Causality patterns and machine learning for the extraction of problem-action relations in discharge summaries

Graphical abstract This figure is our system architecture for the Problem-Action relation extraction. We detect from clinical texts based on conditional random fields (CRF) and lexical patterns in step 1. The clinical event categories include Symptom, Purpose, Finding, Diagnosis, Drug, Treatment, Test, Time, and Visit. The basic semantic structure of a Problem-Action relationship between events consists of a time and the clinical events that occurred at the relevant time. We define this basic unit as a clinical semantic unit. In step 2, we segment clinical semantic units (CSU) using proposed algorithm for finding CSU. The segmented CSUs in step 2 are classified into four classes using multiclass Support Vector Machine (SVM) with event causality features as well as with lexical and context features in step3. These four classes are ‘ TAP' , “ Problem ”, ‘ Action ’, and ‘ Problem-Action ’ relation. In the last step, we extract Problem-Action relations based on event causality patterns. Events in CSU are arranged at classified relations ( TAP ; Problem ; Action ; Problem-Action ) by event causality patterns.