Representation Learning of Finding Codes in Structured Echocardiogram Reporting

Structured echocardiogram reporting system uses predefined finding codes (FCs) that correspond to one-sentence descriptions to report medical findings. Exploring the latent relationships between FCs will facilitate the efficiency and accuracy in reporting by developing strategies such as automatically correlating clinical findings or detecting mutual exclusive findings. In this work, we propose an unsupervised learning model to explore the rich latent relationships between FCs and to learn representations for FCs utilizing narratives of findings. Experimental results on a data set containing structured reports collected from a real clinical institute show the effectiveness of our method.