Suicidal Ideation Detection: A Review of Machine Learning Methods and Applications

Shaoxiong Ji; Shirui Pan; Xue Li; Erik Cambria; Guodong Long; Zi Huang

Source-linked AI summary

Suicidal Ideation Detection: A Review of Machine Learning Methods and Applications

Shaoxiong Ji, Shirui Pan, Xue Li, Erik Cambria, Guodong Long, Zi Huang

arXiv:1910.12611v4Oct 23, 2019cs.CYcs.AIcs.CLcs.LGcs.SI

TL;DR

Suicidal ideation detection is important for identifying risks before suicide attempts, but current AI research has limited benchmarks, incomplete intention understanding, and interpretability concerns. This survey synthesizes clinical and machine-learning approaches, reviews four data-source domains, summarizes tasks and datasets, and discusses limitations and future directions. It concludes that feature-based and deep-learning methods support automatic detection while privacy, ethics, data deficiency, and annotation bias remain important constraints.

Problem
Suicide prevention requires early detection of suicidal ideation, while existing AI research has limited benchmarks, may learn statistical clues without understanding intentions, and often lacks interpretability.
Method
The survey reviews clinical methods, feature engineering, deep learning, and applications using questionnaires, EHR data, suicide notes, and online user content.
Results
The survey summarizes existing suicidal ideation detection methods, domain-specific applications, tasks, datasets, limitations, and future research directions.
Takeaways & Limitations
Feature-based and deep-learning approaches provide ways to detect suicidal ideation automatically from textual and multimodal data, while interpretability and intention understanding remain research directions.
Takeaways & Limitations
Current research is constrained by insufficient annotated data and annotation bias, including limited fine-grained labels and potentially biased manual labeling.

Abstract

from arXiv · show

Suicide is a critical issue in modern society. Early detection and prevention of suicide attempts should be addressed to save people's life. Current suicidal ideation detection methods include clinical methods based on the interaction between social workers or experts and the targeted individuals and machine learning techniques with feature engineering or deep learning for automatic detection based on online social contents. This paper is the first survey that comprehensively introduces and discusses the methods from these categories. Domain-specific applications of suicidal ideation detection are reviewed according to their data sources, i.e., questionnaires, electronic health records, suicide notes, and online user content. Several specific tasks and datasets are introduced and summarized to facilitate further research. Finally, we summarize the limitations of current work and provide an outlook of further research directions.

I. INTRODUCTION

Suicidal ideation detection addresses a critical prevention problem by identifying suicidal thoughts or intentions from clinical interactions, structured data, or online content. The survey reviews machine-learning methods and applications while highlighting limited benchmarks, weak intention understanding, interpretability concerns, privacy, and ethics.

Suicidal ideation detection determines whether a person has suicidal thoughts from tabular data or textual content.
Early detection is presented as an important way to identify risks and support intervention before suicide attempts.
Existing approaches include clinical and psychological studies, questionnaire classification, feature engineering, sentiment analysis, and deep learning on social content.
Privacy and ethical concerns include algorithmic bias, predicting the timing of suicide acts, and the consequences of false-positive and false-negative predictions.
AI-based research faces limited benchmarks, models that learn statistical clues without understanding intentions, and insufficient interpretability.
The survey comprehensively reviews methods and applications across questionnaires, EHR data, suicide notes, and online social content, then summarizes tasks, datasets, limitations, and future directions.

II. METHODS AND CATEGORIZATION

Suicide ideation detection spans clinical interaction, automatic analysis of user-generated content, and a broader categorization of methods and application domains. Online content analysis examines linguistic information and social engagement as potential detection resources.

Automatic detection mainly analyzes user-generated text using machine-learning techniques.
Traditional detection relies on self-reports, face-to-face interviews, and clinical measurements such as speech, heart rate, fMRI, and EEG signals.
Figure 1 organizes suicidal ideation detection by method categories on the left and domain categories on the right, with arrows and solid points marking subcategories.
Exploratory analysis of user-generated content examines language use and linguistic clues through lexicon filtering, statistical linguistic features, and topic modeling.
Researchers manually build suicide-related keyword and phrase lexicons, including terms such as “kill,” “suicide,” “feel alone,” “depressed,” and “cutting myself.”

