Predicting recurrent chat contact in a psychological intervention for the youth using natural language processing

Area of Science:

• Digital Mental Health
• Computational Linguistics
• Adolescent Psychiatry

Background:

• Chat-based counseling hotlines offer accessible mental health support for youth.
• Large text datasets from these services present opportunities for Natural Language Processing (NLP) applications.
• Limited research exists on NLP within youth crisis chat services.

Purpose of the Study:

• To develop and evaluate an NLP model for predicting repeat contact among youth using a German crisis chat service.
• To identify linguistic features associated with recontact to inform personalized care strategies.

Main Methods:

• Utilized approximately 800,000 messages from 19,000 youth consultations with a 24/7 crisis service.
• Trained an XGBoost Classifier on anonymized chat conversation text.
• Employed repeated cross-validation and Bayesian optimization for model training and hyperparameter tuning.
• Applied a SHAPley Additive exPlanations (SHAP) approach for model interpretability.

Main Results:

• The best XGBoost model achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.68 (p < 0.01) in predicting recontact.
• Linguistic indicators of younger age, female gender, self-harm, and suicidal ideation were associated with a higher likelihood of recontact.
• The model demonstrated predictive capability on novel consultation data.

Conclusions:

• NLP-based prediction of recurrent contact is a feasible and promising approach for enhancing personalized care in youth mental health chat hotlines.
• Identifying at-risk individuals through linguistic analysis can facilitate timely interventions and treatment redirection.
• This study highlights the potential of computational methods to improve support delivery in low-threshold mental health services.