Practical approach for disease similarity calculation based on disease phenotype, etiology, and locational clues in disease names

摘要

Disease similarity is a useful measure to improve clinical decision support systems wherein it allows continuous presentation of similar diseases. In a previous study, we demonstrated that etiological and symptomatic information of diseases provide a reasonable approximation for the similarity of diseases. This study extends the previously proposed approach by incorporating the locational information of diseases, which may improve the performance against the baseline achieved only by the etiological and symptomatic features.First, this study attempted to acquire anatomical information from disease names in Japanese and evaluated the relation between those disease names by constructing a graph that represents the relations of anatomical concepts. Then, locational information was used as a supplemental input to the disease similarity calculation in the machine learning framework developed for the previous study. The baseline performance of the algorithm presented in the previous study was 0.907 using the normalized disease similarity measure (NDSM) which utilizes 80 annotated diseases as the gold standard. The proposed approach with the locational feature achieved a higher score: 0.921. As expected, in-depth microscopic analysis revealed that this approach benefits the diseases that have tight association with specific locations in the human body. The results suggest that location is a promising feature for computing disease similarity, which is worth further investigation by acquiring locational knowledge from disease profiles.