Prediction of hypotension during haemodialysis through data analytics and machine learning

Background: Patients undergoing haemodialysis (HD) are often exposed to complications due to their treatment, whether at home or in-centre. Intradialytic hypotension (IDH), the most serious of the adverse effects, is associated with increased mortality. The ability to accurately predict a future episode of IDH would be a clinically useful tool so that preventative measures can be taken, yet no tool for the management and decision-making of IDH has been available until now. This study has aimed to investigate the scope of machine learning (ML) techniques in predicting IDH by analysing various standard clinical variables measured by clinicians. Methods: Using only routinely measured clinical observations, this study has identified the key variables and used machine learning to build an IDH predictor. The dataset comprised 73323 haemodialysis sessions with 36662 IDH events, collected from 3944 patients in 10 centres during 2000–2020. Results: The analysis revealed systolic and diastolic blood pressure (SBP and DBP) as key predictor variables. Patients with IDH had lower pre-dialysis SBP and a greater percentage drop in their SBP during dialysis. They also had lower pre-dialysis DBP and a greater percentage drop in their DBP. The probability of IDH was increased by lower pre-SBP, greater delta-systolic and increased weight loss. A machine-learning model with Random Forest had the highest specificity (73.9%) and highest overall predictive accuracy (75.5%), but a model with Bidirectional Long Short-Term Memory (Bi-LSTM) achieved the highest sensitivity (78.5%) in predicting IDH. Tested on separate validation data, the Bi-LSTM model achieved accuracy 74.09%, sensitivity 74.81%, specificity 73.3% and a ROC-AUC score of 0.816. Using only pre-dialysis data, which would be available at the start of a dialysis session, the prediction performance dropped to accuracy 68.60%, sensitivity 67.40%, specificity 69.81%, and a ROC-AUC score of 0.757. However, as some variables can be measured in real time, the model’s performance can be expected to rise during dialysis toward that of the model using all data. Conclusions: Data analytics and machine learning can help to predict and avoid IDH. A prediction model using ML algorithms offers great promise as a tool in identifying patients at risk of IDH in advance. Moreover, adding the data measured during the dialysis could improve the model further and lead to personalised management and intervention of IDH. Future work will involve building a decision-support system for clinicians and conducting a clinical trial.