Dynamic prediction of extra-corporal circuit clotting risk during hemodialysis sessions using deep learning: a retrospective study

Extra-corporal circuit (ECC) clotting during hemodialysis (HD) is a well-known issue that is hard to detect early, potentially leading to blood loss and reduced treatment duration. Ideally, clotting should be detected in time to administer additional anticoagulants or initiate restitution. However, most studies focus on finding the ideal anticoagulation strategy without the support of any data-driven solution. This study aims to develop an Artificial Intelligence model for dynamic ECC clotting-risk prediction using in-session temporal monitoring data (main data) and patients’ out-session clinical data (contextual data), with the goal of providing early alerts. The model is trained to predict the risk that a dialysis session will end with clotting-related outcome, rather than to detect clot formation at a specific time point. Leveraging Deep Learning for dynamic prediction, this study experiments with Long Short-Term Memory (LSTM) architecture and two contextual LSTM approaches (concatenation and multiplication) to identify the best-fitting architecture. The impact of incorporating patients’ out-of-session clinical data is also assessed by testing an LSTM without these features. Data were collected from electronic records at Reims University Hospital (2015–2021), covering 113,119 sessions with 47 in-session and 28 out-session features. Performance differences between models were minimal. The final model, LSTM with raw concatenation of in- and out-session data, reached an F1-Score of 0.8 and an area under the ROC curve of 0.89 for session-end prediction. In a simulated clinical evaluation using sequential predictions, that model surpassed an F1-Score of 0.67 only after two hours of session time had elapsed (sensitivity: 0.7, precision: 0.64), out of an average session length of four hours. Strong performance in session-end prediction suggests dynamic clotting-risk prediction is feasible with Deep Learning. However, early-session predictions remain challenging, likely due to the granularity of clotting annotations in the study. These results provide a solid foundation for further advancements.