In a retrospective cohort study at Hôpital Militaire de Ouakam in Dakar, Senegal, medical researchers demonstrated that supervised machine learning algorithms can predict post-dialysis blood pressure using data collected during the first three hours of treatment. Analyzing 7,600 sessions from 124 chronic hemodialysis patients, an optimized random forest model achieved the best accuracy, recording a mean absolute error of 12.67 ± 0.80 mmHg for systolic blood pressure and 8.93 ± 0.35 mmHg for diastolic blood pressure. Excluding end-of-session data established a non-invasive forecasting framework designed for proactive clinical decision-making.
Intradialytic blood pressure variability poses a major management challenge in maintenance hemodialysis, frequently causing cardiovascular complications and premature treatment terminations. As noted by the authors, chronic kidney disease constitutes a growing global health issue, with over 2.5 million patients treated with hemodialysis in 2020 and projections estimating 5.4 million by 2030. Cardiovascular complications represent the leading cause of mortality in this population, accounting for up to 50 percent of patient deaths. Furthermore, arterial hypertension affects more than 80 percent of individuals undergoing chronic hemodialysis. The pathophysiology of hypertension in dialysis differs substantially from the general population, arising primarily from chronic fluid overload, neurohormonal hyperactivity, and increased vascular resistance. Sustained hypervolemia increases cardiac output and triggers autoregulatory vasoconstriction, making dry weight adjustment essential during each session.
Anticipating intradialytic blood pressure profiles remains clinically difficult because systolic and diastolic pressures fluctuate unpredictably during four-hour sessions. Fluctuations depend on hydration status, cardiovascular and neurohormonal function, prescribed medications, dialyzer parameters, and individual responses to ultrafiltration. Supervised machine learning models have emerged as practical tools to anticipate these changes. Earlier research demonstrated the feasibility of predicting post-dialysis blood pressure from basic clinical indicators such as pre-dialysis pressure, interdialytic weight gain, and intradialytic weight loss in pediatric cohorts. However, most prior investigations were conducted in high-resource settings. Clinical data from African hemodialysis populations, where access to renal replacement therapy is constrained and logistical challenges are widespread, remain scarce, motivating this localized study.
To evaluate predictive feasibility in Senegal, the authors examined electronic hemodialysis records collected over 20 months between February 8, 2023, and November 9, 2024, at Hôpital Militaire de Ouakam. The initial dataset comprised 29,422 intradialytic logging records from 8,132 sessions across approximately 130 patients. After quality screening removed sessions with significant missingness, implausible values, or missing core variables, 7,600 sessions from 124 patients were retained. The cohort consisted of 52.4 percent male patients with an average age of 49.84 ± 15.45 years, with the youngest patient aged 19. Session frequency varied widely: while the median number of recorded sessions per patient was 22, the cohort mean reached 61.3 ± 61.9 sessions, ranging from one to 153 sessions, with approximately 50 percent of patients having fewer than 20 sessions and 37 percent having over 100.
Hemodialysis sessions lasted approximately four hours, with physiological and machine-related parameters recorded at four intradialytic checkpoints labeled T1 through T4. The prediction targets were post-dialysis systolic and diastolic blood pressure measured at the end of the session at time point T4. To prevent methodological data leakage and reproduce realistic clinical conditions, all T4 measurements were strictly excluded from model inputs. The models utilized predictor variables available exclusively during the first three hours at time points T1, T2, and T3. These inputs encompassed intradialytic hemodynamics including systolic pressure, diastolic pressure, mean arterial pressure, and heart rate, as well as machine metrics such as ultrafiltration rate, transmembrane pressure, venous pressure, and blood flows. Demographics, dry weight, vascular access, anticoagulation, and interdialytic weight gain completed the feature set.
To reflect intradialytic dynamics, the researchers engineered dedicated temporal features from raw measurements recorded during the first three hours. Alongside raw checkpoint values, models incorporated net changes between hours three and one, estimated hourly slopes, and mean three-hour exposures. For machine parameters, cumulative exposure sums over three hours represented treatment load, alongside the ratio of ultrafiltration volume to dry weight. Missing data were managed conservatively: hemodynamic variables underwent linear temporal interpolation within sessions, stable variables were propagated forward, and remaining missing values received median imputation. Categorical variables were processed using one-hot or ordinal encoding. Patient age was categorized into young adults aged 50 and younger, middle-aged adults between 51 and 65, and older adults over 65, with body mass index classified following World Health Organization criteria.
The team used multi-output regression pipelines in scikit-learn, evaluating models through patient-level five-fold GroupKFold cross-validation. This partitioning ensured that sessions from any single patient appeared exclusively in either the training set or the validation set, preventing data contamination and testing true inter-patient generalizability. Initial screening compared six candidate architectures against a naive median baseline predictor. The baseline registered high errors, with mean absolute errors of 21.27 mmHg for systolic pressure and 14.93 mmHg for diastolic pressure alongside negative R² values, demonstrating the substantial added value of machine learning. Tree-based models outperformed linear and kernel methods across all metrics. Hyperparameter optimization using randomized search across 50 iterations, conducted separately from cross-validation to prevent leakage, refined four algorithms for detailed analysis: ridge regression, random forest, LightGBM, and XGBoost.
