Early prediction of post-dialysis systolic and diastolic blood pressure using intradialytic data in Senegalese hemodialysis patients: a single-center retrospective cohort study

Background Intradialytic blood pressure variability poses a major challenge in chronic hemodialysis, especially in resource-limited settings, contributing to cardiovascular complications and session interruptions. This study aimed to develop and evaluate supervised machine learning models to predict post-dialysis systolic blood pressure (SBP) and diastolic blood pressure (DBP) using only data available up to the third hour of the session. Methods In this retrospective, single-center cohort study at Hôpital Militaire Ouakam (Dakar, Senegal; February 2023 to November 2024), we analyzed 7,600 hemodialysis sessions from 124 chronic patients after quality-based exclusions. Predictors included intradialytic hemodynamics, dialysis machine parameters, and demographic and contextual factors recorded up to hour 3, while end-of-session data were excluded to prevent leakage. Four models were retained after initial screening of six candidates (Ridge Regression, Random Forest, LightGBM, and XGBoost) and evaluated using patient-level 5-fold cross-validation, with mean absolute error across SBP and DBP as the primary performance metric. Results Tree-based models outperformed linear and kernel-based methods, with the optimized Random Forest model achieving the best performance, with an MAE of 12.67 ± 0.80 mmHg for SBP and 8.93 ± 0.35 mmHg for DBP, a root mean square error of 14.68 ± 0.75 mmHg, and R² of 0.553 ± 0.033. Diastolic blood pressure predictions were more stable than systolic predictions, with no systematic bias observed in Bland-Altman analyses (bias = 0.015 mmHg for SBP and 0.395 mmHg for DBP). Subgroup performance showed moderate heterogeneity related to age, sex, and body mass index, with the best results observed in younger male patients. Conclusion These findings demonstrate the feasibility of supervised machine learning for unbiased prediction of post-dialysis blood pressure in Senegalese chronic hemodialysis patients, supporting the potential development of proactive interventions during hemodialysis sessions. Prospective and multicenter studies are needed to validate and measure the clinical impact of artificial intelligence use in the management of dialysis patients in resource-limited settings.