Capturing disease-symptom relations using higher-order co-occurrence algorithms

The field of medical informatics has been thriving over the last decade. One critical task in medical informatics is whether computational algorithms allow for predicting diseases from symptoms and vice versa. A niche of algorithms that has not been explored extensively are computational linguistic in nature and focus on higher-order co-occurrences of language units, such as words and paragraphs. The current study explored whether disease-symptom relations can be captured using such higher-order co-occurrences. Results indicated that higher order co-occurrences allow for capturing the semantic relation between disease and symptom. Two algorithms were tested, one using latent semantic analysis (LSA), which typically ignores the role of negations in language, and a customized LSA algorithm that took negations into account. Both algorithms predicted the semantic relations between symptoms and diseases well above chance level, with the customized algorithm outperforming the original LSA algorithm.