Robust Blood Pressure Prediction from Electronic Health Records: A Comprehensive Framework with External Validation and Clinical Utility Assessment

1^st Md Basit Azam
Department of Computer Science & Engineering
Tezpur University
Napaam - 784 028, Tezpur, Assam, INDIA
📧 mdbasit@tezu.ernet.in

2^nd Sarangthem Ibotombi Singh
Department of Computer Science & Engineering
Tezpur University
Napaam - 784 028, Tezpur, Assam, INDIA
📧 sis@tezu.ernet.in

🔬 Abstract

This repository contains a comprehensive machine learning pipeline (mimic_bp_pipelineV3.0.py) for predicting systolic and diastolic blood pressure using readily available clinical variables from Electronic Health Records (EHR). Our methodology addresses critical challenges in clinical ML including data leakage prevention, uncertainty quantification, and external validation across multiple healthcare databases.

Key Contributions:

• 🎯 Rigorous Data Leakage Prevention: Systematic removal of verification features that could lead to unrealistic performance
• 🔀 Advanced Feature Engineering: Clinical interaction features and temporal dynamics analysis
• 📊 Bayesian Optimization: Hyperparameter tuning with uncertainty quantification
• 🏥 Clinical Utility Focus: BHS/AAMI standards compliance and workflow impact simulation
• ✅ External Validation: Cross-database validation between MIMIC-III and eICU databases
• 📈 State-of-the-Art Comparison: Comprehensive benchmarking against existing methods

File Overview: The main implementation is contained in mimic_bp_pipelineV3.0.py - a complete, self-contained pipeline that handles data extraction, feature engineering, model training, evaluation, and clinical validation.

🏆 Performance Highlights

Metric	MIMIC-III (Internal)	eICU (External)	Clinical Standard
SBP RMSE	6.03 mmHg	7.84 mmHg	<8 mmHg (AAMI)
DBP RMSE	7.13 mmHg	9.31 mmHg	<8 mmHg (AAMI)
SBP R²	0.86	0.84	>0.60 (Good)
DBP R²	0.49	0.32	>0.50 (Acceptable)
SBP Within 5mmHg	57.0%	49.7%	>60% (BHS Grade A)
DBP Within 5mmHg	55.7%	43.9%	>60% (BHS Grade A)
SBP Within 10mmHg	91.1%	N/A	>85% (BHS Grade A)
DBP Within 10mmHg	86.7%	N/A	>85% (BHS Grade A)
BHS Grade	B (Both SBP/DBP)	N/A	A (Excellent)
AAMI Standard	✅ PASS	N/A	PASS Required

Key Achievement: 24.7% improvement in SBP RMSE compared to best literature method on MIMIC-III dataset.

🏗️ Methodology Overview

Data Sources

• Primary: MIMIC-III [1] Critical Care Database (n=889 patients)
• External Validation: eICU Collaborative Research Database [2] (n=378 patients)
• Inclusion: ICU patients ≥18 years, >24h stay, valid BP measurements

Feature Categories

• Demographics: Age, gender, ethnicity, BMI (74 final features after alignment)
• Vital Signs: Heart rate, respiratory rate, SpO2, temperature
• Laboratory: Electrolytes, kidney function, cardiac markers, blood gas
• Medications: Antihypertensives, vasopressors, diuretics
• Comorbidities: Hypertension, diabetes, heart failure, CKD
• Temporal: Heart rate variability, BP lability, measurement patterns
• Engineered: Clinical interactions, non-linear transformations

🏗️ Model Architecture Overview

1. Input Layer
74 clinical features extracted from EHR data

2. Preprocessing Pipeline
Variance Threshold → Robust Scaling → Power Transform
(Handles missing values, normalizes distributions, and reduces noise)

3. Core Algorithm
🧠 Gradient Boosting Regressor
(Hyperparameters optimized via Bayesian Search)

4. Output Configuration
→ Multi-Output Prediction: Simultaneous SBP & DBP estimation
→ Uncertainty Quantification: 10th-90th percentile confidence intervals

5. Ensemble Strategy
🧬 Stratified Ensemble Modeling
Separate models trained for:

• Hypertensive range
• Normal range
• Hypotensive range

Validation Strategy

• Internal: 5-fold Group Cross-Validation (patient-wise splitting)
• External: Independent eICU database validation
• Clinical: BHS/AAMI standards, workflow simulation
• Statistical: Bland-Altman analysis, hypothesis testing

