Investigating the Value of Diagnosis Codes in the Primary Care Patient Record

We have used supervised document classification to investigate whether or not the ICPC diagnosis code used in the primary care patient record complements the associated encounter note. Our hypothesis is that diagnosis codes are set independently from the notes and that the code accordingly provides additional information that is not reflected in the note. To investigate this hypothesis we built a set of document classifiers using data from a primary care practice and attempted to predict the correct diagnosis code for each encounter note. At best we achieved a correct prediction rate of 51.6 %. Given our results we discuss possible reasons behind misclassifications and find some indicators that the diagnosis code may add information that can not be inferred from the encounter note.