Introduction In the Netherlands, approximately five thousand people are dependent on maintenance hemodialysis (HD) treatment. The one-year mortality of this patient population in 2021 was as high as 18%. A significant contributor to this high mortality and burden on HD patients is the high incidence of intradialytic hypotension (IDH), which can lead to hypoperfusion of vital organs. Despite advances in fluid management, including technical innovations such as relative blood volume (RBV) monitoring and bioimpedance measurements, the incidence of IDH remains unacceptably high. Therefore, the primary goal of this master thesis is to assess the feasibility of machine learning-based real-time prediction of IDH using patient data from HD sessions at the LUMC with the aim to enhance individualized fluid assessment. In this study, the performance of a recurrent neural network, known for its ability to retain time-varying information, will be compared to other machine learning models. Methods Patient demographics and intradialytic measurements, such as blood pressure and RBV measurements, were obtained from all HD patients of the LUMC with an age ≥18 years. The two definitions of IDH that were assessed were defined as IDH1, a drop of ≥20 mmHg compared to the pre-dialytic SBP and IDH2, is an absolute intradialytic SBP of ≤90 mmHg. Performance of a logistic regression, random forest and extreme gradient boosting (XGBoost) model were assessed and compared to that of a recurrent neural network (RNN). The models provided a prediction of IDH at every 30 minutes during HD treatment. Feature importances were determined for the logistic regression, random forest and XGBoost and two subanalyses were conducted to assess predictions made 60 minutes ahead instead of 30 minutes, and the effect of removal of blood pressure measurements on the prediction. Results 37,025 HD sessions of 436 patients were included in the analysis of IDH1 and 43,722 sessions of 441 patient in the analysis of IDH2. The Area under the Curve (AUC) of the Receiver Operating Characteristics (ROC) curve for the logistic regression, random forest, XGBoost and RNN were 0.80, 0.81, 0.81 and 0.87 respectively for the prediction of IDH1 and 0.93, 0.84, 0.86 and 0.92 for IDH2. At the highest clinically acceptable false positive rate (FPR) of 0.1, the RNN demonstrated a precision, recall, and specificity of 0.56, 0.59, and 0.90, respectively, for IDH1, and 0.10, 0.76, and 0.90 for IDH2, and the logistic regression 0.63, 0.50 and 0.90 for IDH1 and 0.14, 0.80 and 0.90 for IDH2. Conclusion The machine learning algorithms showed the ability to learn from HD data to make a prediction of IDH at different intradialytic timepoints. However, the current models do not meet the criteria for clinical implementation due to their limited performance metrics. Evaluation of the subanalyses revealed that the RNN had the ability to detect patterns that did not only rely on the blood pressure. Therefore, the prospect of real-time predictions of IDH appears promising with further refinement of the RNN and the expansion of the database by incorporating additional features and sessions.