Heart transplantation is a life-saving procedure for children affected by end-stage heart failure. However, despite recent improvements in long-term outcomes, 1-year post-transplantation mortality has remained relatively high. Accurate prediction of post-transplantation mortality is crucial to evaluating risks related to recipient-donor matches. Machine learning techniques can potentially improve the current allocation system through the integration of a larger set of features. In this work, we improve 1-year mortality prediction after pediatric heart transplantation using a new self-training approach based on generating artificial recipient-donor pairs as synthetic unlabeled observations. We tested and compared our approach to several baselines using the cohort of pediatric patients in the UNOS database. Our study suggests that augmenting the dataset with proper synthetic observations can improve the prediction of 1-year mortality after pediatric heart transplantation. Our findings have implications for the future of heart transplantation in children, offering a potential path to refine recipient-donor matching and improve survival rates. This study contributes to the growing field of applying advanced machine learning techniques in medical decision-making, specifically in the context of organ transplantation.
Massive Open Online Courses (MOOCs) have recently attracted students and professionals as complementary tools to academic education. Despite the number of advantages MOOCs provide, such as openness and flexibility regarding learning pace, such courses are characterized by a consistently higher dropout rate than conventional classrooms. A crucial factor that influences dropout is the choice of the appropriate course, hence the need for effective course recommendations. A course recommendation system (RS) that uses dropout information can mitigate course withdrawal and user dissatisfaction. In this paper, an extension of Bayesian Personalized Ranking, which is a learning-to-rank RS, is proposed that uses the pseudo-labels extracted by survival analysis based on dropout information to recommend courses in the context of MOOCs. The proposed approach performs the best compared to six competing RSs on three MOOCs datasets.