Comprehensive Summary
This study looks at how well deep learning can detect and predict suicidal behavior, specifically intentional self-harm (ISH), from clinical notes. Clinical notes from 11,298 patients aged 18-90 were used in this study. These notes were run through two bag-of-words models, Naïve Bayes (NB) and Random Forest (RF), as well as two Convolutional Neural Network (CNN) models which were used for deep learning, one with randomly initialized embeddings (CNNr) and one with Word2Vec-initilized embedding (CNNw). In the results, the CNN models out-performed the bag-of-words models in detection of concurrent ISH, with both CNN models achieving AUC scores of 0.99 compared to the bag-of-words models which achieved an AUC of 0.95. In terms of predicting future ISH events, both CNN models did better than the NB and RF models, achieving an AUC of 0.82 for the CNNr model and an AUC of 0.81 for the CNNw model while the RF model had an AUC of 0.79 and the NB achieved an AUC of 0.79. The authors were able to successfully use clinical notes along with well-defined ICD codes to train deep learning models to detect ISH events. They also noted how this study echoes previous findings showing that the results are reproducible.
Outcomes and Implications
This research is important as developing an algorithm that can detect and predict ISH events can improve efficiency, as existing methods for suicide risk assessment can be costly, time-consuming, and limited in their predictive accuracy. Also, this research paves the way for similar tools being developed for other health outcomes such as drug overdoses. This research is clinically relevant because it shows that deep learning models can flag patients at risk of harming themselves, allowing for early intervention and subsequent prevention. The authors note that while the approach is promising, similar research needs to be done with larger sets of data before clinical implementation.