A genetic algorithm enabled ensemble for unsupervised medical term extraction from clinical letters

Despite the rapid global movement towards electronic health records, clinical letters written in unstructured natural languages are still the preferred form of inter-practitioner communication about patients. These letters, when archived over a long period of time, provide invaluable longitudinal clinical details on individual and populations of patients. In this paper we present three unsupervised approaches, sequential pattern mining (PrefixSpan); frequency linguistic based C-Value; and keyphrase extraction from co-occurrence graphs (TextRank), to automatically extract single and multi-word medical terms without domain-specific knowledge. Because each of the three approaches focuses on different aspects of the language feature space, we propose a genetic algorithm to learn the best parameters of linearly integrating the three extractors for optimal performance against domain expert annotations. Around 30,000 clinical letters sent over the past decade from ophthalmology specialists to general practitioners at an eye clinic are anonymised as the corpus to evaluate the effectiveness of the ensemble against individual extractors. With minimal annotation, the ensemble achieves an average F-measure of 65.65 % when considering only complex medical terms, and a F-measure of 72.47 % if we take single word terms (i.e. unigrams) into consideration, markedly better than the three term extraction techniques when used alone.