NATURAL LANGUAGE PROCESSING-BASED INFORMATION EXTRACTION FOR MEDICAL KNOWLEDGE GRAPH CONSTRUCTION

• Main Author: Reinhart, Richard
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/87686
• Institution: Institut Teknologi Bandung

Every day, hospitals generate new data related to patients. One of the challenges is that patient information is highly diverse and massive, making certain tasks performed by healthcare professionals, such as retrieving patient medical records, repetitive and time-consuming. Harnoune et al. (2021) proposed a solution in the form of a knowledge graph to address this issue. However, the resulting knowledge graph used the same edge labels for every edge connecting two same type entities, thus failing to provide information such as positive or negative relationships. This study develops the construction of a medical knowledge graph by utilizing Named Entity Recognition (NER) to identify entities such as disease names, medications, or medical procedures. Part-of-Speech (POS) Tagging and Dependency Parsing are used to determine the words functioning as verbs and roots. These words are then used as the relationships between entities in the knowledge graph. This approach aims to generate a graph structure with relationships that align with the contextual connections between entities in the medical domain. The resulting knowledge graph is evaluated using both quantitative and qualitative methods. Quantitative evaluation involves measuring metrics such as precision, recall, and F1-score, which achieved results of 0.89, 0.93, and 0.91, respectively. iv Meanwhile, qualitative evaluation is conducted by involving experts in the medical and informatics domains to assess the correctness and informativeness of the constructed knowledge graph, with scores of 4.2 out of 5 and 3.9 out of 5, respectively. The results demonstrate that the NER, POS Tagging, and Dependency Parsing-based approach is capable of constructing an informative and valid medical knowledge graph, yielding favorable evaluation results both quantitatively and qualitatively.