Graph Neural Network Modelling as a potentially effective Method for predicting and analyzing Procedures based on Patient Diagnoses

Abstract

AbstractBackgroundCurrently, the healthcare sector strives to increase the quality of patient management and improve the economic performance of healthcare providers. The data contained in electronic health records (EHRs) offer the potential to discover relevant patterns that aim to relate diseases and therapies, and thus discover patterns that could help identify empirical medical guidelines that reflect best practices in the healthcare system. Based on this pattern identification, it is then possible to implement recommendation systems based on the idea that a higher volume of procedures is associated with high-quality models.MethodsAlthough there are several applications that use machine learning methods to identify these patterns, this identification is still a challenge, in part because these methods often ignore the basic structure of the population, considering the similarity of diagnoses and patient typology. To this end, we have developed graph methods that aim to cluster similar patients. In such models, patients are linked when the same or similar patterns can be observed for these patients, a concept that enables the construction of a network-like structure. This structure can then be analyzed with Graph Neural Networks (GNN) to identify relevant labels, in this case the appropriate medical procedures.ResultsWe report the construction of a patient Graph structure based on basic patient’s information like age and gender as well as the diagnoses and trained GNNs models to identify the corresponding patient’s therapies using a synthetic patient database. We compared our GNN models against different baseline models (using the SCIKIT-learn library of python) and compared the performance of the different model methods. We have found that GNNs are superior, with an average improvement of the f1 score of 6.48% respect to the baseline models. In addition, the GNNs are useful for performing additional clustering analyses that allow specific identification of specific therapeutic clusters related to a particular combination of diagnoses.ConclusionsWe found that GNNs are a promising way to model the distribution of diagnoses in a patient population and thus better model how similar patients can be identified based on the combination of morbidities and comorbidities. Nevertheless, network building is still challenging and prone to prejudice, as it depends on how ICD distribution affects the patient network embedding space. This network setup requires not only a high quality of the underlying diagnostic ecosystem, but also a good understanding of how to identify related patients by disease. For this reason, additional work is needed to improve and better standardize patient embedding in graph structures for future investigations and applications of services based on this technology, and therefore is not yet an interventional study.