The research goals described in the different articles differ very much in detail, but an application often mentioned in the articles was personalized medicine, which was named by 6 of the 11 investigated papers. The other goals were quality improvement, information gaining, predictive modeling, disease diagnosis, patient segmentation (each used in 2 papers), population management, data mining, data warehouse and disease pattern (each used in one paper).
To reach these goals the papers follow very different strategies. Atif et al. used image-based information of brains to create a graph-based cerebral description of brain anatomy. This spatial graph is created manually for every patient and afterwards patients could be compared using these graphs [
19]. In contrast, Campbell et al. used the SNOMED CT concept model in a graph database architecture because of the ontology character and polyhierarchy of SNOMED CT, which makes it difficult to implement electronic health records in relational databases [
26]. The created graphs save SNOMED CT data in a specifically created graph format. Risk prediction is the main goal of Chen et al., so the authors developed a graph-based, semi-supervised learning algorithm to reach this goal [
27]. By modelling the clinical evolution of an individual patient with kidney failure Esteban et al. wanted to develop the basis for a future clinical decision support system. This graph-based model contains thousands of events like laboratory results, ordered tests and diagnoses [
28] and represents a patient in a graph. Hanzlicek et al. described MUDR EHR, a multimedia distributed health record for decision support. This electronic health record contains multiple medical concepts, which should help describing a patient in a structured way, apart from free text records [
34]. Kaur et al. described a model, which combines different data stores of patient data. In this architecture the user creates his request at the interface and the architecture below translates this request into a query to get the data from the most suitable data store for this request [
29]. The resulting graph of this paper helps to get the right information from the data stores. Liu et al. used longitudinal patient data to create so called temporal graphs. These graphs were clustered in different phenotypes, so that using these phenotypes helps improving diagnosis performance [
20]. The resulting graphs represent a patient and his medical events in temporal context. The focus of Müller et al. was the lack of clinical context in other approaches. The authors solve this problem by creating a graph-grammar approach to design and implement a graph-oriented patient model, which allows the representation of the clinical context [
30]. Puentes et al. also used (similar to Atif et al.) graphs to gain information out of image-based brain information to model spatial relationships of brain anatomical singularities of individual patients. This approach is especially used for spatial modelling of cerebral tumors [
31]. Zhang et al. [
32] created a convolutional neural network on heterogeneous attributes of a patient (e.g. diagnoses, procedures and medications) using a graph, which gains its data from electronic health records [
32]. Zhang et al. [
33] created a unified graph representation of the electronic health records of an individual patient in a temporal manner. Using this graph, in the second step a modified algorithm was used to create a temporal profile of each patient. This approach was used for risk prediction [
33].