Patient empowerment is strongly affected by heath literacy, i.e., the patient's capacity to navigate healthcare systems, make informed decisions, and engage effectively in promoting personal and community health [1], [2]. Considering the rapidly increasing availability of patient-facing healthrelated content [3], especially in digital settings, we focus on personalised medical content delivery to support patient empowerment. Our approach distinguishes itself by considering more profound elements affecting a patient's comprehension and engagement with the provided content, including cognitive competencies, socio-economic factors, and cultural insights. We leverage existing medical knowledge resources and patient education guidelines to propose and develop HEALIE (HEALth Information Enhancement), a Knowledge Graph-enabled, personalised medical content generation system.
We build on the integrated conceptual model established by Sørensen and colleagues [4], which consolidates four comprehensive evidence-based dimensions of health literacy, i.e. Accessing, Understanding, Appraising, and Applying information. In this work, we focus only on mapping the second (Understand) and third dimension (Appraise) and we add specific cognitive factors to Sørensen's model. Cognitive factors are characteristics of an individual that affect performance and learning, i.e. memory-related cognitive skills can affect a patient's ability to recall a medical text and require modifications, like repeating information multiple times or offering a bullet-point summary at the end. Or a low Knowledge factor implies the need for layman terms (e.g. replacing "hematological cancer" with "blood cancer"). Our model also includes some social determinants of health that impact well-being and health literacy. Each factor's importance has been ranked and incorporated into the model in the form of weights.
The HEALIE KG's development started with an extensive literature review and identification of available data resources. Then the conceptual model was created, defining five major node clusters and diverse data types, e.g. textual descriptions of medical concepts, coupled with visual aids (diagrams, infographics, images) to accommodate to different cognitive profiles and to aid in conveying complex health information. Finally, the KG was instantiated with data points collected from the literature; ontologies and expert insights were organized and structured as nodes, properties, and relationships within the graph, following data harmonisation. Patient nodes with full profiles (clinical and demographic data, social determinants of health, cognitive factor scores etc.), corresponding to indicative use cases were also incorporated (see Figure 1). Neo4j and its Cypher query language was used to store the KG, execute path navigation queries, and extract the necessary information. For the automatic content creation, we used RosaeNLG, a template-based NLG software. RosaeNLG dynamically ingests the KG queries' results, uses graph processing algorithms and bespoke rules curated by the domain experts to gather information for the text generation components and to identify the appropriate template type. The dynamic content spans from simple synonyms to entire text blocks of cognitive-related information, multimedia, and even customised appearance (font, font size, colour) for the generated content.