🚀 Quick Start

Prerequisites

# Python 3.8+ required
pip install numpy pandas scikit-learn matplotlib seaborn
pip install xgboost scikit-optimize shap google-cloud-bigquery
pip install missingno scipy statsmodels PyYAML

# Google Cloud access for MIMIC-III/eICU (credentials required)
export GOOGLE_CLOUD_PROJECT="your-project-id"

# Download the main pipeline file
wget https://github.com/yourusername/bp-prediction-ehr/raw/main/mimic_bp_pipelineV3.0.py

# Or clone the repository
git clone https://github.com/yourusername/bp-prediction-ehr.git

# Run the complete pipeline with default settings
python mimic_bp_pipelineV3.0.py

# With custom parameters
python mimic_bp_pipelineV3.0.py --log_level DEBUG --output_dir results

# Skip training (evaluation only)
python mimic_bp_pipelineV3.0.py --skip_training --load_models 20250607_125351

⚙️ Configuration Options

The pipeline uses a default configuration defined in the DEFAULT_CONFIG dictionary within the Python implementation. Key parameters include:

DEFAULT_CONFIG = {
    'data_extraction': {
        'project_id': 'd1namo-ecg-processing',  # Your GCP project
        'max_patients': 1000,                   # Maximum patients to extract
        'use_eicu_validation': True             # Enable external validation
    },
    'model': {
        'bayesian_optimization': True,          # Use Bayesian optimization
        'stratified_modeling': True,            # Enable BP range-specific models
        'uncertainty_quantification': True,     # Add prediction intervals
        'use_stacked_ensemble': True           # Ensemble modeling
    },
    'evaluation': {
        'clinical_utility_metrics': True,      # Calculate clinical metrics
        'shap_analysis': True,                 # Feature importance analysis
        'sota_comparison': True                # Compare with literature
    }
}

🛠️ Configuration Flexibility

You can customize the pipeline through:

1. Direct Python edits: Modify DEFAULT_CONFIG dictionary
2. CLI arguments: Override specific parameters via command line
3. External files: Load custom configurations from YAML/JSON

---

📊 Detailed Results & Validation

📈 Model Performance Metrics

Category	Metric	Value Range
Regression	RMSE (SBP/DBP)	6.03 mmHg
	MAE	4.52 mmHg
	R²	0.86
	Pearson Correlation	0.93
Clinical Accuracy	Within 5 mmHg	56.6-69.6%
	Within 10 mmHg	82.4-93.2%
Uncertainty	Interval Coverage	~80%
Classification	Hypertension Sensitivity	92.3%
	Hypotension Specificity	94.1%

🏥 Clinical Validation

✅ BHS Grade B certification for both SBP & DBP (MIMIC-III)
✅ AAMI Compliance:

• Mean difference: <5 mmHg
• Standard deviation: <8 mmHg

⚡ Clinical Impact:

• Potential saving: 424.8 nurse-hours/day
• Risk stratification:
  • Low-risk: 383 patients
  • Medium-risk: 206 patients
  • High-risk: 300 patients

---

🧪 Stratified Performance (MIMIC-III)

SBP Accuracy by Category

• Normal BP: 56.6% within 5mmHg (n=461)
• Prehypertension: 56.3% within 5mmHg (n=302)
• Stage 1 HTN: 63.2% within 5mmHg (n=95)
• Stage 2 HTN: 69.6% within 5mmHg (n=23)

DBP Accuracy by Category

• Hypotension: 60.4% within 5mmHg (n=495)
• Normal: 52.4% within 5mmHg (n=368)

---

📁 Repository Structure

📄 File Descriptions

• README.md: This document containing methodology overview, configuration, and results
• mimic_bp_pipelineV3.0.py: Core implementation of the blood pressure prediction pipeline
• LICENSE: MIT License file granting usage rights

Note: As your project grows, you can expand this structure with directories for data, notebooks, and results.

