Section heading recognition in electronic health records using conditional random fields

Electronic health records (EHRs) contain a wealth of information, such as discharge diagnoses, laboratory results, and pharmacy orders, which can be used to support clinical decision support systems and enable clinical and translational research. Unfortunately, the information is represented in a highly heterogeneous semi-structured or unstructured format with author- and domainspecific idiosyncrasies, acronyms and abbreviations. To take full advantage of health data, text-mining techniques have been applied by researchers to recognize named entities (NEs) mentioned in EHRs. However, the judgment of clinical data cannot be known solely from the NE level. For instance, a disease mention in the section of past medical history has different clinical significance when mentioned in the family medical history section. To obtain high-quality information and improve the understanding of clinical records, this work developed a machine learning-based section heading recognition system and evaluated its performance on a manually annotated corpus. The experiment results showed that the machine learning-based system achieved a satisfactory F-score of 0.939, which outperformed a dictionary-based system by 0.321.