Graph neural network modelling as a potentially effective method for predicting and analyzing procedures based on patients' diagnoses

Background: Currently, the healthcare sector strives to improve the quality of patient care management and to enhance/increase its economic performance/efficiency (e.g., cost-effectiveness) by healthcare providers. The data stored in electronic health records (EHRs) offer the potential to uncover relevant patterns relating to diseases and therapies, which in turn could help identify empirical medical guidelines to reflect best practices in a healthcare system. Based on this pattern of identification model, it is thus possible to implement recommender systems with the notion that a higher volume of procedures is often associated with better high-quality models.

Methods: Although there are several different applications that uses machine learning methods to identify such patterns, such identification is still a challenge, due in part because these methods often ignore the basic structure of the population, or even considering the similarity of diagnoses and patient typology. To this end, we have developed a method based on graph-data representation aimed to cluster 'similar' patients. Using such a model, patients will be linked when there is a same and/or similar patterns are being observed amongst them, a concept that will enable the construction of a network-like structure which is called a patient graph.¹ This structure can be then analyzed by Graph Neural Networks (GNN) to identify relevant labels, and in this case the appropriate medical procedures that will be recommended.

Results: We were able to construct a patient graph structure based on the patient's basic information like age and gender as well as the diagnosis and the trained GNNs models to identify the corresponding patient's therapies using a synthetic patient database. We have even compared our GNN models against different baseline models (using the SCIKIT-learn library of python) and also against the performance of these different model-methods. We have found that the GNNs models are superior, with an average improvement of the f1 score of 6.48 % in respect to the baseline models. In addition, the GNNs models are useful in performing additional clustering analysis which allow a distinctive identification of specific therapeutic/treatment clusters relating to a particular combination of diagnoses.

Conclusions: We found that the GNNs models offer a promising lead to model the distribution of diagnoses in patient population, and is thus a better model in identifying patients with similar phenotype based on the combination of morbidities and/or comorbidities. Nevertheless, network/graph building is still challenging and prone to biases as it is highly dependent on how the ICD distribution affects the patient network embedding space. This graph setup not only requires a high quality of the underlying diagnostic ecosystem, but it also requires a good understanding on how patients at hand are identified by disease respectively. For this reason, additional work is still needed to better improve patient embedding in graph structures for future investigations and the applications of this service-based technology. Therefore, there has not been any interventional study yet.