🛠️ Pipeline Components (within mimic_bp_pipelineV3.0.py)

The single-file implementation contains all necessary components:

1. Data Extractors

• MIMICDataExtractor, eICUDataExtractor
  Classes for BigQuery data extraction from clinical databases

2. Feature Engineering

• EnhancedFeatureProcessor
  Creates clinical feature interactions and derived variables

3. Missing Data Handling

• AdvancedDataImputer
  Supports multiple strategies:
  • MICE (Multivariate Imputation by Chained Equations)
  • KNN imputation
  • Domain-specific fallback values

4. Model Training

• OptimizedBPPredictor
  Implements:
  • Bayesian hyperparameter optimization
  • Gradient boosting with stratified ensembling
  • Multi-output SBP/DBP prediction

5. Evaluation Suite

• ModelEvaluator
  Metrics include:
  • Clinical accuracy thresholds (±5/10/15 mmHg)
  • Uncertainty calibration
  • Risk stratification performance

6. Utility Functions

• Data leakage prevention
• Feature alignment across datasets
• Clinical validation tools (BHS/AAMI standards)

---

🔧 Advanced Usage

Command Line Options

# Full pipeline with custom settings
python mimic_bp_pipelineV3.0.py --log_level INFO --output_dir my_results

# Skip training (use pre-trained models)
python mimic_bp_pipelineV3.0.py --skip_training --load_models 20250607_125351

# Custom timestamp for reproducibility
python mimic_bp_pipelineV3.0.py --timestamp 20250607_125351

Modifying Configuration

Edit the DEFAULT_CONFIG dictionary in the Python file to customize:

# Example: Reduce dataset size for testing
DEFAULT_CONFIG['data_extraction']['max_patients'] = 100

# Example: Disable external validation for faster runs
DEFAULT_CONFIG['data_extraction']['use_eicu_validation'] = False

# Example: Enable enhanced DBP modeling
DEFAULT_CONFIG['model']['enhanced_dbp_modeling'] = True

🔑 Key Classes and Functions

All functionality is contained within the single Python file:

• MIMICDataExtractor
  Extracts patient data from MIMIC-III via BigQuery

• eICUDataExtractor
  Extracts patient data from eICU via BigQuery

• EnhancedFeatureProcessor
  Creates clinical features and interactions (e.g., medication × vitals)

• OptimizedBPPredictor
  Trains ML models with Bayesian optimization and stratified ensembling

• ModelEvaluator
  Evaluates models using clinical standards (BHS grading, AAMI compliance)

• remove_leaky_features()
  Prevents data leakage by removing verification features

• run_eicu_external_validation()
  Performs external validation on eICU dataset

---

🏥 Clinical Significance

📊 Clinical Decision Support Applications

• Risk Stratification: Automatic categorization of patients by BP stability
• Monitoring Optimization: Reduced manual measurements for low-risk patients
• Early Warning: Detection of BP changes 30 minutes earlier than manual methods
• Resource Allocation: Optimized nurse workflow with 424.8 hours saved per day potential

📏 Regulatory Compliance

• AAMI Standards: Mean difference <5 mmHg, SD <8 mmHg ✅ ACHIEVED
• BHS Grading: Grade B achieved (target Grade A: ≥60% within 5 mmHg)
• FDA Guidance: Uncertainty quantification and external validation implemented
• ISO Standards: Clinical evaluation and risk management protocols followed

💡 Clinical Utility Metrics

• Hypotension Detection: 100% specificity, 99.1% NPV
• Hypertension Detection: 100% sensitivity and specificity
• Prediction Intervals: 80.3% coverage with 22.89 mmHg average width

---

📈 Key Findings

• Data Leakage Prevention: Removal of verification features maintains clinical utility while preventing overfitting
• Feature Importance: Heart rate variability, medication interactions, and temporal patterns are key predictors
• Generalizability: External validation shows 30% RMSE increase (SBP) and 31% (DBP) - indicating reasonable but limited generalizability
• Clinical Impact: Potential for 35% reduction in manual BP measurements while maintaining safety
• Uncertainty Quantification: Reliable prediction intervals enable risk-based clinical decisions (80% coverage achieved)
• Stratified Performance: Better performance in hypertensive patients vs hypotensive patients

---

🔬 Reproducibility

📁 Data Access Requirements

• MIMIC-III: Requires PhysioNet credentialed access and approved research project
• eICU: Separate PhysioNet credentialing required
• Google Cloud: BigQuery access configured with appropriate project permissions