B. Feature Engineering

Feature-engineering methods transform tabular or unstructured textual information into inputs for conventional classifiers. The reviewed features include lexical, statistical, syntactic, psychological, contextual, affective, and topic-based representations.

Text-based suicide classification determines whether candidates express suicidal ideation through their posts using machine learning and NLP.
Tabular inputs include questionnaire responses and structured statistics that can serve directly as classification or regression features.
Unstructured-text features include N-grams, knowledge-based, syntactic, context, and class-specific features.
Researchers combine statistical, syntactic, LIWC, word-embedding, and topic features with traditional supervised classifiers.
Feature-engineering models include support vector machines, artificial neural networks, and conditional random fields.
Figure 2 illustrates methods that use feature engineering to construct model inputs for suicidal ideation detection.

3) Affective Characteristics:

Affective characteristics provide important cues for understanding suicidal ideation, while deep learning automates text-feature learning and supports detection across several neural architectures. Recent approaches also combine neural models with aggregation, attention, topic modeling, external features, or knowledge bases.

Affective Characteristics: Emotion categories and fine-grained sentiment features are used to analyze affective characteristics in suicide notes.Reported categories include anger, sorrow, hopefulness, fear, pride, abuse, forgiveness, factual aspects, and sentimental aspects.
Deep Learning: Deep learning learns text features automatically without sophisticated feature engineering for suicidal ideation detection.Some approaches instead feed psycholinguistic features and word occurrence into multilayer perceptrons.
Deep Learning: Recent models combine CNNs, LSTMs, bidirectional sequence encoding, self-attention, recurrent networks, and topic modeling for suicidal ideation detection.These designs target private-chat detection, informative subsequence selection, and joint encoding of text and risk indicators.
Deep Learning: The 2019 CLPsych Shared Task included CNN, LSTM, GRU-based RNN, NeuNetS, hierarchically attentive RNN, and BERT architectures.Studies also evaluated the effect of pretraining across different models.
Hybrid Methods: Hybrid methods combine limited feature engineering or external tabular features with representation learning for suicide-related classification.Examples include behavioral and suicide-language models and D-CNN models for depressed suicide attempters.
Summary: Deep neural networks with automatic feature learning boosted predictive performance, but some methods may learn statistical cues without commonsense.Knowledge bases and a suicide ontology were incorporated for knowledge-aware suicide risk assessment.

III. APPLICATIONS ON DOMAINS

Suicidal ideation detection applications draw on questionnaires, clinical records, and other person-level data, with methods adapted to the characteristics of each source. The reviewed domains include standardized assessment, electronic health records, and machine-learning frameworks for suicide-risk prediction.

Applications on Domains: Applications of suicidal ideation detection are organized by data source, including questionnaires, electronic health records, suicide notes, and online user content.The survey uses data-source categories to structure domain-specific applications.
Questionnaires: Questionnaire-based assessment uses standardized criteria and scales such as DSM-IV, ICD-10, IPDE-SQ, BIS-11, and SRRS.These tools support self-measurement or clinician-patient interviews and comparisons between suicide attempters and non-attempters.
Electronic Health Records: Electronic health records contain demographic and diagnosis-related histories, but their sparsity, variable-length clinical series, and heterogeneity complicate modeling.These characteristics remain challenges for suicide-attempt prediction.
Electronic Health Records: EHR-based studies use feature extraction, risk classifiers, calibration, preprocessing, and machine-learning models for short- and medium-term suicide-risk assessment.Reported models include random forests, gradient boosting machines, and deep neural networks.

C. Suicide Notes

Suicide notes provide text and multimedia material for studying suicidal ideation through content, sentiment, emotion, and linguistic analysis. As online blogs and social platforms expand, related work also examines user-generated posts and behavioral signals for earlier risk identification.

