Inclusion-Exclusion Knowledge Filtering Approach for Conversation-Based Preliminary Diagnosis

Using natural language processing (NLP) techniques, we conducted a preliminary diagnosis of the disease from the patient syndrome description. Because patients are not medical professionals, they cannot accurately describe all symptoms. To solve this issue, we build a medical knowledge graph (KG) by constructing symptom-disease relation triples for pre-processing the patient syndrome description. According to the medical KG, the descriptions were reconstructed into KG embedding representation. To avoid the knowledge noise issue, we investigate an inclusion-exclusion knowledge filtering approach (IKFA) for symptom-to-disease triples to load them to a pretrained language model (PLM), i.e., bidirectional encoder representations from Transformers (BERT). To train the IKFA, we built a medical diagnosis question-answer dataset (MDQA dataset), which contains large-scale and high-quality questions (patient symptom description) and answers (diagnosis) (Q&A) corpus with 1.63 million entries in the size of 213 MB. The KG was built based on 8,731 diseases with detailed syndrome descriptions in the size of 1.98 MB. The experimental results showed that the IKFA preliminarily diagnosed 8,731 different diseases based on the patient's initial symptom description with an accuracy of 0.9894.