Node2VecFuseClassifier: Bridging Perspectives in Modeling Transplantation Attitudes Among Dialysis Patients

Rafaa Aljurbua, Avrum Gillespie, Jumanah Alshehri, Abdulrahman Alharbi, Nouf Albarakati, Zoran Obradovic

Abstract: The study of patient attitudes toward transplantation is crucial for dialysis clinic decision-makers. Understanding the factors influencing patient attitudes toward transplantation helps improving transplant access. While many studies attempt to predict patient attitudes toward transplantation from a patient-centric point of view by focusing on patient information, there is a limited number of studies that explore the influence of social interaction among dialysis patients on predicting their attitudes toward transplantation. In this study, we utilize machine learning-based models and social networks to classify hemodialysis patients' attitudes toward transplantation taking into account the influence of social interaction among dialysis patients. We conduct a questionnaire among 110 participants from two hemodialysis centers. We investigate the performance of machine learning algorithms, including linear, non-linear model, and graph-based machine learning models. We use developed models to classify hemodialysis clinic patients into positive and negative attitudes toward transplantation. We introduce a graph-based model named Node2VecFuseClassifier that integrates both patient inter-actions and patient features. To emphasize the significance of social interaction, we compare the benefits of using a sociodemographic patient-centric versus a social-centric problem representation that considers patient-to-patient and patient-to-staff interactions. We believe that including the social aspect, patient-to-patient, and patient-to-staff network features enhances all machine learning models' performance as compared to relying on patient-centric features alone. By combining patient experiences with staff expertise, the multilevel analysis enhances predictive capabilities by incorporating diverse roles like patients and staff. Thus, we combine patient and staff networks and observe that such a multi-network approach boosts the F1-score in predicting patient attitude toward transplantation by at least 5 % compared to using only the patient network or staff network. The proposed Node2VecFuseClassifier that combines Node2Vec embeddings and features improves the accuracy of transplantation attitudes prediction by 4% - 6%. Overall, our study shows that integrating the interaction between patients and staff provides beneficial insights to accurately predicting transplantation attitudes for dialysis patients.