Early Detection of Mental Health Disorders by Social Media Monitoring: The First Five Years of the eRisk Project

Foreword
Preface
Contents
Early Risk Prediction of Mental Health Disorders
	1 Introduction
	2 Book Content
The eRisk Initiative
The Challenge of Early Risk Prediction on the Internet
	1 Introduction
	2 Text as a Source of Risk Evidence
	3 Datasets
		3.1 Positive Group
		3.2 Control Group
		3.3 Extraction of the User's Submissions
		3.4 eRisk Collections
	4 Evaluation Design
		4.1 Decision-Based Metrics
		4.2 Ranking-based Evaluation
	5 eRisk Task on Automatically Filling a Depression Questionnaire
		5.1 Evaluation Metrics
	6 Conclusions
	References
A Survey of the First Five Years of eRisk: Findings and Conclusions
	1 Introduction
	2 Early Risk Detection Systems Design
		2.1 Feature Extraction
		2.2 Assessment Technologies
		2.3 Decision Policy
	3 Measuring the Severity of the Signs of Depression
	4 Findings and Lessons
		4.1 A Challenging Trade-Off
		4.2 Experimental Design and Evaluation Framework
		4.3 Signals Crossover
	5 Conclusions
	References
The Best of eRisk
From Bag-of-Words to Transformers: A Deep Dive into the Participation in the eRisk Early Risk Detection of Depression Tasks with Classical and New Approaches
	1 Introduction
	2 eRisk 2017/2018 Task for Early Detection of Depression
		2.1 Overview
		2.2 Datasets
	3 Evaluation Criteria and Concerns About ERDE Score
	4 Features and Architectures Used for Depression Detection
		4.1 Hand-Crafted User Features
		4.2 Readability
		4.3 Word and Grammar Usage
		4.4 Metadata Feature Summary
		4.5 Linguistic Metadata
		4.6 Emotions and Sentiment Lexica
		4.7 Word Embeddings
		4.8 Neural Network Architecture
		4.9 Transformers
	5 Chosen Models for the eRisk Tasks, Related Work  and Ongoing Work
		5.1 Bag-of-Words Ensemble - BCSGA - 2017/2018
		5.2 Paragraph Vector - BCSGB - 2017
		5.3 Bag-of-Words Ensemble - BCSGB - 2018
		5.4 LSTM with LSA Vectors - BCSGC - 2017
		5.5 CNN with GloVe Embeddings - BCSGC - 2018
		5.6 LSTM with Paragraph Vectors - BCSGD - 2017
		5.7 CNN with fastText Embeddings - BCSGD - 2018
		5.8 Late LSTM with Paragraph Vectors - BCSGE - 2017
		5.9 CNN and Bag-of-Words Metadata Ensemble - BCSGE - 2018
		5.10 CNN's and metadata features
		5.11 Setup for Transformer Experiments
	6 Results
		6.1 Results eRisk 2017
		6.2 Results eRisk 2018
		6.3 Related results
		6.4 Results of Ongoing Work
		6.5 Results with Transformers
	7 Discussion
	8 Conclusion
	References
Comparison of Machine Learning Models for Early Depression Detection from Users' Posts
	1 Introduction
	2 Related Work
	3 Information Modeling
		3.1 Feature-Based Representation
	4 Machine Learning Models
		4.1 Well-Established Machine Learning Models
		4.2 BERT-Based Model
	5 Experimental Framework
		5.1 Collections
	6 Results and Discussion
		6.1 Depression Detection
		6.2 Early Depression Detection
		6.3 Simplified Models
		6.4 Ablation Analysis
	7 Visualization of Early Detection
	8 Conclusion
	References
Quick and (Maybe Not So) Easy Detection of Anorexia in Social Media: To Explainability and Beyond
	1 Introduction
	2 Related Work
	3 System Overview
		3.1 Sub-models
		3.2 Ensemble Model
	4 Experimental Setup
		4.1 Sub-models Implementation
		4.2 Ensemble Classifiers
		4.3 Submitted Runs
	5 Shared Task Results
	6 Explainability Analysis
		6.1 Experiments
		6.2 Results and Discussion
		6.3 Further Analysis
	7 Conclusion
	References
