Automatic Classification of Cancer Notifiable Death Certificates

The timely notification of cancer cases is crucial for cancer monitoring and prevention. However, the abstraction and classification of cancer from the free-text of pathology reports and other relevant documents, such as death certificates, are complex and time-consuming activities. In this paper we investigate approaches for the automatic detection of cases where the cause of death is a notifiable cancer from free-text death certificates supplied to Cancer Registries. A number of machine learning classifiers were investigated. A large set of features were also extracted using natural language techniques and the Medtex toolkit; features include stemmed words, bi-grams, and concepts from the SNOMED CT medical terminology. The investigated approaches were found to be very effective in identifying death certificates where the cause of death was a notifiable cancer. Best performance was achieved by a Support Vector Machine (SVM) classifier with an overall F-measure of 0.9647 when evaluated on a set of 5,000 free-text death certificates. This classifier considers as features stemmed token bigrams and information from SNOMED CT concepts filtered by morphological abnormalities and disorders. However, our analysis shows that it is the selection of features that most influences the performance of the classifiers rather than the type of classifier or the feature weighting schema. Specifically, we found that stemmed token bigrams with or without SNOMED CT concepts are the most effective feature. In addition, the combination of token bigrams and SNOMED CT information was found to yield the best overall performance.