GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs

Clinical predictive models often rely on patients' electronic health records (EHR), but integrating medical knowledge to enhance predictions and decision-making is challenging. This is because personalized predictions require personalized knowledge graphs (KGs), which are difficult to generate from patient EHR data. Our method extracts knowledge from large language models (LLMs) and external biomedical KGs to build patient-specific KGs, which are then used to train our proposed Bi-attention AugmenTed (BAT) graph neural network (GNN) for healthcare predictions. On MIMIC-III, it boosts AUROC by 17.6% and 6.6% for mortality and readmission, and F1-score by 7.9% and 10.8% for LOS and drug recommendation, respectively. To improve predictive performance and integrate expert knowledge with data insights, clinical knowledge graphs (KGs) were adopted to complement EHR modeling (Chen et al., 2019; Choi et al., 2020; Rotmensch et al., 2017). These KGs represent medical concepts (e.g., diagnoses, procedures, drugs) and their relationships, enabling effective learning of patterns and dependencies. However, existing approaches mainly focus on simple hierarchical relations (Choi et al., 2017; 2018; 2020) rather than leveraging comprehensive relationships among biomedical entities despite incorporating valuable contextual information from established biomedical knowledge bases (e.g., UMLS (Bodenreider, 2004)) could enhance predictions. Moreover, large language models (LLMs) such as GPT (Brown et al., 2020; Chowdhery et al., 2022; Luo et al., 2022; OpenAI, 2023) pre-trained on web-scale biomedical literature could serve as alternative resources for extracting clinical knowledge given their remarkable reasoning abilities on open-world data. There is a substantial body of existing research demonstrating their potential use as knowledge bases (Lv et al., 2022; Petroni et al., 2019; AlKhamissi et al., 2022). To fill the gap in personalized medical KGs, we propose to leverage the exceptional reasoning abilities of LLMs to extract and integrate personalized KG from open-world data. Below: For each patient, we compose a patient-specific graph by combining the concept-specific KGs associated with them and make the graph temporal with sequential data across patient's visits ( 3.2). To utilize the patient graph for predictions, we employ a bi-attention augmented graph neural network (GNN) model, which highlights essential visits and nodes with attention weights ( 3.3). As shown in Figure 1, our patient KG generation module first takes medical concepts as input and generates concept-specific KGs by prompting LLMs or retrieving subgraphs from existing graphs.