Two Simple and Domain-independent Approaches for Early Detection of Anorexia
	1 Introduction
	2 Approaches
		2.1 Flexible Temporal Variation of Terms (FTVT)
		2.2 SS3 Text Classifier
	3 Participation and Results
		3.1 Early Detection of Anorexia—2018 Edition
		3.2 Early Detection of Anorexia—2019 Edition
	4 Conclusion and Future Work
	References
Early Risk Detection of Self-Harm Using BERT-Based Transformers
	1 Introduction
	2 Related Work
	3 Approach
	4 Experiments
		4.1 Provided Training Data
		4.2 Created Training Data
		4.3 Method
		4.4 Results
	5 Additional Experiments
		5.1 Provided Training Data (Anorexia and Depression)
		5.2 Created Training Data (Anorexia and Depression)
		5.3 Method (Anorexia and Depression)
		5.4 Results (Anorexia and Depression)
	6 Summary
	References
Detecting Traces of Self-harm on Reddit Through Emotional Patterns
	1 Introduction
	2 Related Work
	3 From Sentiments to Emotions
	4 Can Emotions Have Shades? A Sub-emotion Based Representation
		4.1 Generating Sub-emotions
		4.2 Analysis of the Novel Sub-emotions
		4.3 Converting Text to Sub-emotions Sequences
	5 Using BoSE to Identify Self-harm
		5.1 BoSE Definition
		5.2 BoSE at eRisk 2020
		5.3 BoSE for Whole Post's History
	6 Learning Sequential Information from Sub-emotions
		6.1 normal upper DeltaΔ-BoSE
	7 Deep Learning for Extracting Sequential Emotion Patterns
		7.1 Convolutional Neural Network and BoSE
		7.2 Recurrent Neural Network and BoSE
		7.3 Adding Attention to the Sub-emotions
	8 Conclusion
	References
On the Estimation of Depression Through Social Mining
	1 Introduction
	2 Background
	3 Measuring Depression
		3.1 Dataset Description
		3.2 Metrics
		3.3 Methods
		3.4 Results
	4 Discussion
	5 Conclusions
	References
Automatically Estimating the Severity of Multiple Symptoms Associated with Depression
	1 Introduction
	2 Task and Data
	3 Evaluation Metrics
	4 Related Work
	5 Approaching the Task as One of Authorship Attribution
		5.1 Topic Models
		5.2 Contextualizer
		5.3 Stylometry
	6 Results and Discussion
	7 Conclusion
	References
Beyond eRisk
Beyond Risk: Individual Mental Health Trajectories from Large-Scale Social Media Data
	1 Introduction
	2 Studying Behavioral Traces of Sleep and Emotions
		2.1 Sampling Approach
		2.2 Behavioral Analysis
		2.3 Emotional Analysis
	3 Detecting Cognitive Markers of Mental Health Disorders
		3.1 Constructing Cognitive Markers
		3.2 Sampling Approach
		3.3 Lexical Analysis of Cognitive Distortion Markers
	4 Conclusion, Discussion, and Future Research
	References
Explainability of Depression Detection  on Social Media: From Deep Learning Models to Psychological Interpretations and Multimodality
	1 Introduction
	2 Previous Work
	3 Datasets
	4 Models
		4.1 Features
		4.2 Hierarchical Attention Networks
		4.3 Baseline Classifiers
		4.4 Classification Results
	5 Explainability of Depression Detection
		5.1 Error Analysis
		5.2 Hidden Layer Analysis
		5.3 Ablation Experiments
	6 Multimodal Depression Detection
		6.1 multiRedditDep Dataset Collection
		6.2 Classification Experiments
		6.3 Image Analyses and Psychological Interpretations
	7 Conclusions and Future Directions
	References
The Future
The Future of eRisk
	1 eRisk so Far
	2 Future Outlines for Early Risk Prediction on the Internet
		2.1 Topic Coverage, Domain Expansion and New Metrics  for Early Risk
		2.2 A Continuous Improvement of the eRisk Cycle
		2.3 New Challenges: Estimating Risk from Standard Questionnaires
	3 Conclusions
	References