Table 1.
Annotation agreement between two clinical annotators.
Annotations were retained as the labelled dataset for predictions if the experts annotators agree on the classifications of their mentions.
Fig 1.
The GATE NLP based ADEPt pipeline comprising four rule-based processing components.
The pipeline takes as input EHR clinical notes documents and a dictionary containing all annotation-related terms. The pipeline sequentially applies the four components accumulating new annotations for the target annotation (ADE). The output of the pipeline is a single ADE annotation with six features (ADE type, Experiencer, Negation, Temporality, Categorical_Value, Refinedment_Rule, ADE_status and clause).
Table 2.
Enrichment of the ConText algorithm trigger terms.
Fig 2.
The retention rules pattern used in the ADEPt pipeline.
Fig 3.
The removal rules pattern used in the GATE NLP based ADEPt pipeline.
Table 3.
Incremental results of akathisia, galactorrhea, nausea and myocarditis.
Fig 4.
Receiver operating characteristic curves representing the performance of the ADEPt pipeline in identifying akathisia, nausea galactorrhea and myocarditis ADEs from free text.
The increments in each graph correspond to 1) our previous work [30], 2) using paragraph boundaries, 3) using clause-boundaries, 4) using unrefined (off-the-shelf) ConText algorithm, 5) adding domain-specific vocabulary to ConText and 6) final refined ConText algorithm.
Table 4.
Results showing the performance of the ADEPt pipeline in identifying a selection of rare to common ADEs related to antipsychotics and antidepressants drugs.