Biomedical Engineering
Evaluating cross-dataset transfer learning for photoplethysmography-based blood pressure estimation
This study demonstrates that cross-dataset transfer learning can significantly improve blood pressure (BP) estimation from photoplethysmography (PPG) signals, particularly in environments with limited data. By pre-training models on large datasets and fine-tuning with small, subject-specific data, the approach achieves high accuracy (MAE for SBP 3.36 mmHg, DBP 1.81 mmHg), meeting AAMI and BHS standards. This method offers a robust solution for continuous, non-invasive BP monitoring, addressing challenges of data scarcity and domain shifts.
Executive Impact at a Glance
Key quantifiable benefits and advancements from this research, translated into enterprise value.
0%
Improvement in BP Estimation Accuracy
0
Trainable Parameters (Reduced)
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Improved BP Estimation Accuracy with Transfer Learning
Cross-dataset transfer learning achieved mean absolute errors (MAEs) of 3.36 mmHg for systolic BP (SBP) and 1.81 mmHg for diastolic BP (DBP). This represents an approximate 13% improvement over models trained without transfer learning, satisfying AAMI and BHS standards. The performance gap between cross-dataset and intra-dataset transfer learning was statistically insignificant, demonstrating strong generalizability.
Efficient Training with a Compact Model Architecture
A modified BP-CRNN model was developed with fewer than half the trainable parameters of the original model (reduced from ~250,000 to ~105,000). Despite the reduced size and using only half the training data (2.5 hours per patient compared to 5 hours), the model maintained comparable performance, indicating improved data efficiency without degradation.
Cross-Dataset Preprocessing Workflow
The study utilized a comprehensive preprocessing pipeline for both MIMIC-III and VitalDB datasets. This involved delay compensation, 5-second segmentation, autocorrelation filtering for PPG, SAI algorithm filtering for BP, downsampling to 25Hz, normalization, and peak detection to extract SBP/DBP values. This rigorous process ensured data quality and consistency across varied sources.
Comparison of Transfer Learning Approaches
The study compared Non-transfer, Intra-dataset (ID) transfer, and Cross-dataset (CD) transfer learning models. CD transfer models achieved MAEs of 3.36 mmHg (SBP) and 1.81 mmHg (DBP), showing a 13% improvement over non-transfer models and performance comparable to ID transfer learning (within 1% difference). This highlights the robustness of CD transfer learning in leveraging knowledge across different data sources.
Application in Diverse Measurement Environments
The cross-dataset transfer learning approach was evaluated across MIMIC-III (Philips CareVue) and VitalDB (GE Healthcare) datasets, which represent different sensor hardware and recording environments. This demonstrated the method's generalizability and ability to mitigate domain shift, making it suitable for real-world scenarios with varying measurement conditions and devices.
Improved BP Estimation Accuracy with Transfer Learning
Cross-dataset transfer learning achieved mean absolute errors (MAEs) of 3.36 mmHg for systolic BP (SBP) and 1.81 mmHg for diastolic BP (DBP). This represents an approximate 13% improvement over models trained without transfer learning, satisfying AAMI and BHS standards. The performance gap between cross-dataset and intra-dataset transfer learning was statistically insignificant, demonstrating strong generalizability.
13% Improvement in BP Estimation AccuracyEfficient Training with a Compact Model Architecture
A modified BP-CRNN model was developed with fewer than half the trainable parameters of the original model (reduced from ~250,000 to ~105,000). Despite the reduced size and using only half the training data (2.5 hours per patient compared to 5 hours), the model maintained comparable performance, indicating improved data efficiency without degradation.
105,000 Trainable Parameters (Reduced)Cross-Dataset Preprocessing Workflow
The study utilized a comprehensive preprocessing pipeline for both MIMIC-III and VitalDB datasets. This involved delay compensation, 5-second segmentation, autocorrelation filtering for PPG, SAI algorithm filtering for BP, downsampling to 25Hz, normalization, and peak detection to extract SBP/DBP values. This rigorous process ensured data quality and consistency across varied sources.
Raw PPG/BP Signal Input
→
Delay Compensation
→
5-Second Segmentation
→
Signal Filtering & Quality Check (Autocorrelation/SAI)
→
Downsampling & Normalization
→
Peak Detection & SBP/DBP Value Extraction
→
Processed Dataset
| Model Type | SBP MAE (mmHg) | DBP MAE (mmHg) | Key Advantage |
|---|---|---|---|
| Non-Transfer Learning | 3.87 | 2.11 |
|
| Intra-Dataset (ID) Transfer Learning | 3.32 | 1.80 |
|
| Cross-Dataset (CD) Transfer Learning | 3.36 | 1.81 |
|
Application in Diverse Measurement Environments
The cross-dataset transfer learning approach was evaluated across MIMIC-III (Philips CareVue) and VitalDB (GE Healthcare) datasets, which represent different sensor hardware and recording environments. This demonstrated the method's generalizability and ability to mitigate domain shift, making it suitable for real-world scenarios with varying measurement conditions and devices.
Robustness Across Clinical Settings
The transfer learning model was successfully tested on data from two major clinical databases: MIMIC-III, utilizing Philips CareVue systems, and VitalDB, using GE Healthcare devices. This diverse evaluation confirms the model's adaptability to different medical devices and recording protocols. It suggests that knowledge gained from one clinical environment can be effectively transferred to another, enhancing diagnostic consistency.
- MIMIC-III Device: Philips CareVue
- VitalDB Device: GE Healthcare
Advanced ROI Calculator
Estimate the potential savings and reclaimed hours by integrating this AI solution into your operations.
Annual Cost Savings
$0
Annual Hours Reclaimed
0
Your AI Implementation Roadmap
A phased approach to integrate this advanced AI solution seamlessly into your enterprise.
Phase 1: Discovery & Strategy
In-depth analysis of existing infrastructure, data sources, and business objectives. Development of a tailored AI strategy and detailed project plan.
Phase 2: Data Preparation & Model Training
Collection, cleaning, and preprocessing of relevant datasets. Initial model training and validation using identified transfer learning techniques.
Phase 3: Integration & Customization
Seamless integration of the AI model into your current systems. Customization and fine-tuning to align with specific operational workflows and user requirements.
Phase 4: Deployment & Monitoring
Full-scale deployment with continuous monitoring of performance. Ongoing optimization and iterative improvements based on real-world feedback and data.
Phase 5: Scaling & Future Innovations
Expansion of the AI solution to additional departments or use cases. Exploration of advanced features and new research integrations for sustained competitive advantage.
Ready to Transform Your Enterprise?
Schedule a personalized consultation with our AI experts to discuss how this solution can be tailored to your specific business needs and drive measurable impact.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat