PHQ-9 depression survey responses from 212 participants, including 99 with moderate to severe depression.
• Used distributional phrase embeddings to extract semantic representations from search queries and video titles, enabling semantic modeling of online activities.
• Applied a multidimensional Hawkes process to model the mutual-exciting dynamics between different implicit topics in online behavior, capturing both temporal and semantic interdependencies.
• Trained the Hawkes process on truncated time windows (D = 2 weeks to 12 months) to identify optimal behavioral patterns preceding clinical assessment.
• Extracted two key features: baseline intensity vector (μ) and topic weight ratio (φ), which quantifies incoming vs. outgoing influence in the topic interaction graph.
• Classified depression status using a linear SVM with L2 regularization, trained on μ and φ features separately, with hyperparameters tuned via 5-fold cross-validation.

Experimental results

Research questions

• RQ1 Can longitudinal personal search and YouTube history logs reliably predict depressive disorder symptoms in individuals?
• RQ2 How do semantic and temporal patterns in online behavior differ between individuals with and without moderate to severe depression?
• RQ3 To what extent can a multidimensional Hawkes process model the mutual-excitation between distinct online activity topics to improve classification?
• RQ4 Does incorporating semantic embeddings from distributional phrase models improve detection performance over traditional keyword or rule-based features?
• RQ5 What is the optimal time window (D) for data truncation that maximizes predictive performance for depression detection?

Key findings

• The model using the topic weight ratio feature (φ) achieved the highest average weighted F1 score of 0.77±0.04, outperforming the baseline intensity feature (μ).
• The φ-based model achieved an AUC ROC of 0.81±0.02, indicating strong discriminative performance in distinguishing individuals with moderate to severe depression.
• The optimal hyperparameters across all configurations were σ=0.01, C=1, K=10, and D=6 months, which maximized classification performance.
• The φ feature, which captures the balance of incoming versus outgoing influence in topic interaction networks, proved more informative than raw intensity patterns.
• The study demonstrated that semantic-aware, stochastic modeling of online behavior via Hawkes processes significantly improves depression detection over rule-based or non-semantic approaches.
• Despite strong performance, the model's generalizability remains limited to college students, and ethical and privacy concerns around longitudinal data use remain significant barriers to real-world deployment.