Fig. 4: Side-by-side comparison of a psychiatrist annotation of risk factors for violence or threats compared to token-level SHAP highlights correlated to the prediction from our fine-tuned document classification model. | npj Mental Health Research

Fig. 4: Side-by-side comparison of a psychiatrist annotation of risk factors for violence or threats compared to token-level SHAP highlights correlated to the prediction from our fine-tuned document classification model.

From: Deep learning models can predict violence and threats against healthcare providers using clinical notes

Fig. 4

The automatically learned phrases strongly predictive of Code Gray or PSN by the model (right) are in certain areas similar to psychiatrist annotations (left), despite being trained on raw clinical note text rather than annotations. In other cases, psychiatrist-annotated risk factors, such as “bilateral acute on chronic SDH'', appear to not strongly influence prediction for the model. This example has been abbreviated and modified to prevent patient re-identification.

Back to article page