Environment Setup

# Create isolated environment
conda create -n bp-prediction python=3.8
conda activate bp-prediction

# Install required packages
pip install numpy pandas scikit-learn matplotlib seaborn xgboost scikit-optimize shap google-cloud-bigquery missingno scipy statsmodels PyYAML

# Set up Google Cloud credentials
export GOOGLE_APPLICATION_CREDENTIALS="path/to/your/credentials.json"
export GOOGLE_CLOUD_PROJECT="your-project-id"

Running Experiments

# Download the pipeline
wget https://raw.githubusercontent.com/yourusername/bp-prediction-ehr/main/mimic_bp_pipelineV3.0.py

# Run with default settings (requires MIMIC-III access)
python mimic_bp_pipelineV3.0.py

# Run with custom parameters
python mimic_bp_pipelineV3.0.py --log_level DEBUG --output_dir my_experiment

# Reproduce specific experiment
python mimic_bp_pipelineV3.0.py --timestamp 20250607_125351

📤 Output Files

The pipeline generates timestamped output files to ensure reproducibility and traceability:

• bp_prediction_YYYYMMDD_HHMMSS.log
  Detailed execution log including warnings, errors, and progress tracking

• feature_importance_YYYYMMDD_HHMMSS.csv
  Ranked list of clinical features with SHAP-based importance scores

• MIMIC-III_evaluation_YYYYMMDD_HHMMSS.json
  Internal validation metrics (RMSE, R², clinical thresholds)

• eICU_evaluation_YYYYMMDD_HHMMSS.json
  External validation results with stratified performance metrics

• figures_YYYYMMDD_HHMMSS/
  Directory containing:

  • Calibration plots
  • Prediction interval visualizations
  • Feature importance charts
  • Bland-Altman agreement plots

• models/
  Trained model artifacts for reuse:

  • Gradient boosting ensemble
  • Imputation strategy pickles
  • Feature processor configurations

• imputation_stats_*_YYYYMMDD_HHMMSS.csv
  Comparative analysis of MICE vs KNN imputation performance

• *_stratified_results_*_YYYYMMDD_HHMMSS.csv
  Performance breakdown by BP category (hypotension/normal/hypertension)

Note: Timestamps (YYYYMMDD_HHMMSS) ensure unique filenames across runs while maintaining traceability to specific experiments.

💡 Tip: Use --output_dir parameter to customize the output directory location for organized storage across multiple experiments.

📚 Citation

If you use this work, please cite our paper:

@misc{azam2025clinicalgradebloodpressureprediction,
      title={Clinical-Grade Blood Pressure Prediction in ICU Settings: An Ensemble Framework with Uncertainty Quantification and Cross-Institutional Validation}, 
      author={Md Basit Azam and Sarangthem Ibotombi Singh},
      year={2025},
      eprint={2507.19530},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2507.19530}, 
}

🤝 Contributing

We welcome contributions! Please see our Contributing Guidelines for details.

Steps to contribute:

1. Fork the repository
2. Create a feature branch

   git checkout -b feature/amazing-feature
   ##Commit changes
   git commit -m 'Add amazing feature'
   ##Push to branch
   git push origin feature/amazing-feature

3. Open a Pull Request

🔄 Updates

• v3.0: Enhanced pipeline with external validation and clinical utility metrics
• v2.1: Added uncertainty quantification and stratified modeling
• v2.0: Implemented Bayesian optimization and SHAP analysis
• v1.0: Initial release with basic BP prediction pipeline

📄 License

This project uses an MIT License. See the LICENSE file for details.

🙏 Acknowledgments

The authors acknowledge support from the Google Cloud Research Credits program under Award GCP19980904 and partial compute resources from Google’s TPU Research Cloud (TRC), both of which provided critical infrastructure for this research.

Funding:

The authors declare no funding was received for this research.

References

[1] Johnson, A., Pollard, T., & Mark, R. (2016). MIMIC-III Clinical Database (version 1.4). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/C2XW26

[2] Pollard, T., Johnson, A., Raffa, J., Celi, L. A., Badawi, O., & Mark, R. (2019). eICU Collaborative Research Database (version 2.0). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/C2WM1R