Real-World Hemodialysis Data Analysis for Intradialytic Hypotension Risk and Patient-Level Phenotyping
Patient-level clustering based on intradialytic instability, blood pressure drop, ultrafiltration rate, and saline use.
Intradialytic hypotension (IDH) is one of the most frequent and clinically significant complications during hemodialysis. It is associated with increased morbidity, impaired dialysis tolerance, and potential long-term cardiovascular consequences.
Traditionally, IDH has been attributed to session-level factors such as ultrafiltration rate (UFR), interdialytic weight gain (IDWG), and anemia. However, these variables often fail to fully explain the variability observed in clinical practice.
To analyze real-world hemodialysis data in order to:
- Identify factors associated with intradialytic hypotension (IDH)
- Evaluate the role of session-level vs patient-level variables
- Develop a simple clinical risk score
- Characterize patient phenotypes using clustering techniques
- IDH rate was high (43%), confirming significant intradialytic instability
- Traditional predictors (UFR, IDWG, hemoglobin) showed limited discrimination
- Patient-level clustering identified 3 distinct hemodynamic phenotypes
- The most unstable phenotype was NOT driven by high UFR
- Suggests IDH is primarily patient-specific, not session-driven
π This challenges the traditional dialysis paradigm and supports a phenotype-based risk stratification approach
This study is based on real-world hemodialysis session data collected from the MARVESA dialysis center in the Dominican Republic.
Data were extracted directly from Nikkiso DBB-06 dialysis machines and include:
- Blood pressure measurements at multiple time points
- Ultrafiltration parameters
- Dialysis session characteristics
- Clinical interventions such as saline administration, UF pause, and early session termination
The dataset includes:
- 394 dialysis sessions
- 52 patients
- Longitudinal patient-level data
- Observational retrospective analysis
- Real-world dialysis machine-derived data
- Session-level and patient-level feature extraction
- Longitudinal aggregation per patient
- Unsupervised clustering for phenotype discovery
- Cleaning and normalization of raw dialysis data
- Handling missing values and inconsistent entries
- Feature engineering:
- Ξ Mean Arterial Pressure (MAP)
- Maximum systolic drop
- IDH binary classification
- Interdialytic weight gain (IDWG)
- Ultrafiltration rate (UFR)
- Descriptive statistics by IDH status
- Comparison of session-level variables:
- UFR
- IDWG
- Hemoglobin
- Analysis of categorical variables:
- Vascular access
- Left ventricular function (FEVI)
- Infection status
A simple clinical risk score was constructed using:
- UFR > 10 ml/kg/h
- IDWG > 3 kg
- Hemoglobin < 10 g/dL
- Catheter access
- Reduced FEVI
- Recent infection
Patient-level aggregation was performed using:
- IDH rate per patient
- Mean Ξ MAP
- Mean systolic drop
- Mean UFR
- Saline use frequency
K-means clustering (k=3) was applied to identify patient phenotypes.
- Dialysis sessions analyzed: 394
- Unique patients: 52
- IDH events: 170
- Overall IDH rate: 43.15%
- Mean UFR: 7.59 ml/kg/h
- Mean IDWG: 1.99 kg
- Mean Hemoglobin: 9.4 g/dL
Classical session-level variables showed limited discriminatory ability between IDH and non-IDH sessions.
- UFR was only slightly higher in IDH sessions
- IDWG showed minimal difference
- Hemoglobin did not meaningfully separate the groups
These findings suggest that traditional session-level variables alone do not fully explain intradialytic instability in this cohort.
Patient-level clustering identified three distinct hemodynamic phenotypes:
- Lower IDH burden
- Mild blood pressure decline
- Moderate ultrafiltration exposure
- Highest average UFR
- No proportional increase in hemodynamic collapse
- Suggests preserved cardiovascular compensation despite fluid removal stress
- Largest drop in mean arterial pressure
- Highest systolic blood pressure decline
- Highest saline rescue use
- Instability not explained by UFR alone
This pattern supports the concept that intradialytic hypotension may reflect patient-specific hemodynamic susceptibility rather than dialysis prescription alone.
In this analysis of real-world hemodialysis data from a dialysis unit in the Dominican Republic, we observed a high prevalence of intradialytic hypotension (IDH), affecting approximately 43% of sessions.
Data were extracted directly from Nikkiso DBB-06 dialysis machines, allowing detailed characterization of intradialytic physiological changes and treatment-related interventions.
Contrary to traditional assumptions, classical variables such as ultrafiltration rate, interdialytic weight gain, and hemoglobin did not meaningfully differentiate IDH and non-IDH sessions in this cohort.
This suggests that IDH is not primarily driven by isolated session-level parameters, but rather reflects patient-specific hemodynamic susceptibility.
To explore this hypothesis, patient-level clustering was performed using longitudinal hemodynamic features. This analysis identified three clinically meaningful phenotypes: a relatively stable group, a high ultrafiltration exposure group without proportional hemodynamic collapse, and a clearly unstable group characterized by larger blood pressure declines and greater dependence on saline rescue.
Importantly, the most unstable phenotype was not explained by the highest ultrafiltration rates, reinforcing the concept that intrinsic patient factors may play a central role in intradialytic instability.
These findings support a shift from a session-centered model of intradialytic hypotension to a patient-centered hemodynamic phenotype model.
From a clinical perspective, this approach may help guide individualized dialysis prescriptions, risk stratification, and closer monitoring of vulnerable patients.
This study is limited by its observational design, single-center setting, and relatively small sample size. However, its strength lies in the use of real-world machine-derived longitudinal dialysis data.
- IDH risk may be patient-specific rather than session-dependent
- Standard variables (UFR, IDWG) are insufficient predictors alone
- Identification of unstable phenotypes may allow:
- Personalized ultrafiltration strategies
- Improved hemodynamic monitoring
- Targeted clinical interventions
- Observational study design
- Single-center data
- Limited sample size
- Potential residual confounding
Intradialytic hypotension appears to be driven more by patient-level hemodynamic phenotypes than by traditional session-level variables.
This approach provides a more clinically meaningful framework for understanding and managing hemodynamic instability in hemodialysis.
.
βββ data/
β βββ raw/
β βββ processed/
βββ scripts/
β βββ 01_cleaning.py
β βββ 02_analysis.py
β βββ 03_visualization.py
β βββ 04_risk_score.py
β βββ 05_clustering.py
βββ results/
β βββ figures/
β βββ tables/
βββ README.md
## Reproducibility
Requirements:
pip install -r requirements.txt
## Execution Order :
-python scripts/01_cleaning.py
-python scripts/02_analysis.py
-python scripts/03_visualization.py
-python scripts/04_risk_score.py
-python scripts/05_clustering.py
## Outputs generated :
Clean dataset: data/processed/hd_clean.csv
Tables: results/tables/
Figures: results/figures/
Risk score dataset: 07_hd_with_risk_score.csv
Patient clustering: 10_patient_clusters.csv
The raw source file is not shared publicly. Processed data and derived outputs are included to support project reproducibility.
Author
Cristian Arias, MD
Nephrologist | Healthcare and Clinical Data Analyst
Bioinformatics MSc