All four tuned machine learning models demonstrated broadly similar predictive accuracy across the cohort. Systolic blood pressure mean absolute errors ranged from 12.67 mmHg for random forest to 12.74 mmHg for LightGBM, while diastolic mean absolute errors ranged from 8.93 mmHg for random forest to 9.22 mmHg for ridge regression. Pairwise Wilcoxon signed-rank tests revealed no statistically significant differences between any pair of tuned models, with all p-values remaining at or above 0.063, partly attributable to limited statistical power across five folds. The random forest model was selected because it recorded the lowest overall mean absolute error at 10.80 ± 0.53 mmHg, achieved superior cross-validation stability (standard deviation 0.53 vs. 0.63–0.67 mmHg), and showed minimal systematic bias.
Target-specific evaluations highlighted greater predictive stability for diastolic blood pressure than systolic blood pressure. For the random forest model, root mean square error reached 12.28 ± 0.55 mmHg for diastolic pressure and 17.09 ± 1.02 mmHg for systolic pressure, with coefficients of determination of 0.536 ± 0.054 and 0.569 ± 0.054, respectively, producing an overall R² of 0.553 ± 0.033. Diastolic predictions achieved a median absolute error of 6.67 mmHg, with 67.2 ± 1.5 percent of predictions falling within ±10 mmHg. Systolic predictions recorded a median absolute error of 9.45 mmHg, with 52.0 ± 1.8 percent within ±10 mmHg. Within a ±30 mmHg bound, the model captured 97.1 percent of diastolic and 92.3 percent of systolic outcomes, while mean absolute percentage error registered at 9.50 percent for diastolic pressure and 7.89 percent for systolic pressure.
Bland-Altman analysis confirmed that the random forest model possessed negligible systematic bias, recording a mean bias of 0.015 mmHg for systolic pressure and 0.395 mmHg for diastolic pressure. However, the 95 percent limits of agreement were wide, spanning from -33.54 to +33.57 mmHg for systolic pressure and -23.67 to +24.46 mmHg for diastolic pressure, reflecting substantial individual-level hemodynamic variability. Shapley Additive Explanations analysis verified that the algorithm relied on clinically coherent features rather than statistical artifacts. Blood pressure measurements recorded at the third hour of the session were dominant predictors for both targets. Third-hour systolic pressure had a mean absolute SHAP value of 11.90 mmHg, far exceeding hour two at 4.35 mmHg and pre-dialysis pressure at 1.27 mmHg, while third-hour diastolic pressure had a SHAP value of 6.57 mmHg.
Subgroup analyses revealed performance heterogeneity across strata defined by sex, age, and body mass index categories. Younger male patients showed better predictions, notably the underweight young male subgroup (systolic MAE 10.46 mmHg, 60.3 percent within ±10 mmHg) and the overweight middle-aged male subgroup (MAE 10.11 mmHg). Conversely, poorer performance was observed in middle-aged and older female cohorts; the normal-weight middle-aged female subgroup recorded the highest systolic error at 17.07 mmHg, with only 36.4 percent within ±10 mmHg. In obese older women, the model registered a negative R² for systolic pressure, failing to explain variance better than the global mean. Across all subgroups, diastolic blood pressure was consistently better predicted than systolic blood pressure, aligning with the greater physiological variability of systolic pressure under ultrafiltration.
The study’s predictive accuracy broadly parallels international findings, including research by Bou-Matar and colleagues reporting R² values of 0.58 for systolic and 0.52–0.54 for diastolic pressure in pediatric cohorts, and Huang and colleagues reporting a test-set systolic MAE of 11.97 mmHg and R² of 0.50 in adults. However, the authors noted several key limitations. The single-center retrospective design, session imbalance, lack of continuous blood volume monitoring, wide limits of agreement, and reduced accuracy at extreme blood pressures preclude immediate clinical application. Future implementations could automatically retrieve in-session parameters from electronic records to display predicted end-of-session pressures, flagging out-of-range trajectories for clinical review rather than triggering automated interventions. Prospective multicenter studies across diverse African hemodialysis centers remain necessary to validate these models and assess their real-world clinical impact in resource-limited nephrology care.
Subject of Research: Machine learning prediction of post-dialysis blood pressure from intradialytic clinical data
Article Title: Early prediction of post-dialysis systolic and diastolic blood pressure using intradialytic data in Senegalese hemodialysis patients: a single-center retrospective cohort study
Article References: Diankha, M. L., Ndongo, M., Sarr, I. L., Coundoul, B., Mbodj, B., Ndiaye, I., Fall, E. B., Gaye, B., Deme, C. B., & Seck, S. M. (2026). Early prediction of post-dialysis systolic and diastolic blood pressure using intradialytic data in Senegalese hemodialysis patients: a single-center retrospective cohort study. Frontiers in Digital Health, 8, Article 1919498. https://doi.org/10.3389/fdgth.2026.1919498
Image Credits: AI Generated
DOI: 10.3389/fdgth.2026.1919498
Keywords: hemodialysis, blood pressure prediction, systolic blood pressure, diastolic blood pressure, random forest, machine learning in nephrology, intradialytic hemodynamics, GroupKFold cross-validation, Hôpital Militaire de Ouakam, chronic kidney disease, ultrafiltration rate, SHAP feature importance, Bland-Altman analysis, Senegal digital health, mean absolute error, dry weight adjustment, resource-limited nephrology, clinical decision support
News Source: Jerry Hayes. (November 13, 2026). Machine Learning Predicts Post-Dialysis Blood Pressure Using Early In-Session Clinical Records. Scienmag.



