Identifying and Predicting Intentional Self-Harm in Electronic Health Record Clinical Notes: Deep Learning Approach

Area of Science:

• Computational psychiatry
• Artificial intelligence in healthcare
• Clinical informatics

Background:

• Suicide and intentional self-harm represent a significant global public health challenge.
• Machine learning (ML) and deep learning (DL) are increasingly explored for suicide risk assessment.
• Advancements in computing power are driving the adoption of DL in healthcare applications.

Purpose of the Study:

• To utilize deep neural networks (DNNs) to analyze clinical notes for improved identification of intentional self-harm.
• To predict future self-harm events using information extracted from electronic health records (EHRs).

Main Methods:

• Extracted clinical text notes from EHRs of patients with intentional self-harm codes and matched controls.
• Developed and tested traditional ML models, a convolutional neural network (CNN), and a long short-term memory (LSTM) model.
• Evaluated predictive performance using data preceding self-harm events and assessed the impact of Word2vec (W2V) pretraining.

Main Results:

• The CNN achieved near-perfect performance (AUC=0.999, F1=0.985) in identifying intentional self-harm from concurrent clinical notes.
• For predicting future self-harm, the CNN demonstrated strong performance (AUC=0.882, F1=0.769).
• Word2vec pretraining reduced DNN training time but did not enhance model performance.

Conclusions:

• DNNs applied to clinical text show high efficacy for phenotyping and surveillance of intentional self-harm.
• While predictive performance was modest, DNNs using clinical text are competitive with models using structured EHR data.
• These findings support the potential of AI-driven analysis of clinical notes for suicide prevention strategies.