Finding Negative Symptoms of Schizophrenia in Patient Records

This paper reports the automatic extraction of eleven negative symptoms of schizophrenia from patient medical records. The task offers a range of difficulties depending on the consistency and complexity with which mental health professionals describe each. In order to reduce the cost of system development, rapid prototypes are built with minimal adaptation and configuration of existing software, and additional training data is obtained by annotating automatically extracted symptoms for which the system has low confidence. The system was further improved by the addition of a manually engineered rule based approach. Rule-based and machine learning approaches are combined in various ways to achieve the optimal result for each symptom. Precisions in the range of 0.8 to 0.99 have been obtained.