Impact of Clinical Features on Disease Diagnosis Using Knowledge Graph Embedding and Machine Learning: A Detailed Analysis

Disease diagnosis is a primary but most critical step in the effective treatment of a patient. Research has shown the significance of Artificial Intelligence (AI) in diagnosing diseases using patient’s clinical features. However, the extent of their impact on AI-based disease predictions is a relatively unexplored area of research. Therefore, this paper explores the importance of various clinical features in disease diagnosis using a Knowledge Graph (KG)-based framework and presents a detailed analysis by applying different embedding methods and Machine Learning (ML) algorithms. The framework semantically organizes patient’s clinical details in a KG which is further mapped to embeddings using TransE, ComplEx, PairRE, AutoSF, and DistMA. Three ML models, Logistic Regression (LR), Support Vector Machine (SVM) and eXtreme Gradient Boosting (XGBoost), are applied to KG embeddings for disease diagnosis. The diagnosis is performed with seven different combinations of clinical features to analyse their impact on disease diagnosis. The experimental results are obtained with the MIMIC-III dataset. PairRE gives the best performance among its peers and hence was used for further experiments to generate KG embeddings and train ML models. The results show that the performance of a model increases by 12.5% when all clinical features are taken as input compared to a single feature input. Hence, it can be concluded that all the features are important. However, LR and SVM-based predictions show that disease history plays a comparatively more important role in disease diagnosis.