Suicide Notes: Suicide notes appear in letters, online blogs, audio, and video, providing material for NLP-based analysis.Prior work examines their content, sentiment, and emotional characteristics.
Suicide Notes: Researchers apply content analysis, sentiment analysis, emotion detection, and hybrid machine-learning or rule-based methods to suicide notes.The i2b2 task is used for fine-grained emotion and sentiment classification.
Online Blogs: Online blogs can be monitored to identify at-risk bloggers and distinguished suicide notes using linguistic and sentiment features.Examples include MySpace monitoring and comparisons with a non-suicide-note corpus.
Online User Content: Social websites provide anonymous spaces where people may post suicidal thoughts, creating data for language-pattern analysis and early warnings.The survey discusses Facebook, Twitter, Reddit, and MySpace as sources of user-generated content.
Twitter: Twitter studies use public tweets, TF-IDF features, logistic regression, SVM, and feature engineering to detect suicidal content.Psychology-based analysis has also examined conversations suggesting suicidal tendencies.
Reddit: Reddit research includes SuicideWatch analyses of suicidal intention and transitions from mental-health issues to suicidality.The platform supports topic-specific discussion research on mental health and suicide risk.
Other Platforms: Research also covers Chinese microblogs, MySpace adolescent content, and user- or post-centric behavioral analysis on other platforms.These studies use profile, linguistic, content, and behavioral features for suicide-related assessment.

E. Summary

The survey summarizes suicidal ideation detection across four application domains and contrasts their intervention potential and data characteristics. Online user content is presented as especially promising for early warning, while suicide notes are limited for immediate prevention and wearable health data is identified as a future source.

Summary: The four main application domains are questionnaires, electronic health records, suicide notes, and online user content.The survey organizes studies by intervention category, data source, and method.
Summary: Questionnaires and EHRs rely on self-report measurement or patient-clinician interactions involving social workers or mental-health professionals.Their use therefore depends substantially on human assessment settings.
Summary: Suicide notes have limited immediate-prevention value because many attempters act shortly after writing them, but they support content analysis and study of suicide factors.The limitation concerns timing rather than their usefulness as research material.
Summary: Online user content is described as one of the most promising routes for early warning and suicide prevention when combined with machine-learning techniques.The survey expects user-generated content to become more important as digital technology develops.
Future Directions: Wearable-device health data is identified as a likely future source for suicide-risk monitoring.The survey presents this as a future research direction.
Data Sources: Examples of domain data include questionnaire items, patient records, suicide-note text, tweets, and corresponding comments.These examples illustrate the range of inputs used for suicidal ideation detection.

IV. TASKS AND DATASETS

The survey organizes suicidal ideation detection into text classification, multi-aspect risk assessment, interpretable reasoning, and user-level detection tasks, while identifying benchmarks and datasets as research needs.

IV. TASKS AND DATASETS: The survey calls for benchmarks and datasets covering specific suicidal ideation and suicide-related tasks, including tasks that currently lack evaluation resources.These tasks include reasoning suicidal messages, generating responses, and detecting suicide attempters on social graphs.
Suicide Text Classification: Suicide text classification ranges from binary identification of suicidal ideation to multi-class assessment of suicide risk.The supplied passage describes binary and fine-grained multi-class classification, while another passage lists four behavioral categories.
Risk Assessment: Multi-aspect risk assessment learns from labels such as mental pain, cognitive attribution, and suicidal-risk level.Mental pain is scaled from 0 to 7, while cognitive attribution captures frustration of interpersonal needs or no indicated attribution.
Reasoning Suicidal Messages: Reasoning suicidal messages aims to use interpretable methods to investigate suicidal factors beyond predictive feature sets and classification models.The survey links this task to understanding factors and the innermost being expressed in textual posts.
Suicide Attempter Detection: User-level detection aggregates individuals’ posts or social relationships to identify suicide attempters through multi-instance learning or graph node classification.The former uses a bag of posts, while the latter operates on a social graph built from user interactions.

4) Generating Response:

The survey connects response generation with intervention-oriented suicidal ideation detection and situates it alongside mental-disorder and self-harm detection tasks and social-content datasets.

