4 - (SYM 66) Scaling Support: Ai-driven Innovations for 988 and Crisis Helplines

Suicide risk assessment is a critical component of crisis counseling. National standards require continuous quality improvement of 988 and crisis counseling services, and the need for accurate and thorough risk assessment in community treatment is integral to client well-being. Traditional quality improvement efforts rely on human evaluation of sessions, a process that is challenging to scale. Training providers to conduct a thorough risk assessment is severely lacking due to a dearth of resources. Advances in machine learning (ML) and artificial intelligence (AI) offer the potential to automate the detection of suicide risk assessment, thereby enhancing the scalability and efficiency of training and quality improvement initiatives. 

In collaboration with ProtoCall Services, we conducted a two phase research study to develop and evaluate machine learning algorithms to support the automatic identification of risk assessments that occur during crisis interactions (phase one), and to explore the impact of this feedback on caller outcomes (phase two). Through the development of these ML models, we collaborated with content experts in the field of suicideology (Drs. David Jobes and Samantha Chalker) to create a suicide prevention training, which includes automatically generated feedback on key risk assessment skills.  

To create and train our risk assessment classification AI models, a coding team manually labeled 193,257 statements across 476 crisis counseling calls, identifying core elements of risk assessment (based on the Crisis Chat Abstraction Form; Lake et al, 2022). This labeled dataset was used to fine-tune a transformer-based ML model, with separate training, validation, and test datasets employed to assess model performance.

For detecting any risk assessment, the model achieved 98% agreement with human ratings, relative to human interrater agreement. At the call level, the average F1 score—a harmonic mean of precision and recall—was 0.86, while at the statement level, it was 0.66. Variations in F1 scores across specific labels were often due to low base rates for certain risk assessment components. Key components of risk assessments and empathic communication were selected as modules for training.

The findings indicate that ML models can reliably detect suicide risk assessment, presenting a viable solution to scale quality improvement and training efforts in this domain.