Lightweight Calibration-Based Blood Pressure Tracking from Photoplethysmography Features for Smart Devices: Accuracy, Model Size and Recalibration Interval

Calibration-free PPG regressors were barely better than a constant (SBP MAE 13.0–13.3 vs 14.2 mmHg on unseen subjects). A cuff-anchored delta model reached 11.8 mmHg with one cuff and 10.8 with five, and recalibrating about every 15 min gave 10.3 mmHg on VitalDB. A 47 KB tree model lost only 0.2 mmHg vs a 1.5 MB reference.

• Form factor: finger/clinical PPG (MIMIC-BP ICU, VitalDB surgical) evaluated for a smart-device budget

Background/Objectives: Smart devices estimate blood pressure (BP) from photoplethysmography (PPG), but models trained on other people generalize poorly, and device constraints limit model size. This study compared calibration strategies and model sizes for PPG-based BP estimation. Methods: We used 65 hand-crafted PPG features with MIMIC-BP (1295 training and 229 subject-disjoint test subjects, ten repeats) and 17 external VitalDB surgical cases. We compared calibration-free regressors, calibration with one to five cuffs, recalibration policies and calibration models of decreasing size. Results: Without calibration, six regressors reached a systolic BP (SBP) mean absolute error (MAE) of 13.0–13.3 mmHg on unseen subjects, close to a constant predictor (14.2 mmHg). A delta model that predicts the change from the cuff reading reduced the SBP MAE to 11.8 mmHg with one cuff and 10.8 mmHg with five, but followed only 30–50% of large changes. With five calibration readings, the model achieved a pooled area under the curve (AUC) of 0.94 for new-onset SBP ≥160 mmHg on MIMIC-BP, but sensitivity was zero at an estimated-SBP threshold of 160 mmHg. Accuracy was governed by the recalibration interval: recalibration nominally every 15 min gave 10.3 mmHg on VitalDB, and event-triggered cuffs were not better than periodic ones. A 47 KB tree model lost 0.2 mmHg against a 1.5 MB reference. Conclusions: For these data, calibration was required, the recalibration interval mattered more than model complexity, and a compact tree model was sufficient; on-device use requires further validation. Accuracy remained poor at blood-pressure extremes, and external validation in only 17 VitalDB cases remains preliminary.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: https://creativecommons.org/licenses/by/4.0/

1 row is credited to this paper. Ranks apply within each leaderboard only, as reported by the source.

• MIMIC-BP1,524 ICU subjects × 30 × 30 s ABP/PPG/ECG/RESP (Samsung, ODbL)TrainingEvaluation1,524 subjects · PPG · ECG · BP
• VitalDBintraoperative multi-parameter vitals (PhysioNet mirror)6,388 subjects · PPG · ECG · BP