Benchmarking and Enhancing PPG-Based Cuffless Blood Pressure Estimation Methods

Cuffless blood pressure screening based on easily acquired photoplethysmography (PPG) signals offers a practical pathway toward scalable cardiovascular health assessment. Despite rapid progress, existing PPG-based blood pressure estimation models have not consistently achieved the established clinical numerical limits such as AAMI/ISO 81060-2, and prior evaluations often lack the rigorous experimental controls necessary for valid clinical assessment. Moreover, the publicly available datasets commonly used are heterogeneous and lack physiologically controlled conditions for fair benchmarking. To enable fair benchmarking under physiologically controlled conditions, we created a standardized benchmarking subset NBPDB comprising 101,453 high-quality PPG segments from 1,103 healthy adults, derived from MIMIC-III and VitalDB. Using this dataset, we systematically benchmarked several state-of-the-art PPG-based models. The results showed that none of the evaluated models met the AAMI/ISO 81060-2 accuracy requirements (mean error $<$ 5 mmHg and standard deviation $<$ 8 mmHg). To improve model accuracy, we modified these models and added patient demographic data such as age, sex, and body mass index as additional inputs. Our modifications consistently improved performance across all models. In particular, the MInception model reduced error by 23% after adding the demographic data and yielded mean absolute errors of 4.75 mmHg (SBP) and 2.90 mmHg (DBP), achieves accuracy comparable to the numerical limits defined by AAMI/ISO accuracy standards. Our results show that existing PPG-based BP estimation models lack clinical practicality under standardized conditions, while incorporating demographic information markedly improves their accuracy and physiological validity.

💡 Research Summary

This paper addresses a critical gap in cuffless blood pressure (BP) estimation research: the lack of standardized, physiologically controlled benchmarking that aligns with clinical accuracy standards such as AAMI/ISO 81060‑2. The authors first construct a new benchmark subset, NBPDB (Normal Blood Pressure Database), by extracting high‑quality photoplethysmography (PPG) segments from two large public ICU repositories—MIMIC‑III and VitalDB. They apply strict inclusion criteria (SBP 90‑130 mmHg, DBP 60‑85 mmHg, age 18‑65 years, BMI 18.5‑25 kg/m²) and a rigorous preprocessing pipeline (artifact removal, 2‑second segmentation, signal normalization) to obtain 101,453 PPG segments from 1,103 healthy adults. The dataset is split into training (≈81 k segments), calibration‑based validation (≈9 k), and calibration‑free validation (≈11 k) to enable evaluation of both personalized and fully cuffless scenarios.

Next, the study systematically re‑implements several state‑of‑the‑art deep learning architectures that have been used for PPG‑only BP estimation: 1‑D ResNet‑18/50, Inception‑1D, LeNet‑1D, and Structured State Space Sequence (S4). All models share a multimodal design: a dedicated 1‑D CNN or sequence encoder processes the raw PPG waveform, while a small multilayer perceptron encodes demographic variables (age, sex, BMI). The two latent representations are concatenated in a late‑fusion layer that outputs systolic (SBP) and diastolic (DBP) pressures simultaneously. Both calibration‑based (subject‑specific reference BP available) and calibration‑free (no reference) settings are evaluated using mean error, standard deviation, and mean absolute error (MAE) as metrics.

The benchmark results reveal that none of the original models meet the AAMI/ISO thresholds (mean error < 5 mmHg, SD < 8 mmHg). In the calibration‑free condition, MAE ranges from 6 to 9 mmHg, indicating substantial performance degradation without a personal reference. Incorporating demographic information consistently improves all models, reducing MAE by 15‑23 % across both settings. The most notable improvement is observed for the modified Inception‑based model (named MInception), which achieves MAE of 4.75 mmHg for SBP and 2.90 mmHg for DBP, thereby approaching the clinical limits. This demonstrates that age, sex, and BMI are powerful contextual modifiers of the PPG‑BP relationship and that their inclusion mitigates inter‑subject variability.

The authors discuss the implications of these findings. They argue that many prior studies overstate clinical readiness because they rely on heterogeneous ICU cohorts and non‑standard metrics, obscuring true generalization performance. NBPDB provides a reproducible, physiologically controlled benchmark that can be shared across the community. The demographic‑aware multimodal architecture not only boosts accuracy but also offers a practical pathway for real‑world wearable devices, where calibration‑free operation is desirable. Limitations include the hospital‑origin of the data (potentially different motion‑artifact profiles than ambulatory settings) and the exclusion of pathological populations (e.g., hypertension, diabetes). Future work should extend validation to diverse cohorts, incorporate motion‑robust preprocessing, and explore lightweight models suitable for on‑device inference.

In conclusion, this work delivers a rigorously curated benchmark dataset, a clear evaluation protocol aligned with medical device standards, and a demonstrable method—demographic‑augmented deep learning—that brings cuffless BP estimation closer to clinical applicability.