Using weak supervision and deep learning to classify clinical notes for identification of current suicidal ideation

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Mental Health Research

Background:

• Mental health conditions, including suicidal thoughts, are often documented in unstructured clinical notes rather than structured data.
• Accurate identification of suicidal ideation is crucial for timely intervention and patient safety.

Purpose of the Study:

• To evaluate weakly supervised machine learning methods for detecting "current" suicidal ideation from unstructured electronic health record (EHR) clinical notes.
• To compare the performance of various machine learning models, including a convolutional neural network (CNN), for this task.

Main Methods:

• A cohort of 600 patients at risk for suicidal ideation was identified.
• A rule-based natural language processing (NLP) approach was used to label 17,978 training and validation notes.
• Statistical models (logistic classifier, SVM, Naive Bayes) and a CNN were trained and evaluated on a 837-note manually-reviewed test set.

Main Results:

• The CNN model achieved the highest performance, with 94% accuracy and a F1-score of 0.82 for detecting "current" suicidal ideation.
• The algorithm identified 42 additional encounters and 9 patients with suicidal ideation missed by structured diagnosis codes.
• When applied to 5,000 notes, the algorithm flagged 23 cases (0.46%) of "current" suicidal ideation, with 87% confirmed by manual review.

Conclusions:

• Weakly supervised machine learning, particularly CNNs, can effectively detect "current" suicidal ideation in unstructured clinical notes.
• This approach can enhance the identification of at-risk individuals, potentially improving suicide prevention interventions.
• Automated screening of clinical notes holds promise for improving patient care and research in mental health.