Generalizable deep learning for photoplethysmography-based blood pressure estimation-A benchmarking study.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Healthcare
• Cardiovascular Monitoring

Background:

• Photoplethysmography (PPG) offers a cuffless alternative for blood pressure (BP) monitoring.
• Deep learning (DL) models are increasingly used for inferring BP from PPG waveforms.
• Current DL models for BP estimation are often evaluated on limited in-distribution (ID) data, questioning their real-world generalizability.

Purpose of the Study:

• To evaluate the out-of-distribution (OOD) generalization of DL models for PPG-based BP estimation.
• To benchmark DL model performance on the PulseDB dataset and external datasets.
• To investigate methods for improving OOD performance, such as domain adaptation.

Main Methods:

• Five DL models were trained on the PulseDB dataset.
• ID performance was benchmarked on PulseDB.
• OOD performance was assessed on multiple external datasets.
• Sample-based domain adaptation was explored to enhance generalization.

Main Results:

• The best model (XResNet1d101) achieved low ID MAEs (9.0/5.8 mmHg SBP/DBP with calibration).
• OOD MAEs without calibration ranged from 10.0-18.6 mmHg (SBP) and 5.9-10.3 mmHg (DBP) across datasets.
• Significant performance drops were observed on OOD datasets, influenced by BP distribution differences.

Conclusions:

• DL models for PPG-based BP estimation exhibit limited generalizability to datasets with different BP distributions.
• Subject-specific calibration improves ID performance but OOD generalization remains a challenge.
• Domain adaptation techniques show promise for improving model robustness and generalizability.