ABSTRACT
Chat-based counseling hotlines emerged as a promising low-threshold intervention for youth mental health. However, despite the resulting availability of large text corpora, little work has investigated Natural Language Processing (NLP) applications within this setting. Therefore, this preregistered approach (OSF: XA4PN) utilizes a sample of approximately 19,000 children and young adults that received a chat consultation from a 24/7 crisis service in Germany. Around 800,000 messages were used to predict whether chatters would contact the service again, as this would allow the provision of or redirection to additional treatment. We trained an XGBoost Classifier on the words of the anonymized conversations, using repeated cross-validation and bayesian optimization for hyperparameter search. The best model was able to achieve an AUROC score of 0.68 (p < 0.01) on the previously unseen 3942 newest consultations. A shapely-based explainability approach revealed that words indicating younger age or female gender and terms related to self-harm and suicidal thoughts were associated with a higher chance of recontacting. We conclude that NLP-based predictions of recurrent contact are a promising path toward personalized care at chat hotlines.
ABSTRACT
With the need for psychological help long exceeding the supply, finding ways of scaling, and better allocating mental health support is a necessity. This paper contributes by investigating how to best predict intervention dropout and failure to allow for a need-based adaptation of treatment. We systematically compare the predictive power of different text representation methods (metadata, TF-IDF, sentiment and topic analysis, and word embeddings) in combination with supplementary numerical inputs (socio-demographic, evaluation, and closed-question data). Additionally, we address the research gap of which ML model types - ranging from linear to sophisticated deep learning models - are best suited for different features and outcome variables. To this end, we analyze nearly 16.000 open-text answers from 849 German-speaking users in a Digital Mental Health Intervention (DMHI) for stress. Our research proves that - contrary to previous findings - there is great promise in using neural network approaches on DMHI text data. We propose a task-specific LSTM-based model architecture to tackle the challenge of long input sequences and thereby demonstrate the potential of word embeddings (AUC scores of up to 0.7) for predictions in DMHIs. Despite the relatively small data set, sequential deep learning models, on average, outperform simpler features such as metadata and bag-of-words approaches when predicting dropout. The conclusion is that user-generated text of the first two sessions carries predictive power regarding patients' dropout and intervention failure risk. Furthermore, the match between the sophistication of features and models needs to be closely considered to optimize results, and additional non-text features increase prediction results. Supplementary Information: The online version contains supplementary material available at 10.1007/s41666-023-00148-z.