A machine learning approach to identifying suicide risk among text-based crisis counseling encounters

Meghan Broadbent, Mattia Medina Grespan, Katherine Axford, Xinyao Zhang, Vivek Srikumar, Brent M. Kious, Zac E. ImelFrontiers in Psychiatry 2023 · DOI

Abstract

The neural model demonstrated greater sensitivity in the detection of client suicide risk. A manual assessment of errors and model performance reflected these same findings, detecting higher levels of risk in many of the false-positive encounters and lower levels of risk in many of the false negatives. NLP-based models can detect the suicide risk of text-based crisis encounters from the encounter's content.

Introduction

Text-based crisis hotlines have emerged as a vital resource for mental health support, particularly among younger demographics. However, the high volume of messages places significant pressure on counselors who must manage multiple complex conversations simultaneously. While clinical guidelines mandate rigorous risk assessments, the fast-paced nature of text-based communication and counselor burnout often lead to inconsistent documentation. There is a pressing need for automated Natural Language Processing (NLP) tools that can serve as a "second pair of eyes," flagging high-risk encounters in real-time to ensure critical interventions are not missed.

Methods

This study compared two machine learning approaches for classifying crisis encounters as "lower risk" or "higher risk." The primary dataset was sourced from SafeUT, a regional crisis application serving Utah, Idaho, and Nevada, comprising 5,992 evaluation encounters and 120,000 pre-training encounters.

The Neural Model (RoBERTa)

The researchers developed a model based on the "Robustly Optimized BERT" (RoBERTa) architecture. To tailor the model to the crisis counseling domain, it was pre-trained on 2.5 million messages from unlabeled encounters. A key innovation involved prepending "originator tokens" (e.g., [#T] for therapist, [#C] for client) to messages and processing them in consecutive pairs to capture conversational context.

Architecture of the neural model utilizing RoBERTa and a transformer encoder for suicide risk classification.

The architecture utilized a 6-layer transformer encoder to process message embeddings, which were then averaged into a single encounter embedding for final classification via a Multi-Layer Perceptron (MLP).

Baseline Model

The baseline consisted of a traditional TF-IDF (Term Frequency-Inverse Document Frequency) vectorization of unigrams and bigrams, followed by a logistic regression classifier.

Key Results

The neural model demonstrated superior performance across several critical metrics, particularly in its sensitivity to high-risk signals.

ROC curves comparing the performance of the neural model against the TF-IDF baseline.
  • Predictive Accuracy: The neural model achieved an Area Under the Curve (AUC) of 90.37%, outperforming the TF-IDF baseline (88.18%).
  • Sensitivity and Safety: Crucially for suicide prevention, the neural model reached a sensitivity of 62.02%, a massive improvement over the baseline's 39.52%. This resulted in a significantly lower False Negative Rate (37.98% vs. 60.47%), meaning fewer high-risk individuals were missed.
  • Human-Model Discrepancy: A manual review of "False Positives" (where the model predicted high risk but counselors did not) revealed that in 60.6% of these cases, independent clinical reviewers identified clear suicidality signals that the original counselor had likely overlooked.

Discussion and Limitations

The findings suggest that NLP models can effectively track dynamic risk throughout a conversation, identifying "spikes" in distress when specific plans are mentioned and "diffusion" when risk subsides. However, several limitations remain. The data is geographically limited to the SafeUT service area, which may introduce regional linguistic biases. Additionally, the lack of demographic data prevents an assessment of algorithmic fairness across different groups.

Future research should focus on using more objective outcomes, such as hospital records, as the "gold standard" for training, rather than relying solely on subjective counselor labels. The ultimate goal is the integration of these models into real-time decision-support systems to assist counselors during active crises.

Summary updated

Abstract

The neural model demonstrated greater sensitivity in the detection of client suicide risk. A manual assessment of errors and model performance reflected these same findings, detecting higher levels of risk in many of the false-positive encounters and lower levels of risk in many of the false negatives. NLP-based models can detect the suicide risk of text-based crisis encounters from the encounter's content.
BibTeX
@inproceedings{broadbent2023machine,
  author = {Meghan Broadbent and Mattia Medina Grespan and Katherine Axford and Xinyao Zhang and Vivek Srikumar and Brent M. Kious and Zac E. Imel},
  title = {{A machine learning approach to identifying suicide risk among text-based crisis counseling encounters}},
  booktitle = {Frontiers in Psychiatry},
  year = {2023},
  doi = {10.3389/fpsyt.2023.1110527},
}
Figure