Machine learning for intradialytic hypotension prediction in haemodialysis patients

Introduction: Patients undergoing haemodialysis (HD) are exposed to other morbidities due to their treatment, whether at home or in a clinical setting. Intradialytic hypotension (IDH), the most serious of the adverse effects, is associated with increased mortality. Early intervention of such a dangerous clinical event is highly desirable for prevention, 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: The dataset consists of 73323 HD sessions with 36662 IDH events. The data have been collected from 3944 patients in 10 centres during 2000-2020. Different data analytics and ML techniques were applied to explore the important variables associated with IDH. The probabilities of having an IDH were investigated using different combinations of variables. The prediction model was built to predict the IDH using various ML algorithms. The dataset was split 80:20 for training and testing purposes. Moreover, to validate the models further, we used a previously unseen validation dataset containing 6304 HD sessions collected from 748 patients from 10 sites during 2020-2021.

Results: Feature analysis using ML techniques (extra tree classifier and correlation matrix with heat map) indicated that pre-and post-systolic and diastolic blood pressure are useful predictor variables for IDH during an HD session. Statistical analysis suggests that IDH patients had significantly lower pre-dialysis systolic blood pressures (124 mmHg compared with 141 mmHg; p<0.05) and demonstrated a more significant percentage drop in their systolic blood pressure (delta systolic) before and after dialysis (9.18% vs 5.33%; p<0.05). Statistical analysis also showed lower pre-dialysis diastolic blood pressures (66 mmHg compared with 73 mmHg; p<0.05) and demonstrated a significant percentage drop in the patients’ diastolic blood pressure (delta diastolic) before and after dialysis (4.02% vs 2.31%; p<0.05). Investigation of different combinations of variables shows that lower pre-dialysis systolic blood pressures increase the probabilities of having an IDH. An ML model with random forest (RF) outperformed other algorithms in predicting events with an accuracy of 75.5%. An ML model with bidirectional long short-term memory (Bi-LSTM) outperformed different algorithms in predicting IDH events with an accuracy of 78.5%. The model with the RF algorithm gained a ROC-AUC score of 0.832 in comparison with Bi-LSTM’s 0.822. The ML model with Bi-LSTM also performed well on the validation dataset with an accuracy of 74.09%, sensitivity of 74.81%, specificity of 73.37% and ROC-AUC score of 0.816.

Conclusions: 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.