Generating Response: Response generation is intended to provide immediate social care and relieve suicidal intention through thoughtful counseling replies.The survey notes that people with suicidal intentions often post their suffering at midnight and describes eight response strategies.
Related Mental-Health Tasks: Related tasks include depression detection, self-harm detection, stressor-event detection, depression knowledge graphs, and depression–anxiety correlation analysis.The survey presents these as tasks associated with mental health issues and self-harm risks.
Online Data Sources: The reviewed online-content sources include Reddit, Twitter, and the ReachOut peer-support forum.The sources differ in platform structure, post characteristics, annotation procedures, and available datasets.
Reddit: Reddit datasets include 3,549 suicidal-ideation posts, 11,129 users with 1,556,194 posts, and a collection of 508,398 posts with 785 manually annotated posts.These figures come from distinct datasets and collection efforts described in the survey.

2) EHR:

The survey reviews electronic health records as a suicidal-behavior data source and emphasizes broader limitations involving data scarcity, annotation bias, imbalance, and weak intention understanding.

EHR: EHR data combine demographic information, admissions, diagnostic reports, and physician notes for suicide-risk research.One California collection contains 522,056 anonymous EHR records from adolescents, but is not publicly accessible.
Limitations: The survey identifies data deficiency as the most critical limitation, with too few annotated instances for further research.Fine-grained suicide-risk labels are limited, and multi-aspect and socially relational data are unavailable.
Limitations: Annotation bias arises because little evidence confirms suicide actions as ground truth and manual rules or crowdsourcing can bias labels.The survey also reports limited expert-labeled instances and uncertainty in demographic mortality data.
Limitations: Suicidal-intention posts form a tiny fraction of social content, but most studies construct approximately balanced positive and negative datasets.This differs from the imbalanced distribution found in real-world social data.
Limitations: Current statistical methods learn predictive clues but do not adequately reason about suicidal risk factors or the psychology behind suicide attempts.The survey contrasts feature selection and complex neural architectures with incorporating psychological understanding.

B. Future Work

Future work proposes richer representation learning, methods for imbalanced and temporal data, commonsense reasoning, and interpretable models to improve suicidal ideation detection.

Emerging Learning Techniques: Attention mechanisms and graph neural networks are proposed for suicide-text representation learning.The survey also mentions transfer learning, adversarial training, reinforcement learning, and adversarial data augmentation.
Emerging Learning Techniques: Few-shot learning could address the long-tail distribution of suicidal-ideation posts in real-world social networks.Its proposed use is training from a few labeled suicidal posts within a much larger social corpus.
Emerging Learning Techniques: Commonsense reasoning with external suicide-related knowledge bases is proposed to improve understanding of suicidal intention.The survey links this direction to multiple factors correlated with suicide, including mental health and economic conditions.
Emerging Learning Techniques: Because deep-learning predictors can be black boxes, future work should develop interpretable models for understanding suicidal intentions and reliable prediction.The proposed direction addresses both interpretability and understanding of intentions.
Emerging Learning Techniques: Temporal modeling of users’ posts is proposed to monitor changing mental states and detect early suicidal-ideation signals.The survey describes stages including stress, depression, suicidal thoughts, and suicidal plans.

4) Proactive Conversational Intervention:

Proactive intervention extends suicidal ideation detection toward timely conversations that may help relieve suicidal thoughts. The survey identifies automatic response generation and reinforcement learning as promising approaches for conversational suicide intervention.

Proactive Suicide Prevention Online combines suicidal identification with crisis management to support intervention beyond detection.
Automatic response generation is proposed as a technical solution for timely intervention with people expressing suicidal thoughts.
Natural language generation can produce counseling responses intended to comfort depression or suicidal ideation.
Reinforcement learning can model conversational intervention in which responses are selected to maximize rewards based on attempters’ reactions.The described policy-gradient approach treats the initial suicide message as the starting state and observed reactions as rewards.
The survey identifies proactive conversational intervention as a future direction alongside emerging learning techniques, interpretable intention understanding, and temporal detection.