Tracking medication information across medical records

Abstract

A patient's electronic medical record can consist of a large number of reports, especially for an elderly patient or for one affected by a chronic disease. It can thus be cumbersome for a physician to go through all of the reports to understand the patient's complete medical history. This paper describes work in progress towards tracking medications and their dosages through the course of a patient's medical history. 923 reports associated with 11 patients were obtained from a university hospital. Drug names were identified using a dictionary look-up approach. Dosages corresponding to these drugs were determined using regular expressions. The state of a drug (ON, OFF), which determines whether or not the drug was being taken, was identified using a support vector machine with features based on expert knowledge. Results were promising: prec. approximately recall approximately 87%. The output is a timeline display of the drugs which the patient has been taking.

Figure 1.

Java-based annotation tool. Word labels are assigned to generate a gold standard and for training data.

Figure 2.

Timeline visualization with the results of the drug tracking process. A dark cell means that a drug is ON.

Figure 3.

ROC curve for the drug state classifier. The area under the curve is A_z=0.856. The operating point has been marked with a cross.