Single-Beat Cuffless Blood Pressure Estimation Using Ear-PPG and ECG with a Lightweight Hybrid Learning Framework

AI summary

Overview

• Research area: Cuffless blood pressure estimation from wearable physiological sensing, combining photoplethysmography (PPG), electrocardiography (ECG), and machine learning.
• Technical level: Advanced. The paper assumes familiarity with pulse transit time (PTT), heart rate variability (HRV), convolutional neural networks, and gradient-boosted tree regression.
• Scope: The paper presents a synchronized chest-ECG plus ear-clip PPG wearable platform and a hybrid CNN + LightGBM model that estimates systolic and diastolic blood pressure from individual PPG beats, evaluated on a 10-participant stress protocol and the public PulseDB dataset.

What This Paper Is About

Continuous blood pressure monitoring without a cuff is difficult because motion, posture changes, and shifting autonomic states corrupt the PPG signal, and many existing algorithms need multi-second windows of clean beats to produce a stable estimate. This paper asks whether blood-pressure-relevant information survives at the level of a single heartbeat, and builds a lightweight two-branch model — a small CNN for beat morphology fused with 30 physiology-derived features — to test that question. The goal is accurate, low-cost BP estimation that can plausibly run on a wearable or mobile device rather than a server.

Key Contributions

1. A synchronized multi-modal sensing platform. A chest unit providing single-lead ECG (sampled at 512 Hz) and a 6-axis IMU (104 Hz), plus an ear-clip reflectance PPG module with dual wavelengths (red 740 nm, infrared 850 nm) and two photodiodes at 12 mm and 17 mm from the LED pair, yielding four raw PPG channels at 128 Hz, with its own identical 6-axis IMU. Both units stream via Bluetooth Low Energy to a mobile gateway with measured inter-device synchronization latency under 10 ms.
2. Single-beat blood pressure estimation. The paper argues and shows experimentally that BP-relevant morphological information is preserved at the single-beat level, removing the need for long temporal context and avoiding the failure mode where one corrupted beat invalidates an entire estimation window.
3. A lightweight hybrid learning architecture. A one-dimensional CNN maps a 128-sample normalized PPG beat to a 64-dimensional embedding, which is concatenated with 30 hand-crafted physiological features into a 94-dimensional vector, then passed to LightGBM regressors (separate models for systolic and diastolic BP).
4. A robustness evaluation under physiological stress. The system is tested on a 20-minute, four-phase protocol (rest, exercise, recovery, cold pressor) with a 10-participant cohort, plus subject-disjoint validation on the public PulseDB dataset, and open-source code is released at https://github.com/SYMBIOX-Lab/BP-wireless.

Main Findings

• Systolic and diastolic accuracy over 30 runs: Mean absolute error of 4.02 ± 0.21 mmHg for systolic BP and 1.79 ± 0.05 mmHg for diastolic BP across 30 independent subject-level splits. Each experiment randomly sampled 80 subjects from PulseDB and 7 subjects from the wearable dataset, with subject-disjoint partitioning strictly enforced.
• Additional error statistics: RMSE was 10.79 ± 0.93 mmHg for systolic BP and 2.83 ± 0.06 mmHg for diastolic BP. The MAE ranges across runs were [3.60, 4.42] mmHg for systolic and [1.68, 1.90] mmHg for diastolic. The best-performing run achieved a systolic MAE of 3.60 mmHg and a diastolic MAE of 1.78 mmHg.
• Ablation results: LightGBM with hand-crafted features alone produced a combined MAE (systolic plus diastolic) of 7.49 mmHg (5.49 systolic, 2.00 diastolic). A CNN-only regressor on single-beat PPG waveforms performed far worse, with combined MAE of 77.76 mmHg (71.48 systolic, 6.28 diastolic), a −938.0% change. The hybrid model reached a combined MAE of 5.38 mmHg, a +28.2% improvement over the LightGBM-only baseline. The paper states beat-level waveform morphology alone is insufficient for reliable BP estimation without explicit physiological context.
• Trend tracking across physiological stages: In representative continuous predictions, the model captured stage-dependent BP dynamics — systolic and diastolic elevation during exercise and cold exposure followed by gradual post-exercise recovery — and preserved the expected directionality of BP changes without stage-specific recalibration. Systolic BP showed larger and more consistent excursions across stages than diastolic BP.
• Bland–Altman agreement: Analysis showed small mean bias (for example, −1.62 mmHg for diastolic BP) and relatively narrow 95% limits of agreement, with no evident systematic error trends across the measured BP range. The vertical clustering in the plots is attributed to repeated cuff reference values being compared against multiple beat-level predictions within a short window, not to model instability. A small fraction of extreme outliers was excluded from the plot for readability, but reported bias and limits of agreement were computed on the full test dataset.
• Computational footprint: The hybrid model comprises 152,802 trainable CNN parameters and 12,400 decision nodes in the LightGBM regressors, with an overall footprint below 2 MB. End-to-end inference takes approximately 3 ms per beat on GPU hardware, exceeding 300,000 beats/s throughput.
• Mobile deployment benchmark: The CNN feature extractor was converted from PyTorch to a Core ML–compatible format, occupying only 0.58 MB. Core ML benchmarking on Apple silicon showed a mean inference latency of 0.153 ms per beat (P95: 0.205 ms), peak memory usage of 97.9 MB, and batch throughput exceeding 9,500 beats/s, with the tree-based regressor executed on the CPU.
• Signal quality during exercise: During the exercise phase — the highest motion and autonomic variability — ear-PPG signal quality remained sufficient for beat-level inference, supporting the feasibility of the ear-clip fixation approach with adhesive reinforcement.

Methodology in Plain English

The researchers built two small wearables that talk to each other. A chest patch records the heart's electrical signal (ECG) plus body motion, and a clip on the earlobe shines two wavelengths of near-infrared light through tissue to record the pulse waveform (PPG), also with motion sensing. An earlobe was chosen over the finger or wrist because it is anatomically anchored, stays perfused when the body constricts peripheral vessels, and is less disturbed by tendon motion and hydrostatic pressure effects.

Ten healthy volunteers wore the system through a 20-minute protocol with four 5-minute phases: seated rest, moderate-intensity stationary cycling targeting a heart rate above 100 bpm, post-exercise recovery, and a cold pressor test with the right hand immersed in water at roughly 4 degrees Celsius. Reference blood pressure came from an FDA-cleared clinical oscillometric device (GE Dash 4000) at 60-second intervals, except during cycling, when cuff measurement was omitted. The study was approved under IRB #212305.

For modeling, the team computed the delay between the ECG R-peak and the PPG systolic peak — the pulse transit time — for each beat, and derived heart rate variability measures from R-R intervals, giving 30 hand-crafted features in total. Separately, each PPG beat was resampled to 128 samples and normalized, then fed through a three-block 1D CNN (channels 1 → 16 → 32 → 64, kernels of 7, 5, and 3) to produce a 64-dimensional embedding. The CNN was trained first with Smooth L1 (Huber) loss, the Adam optimizer, an initial learning rate of 1×10⁻³, weight decay of 1×10⁻⁴, and batch size 256, for up to 60 epochs with early stopping on validation MAE. Then the CNN was frozen, its embeddings were concatenated with the 30 physiological features, and LightGBM regressors (max depth 7, 31 leaves, learning rate 0.1, minimum leaf size 20, instance and feature subsampling both 0.8, up to 100 boosting iterations with early stopping) were trained separately for systolic and diastolic BP. A large public dataset, PulseDB — 100 subjects with at least 5 minutes of continuous ECG, PPG, and arterial blood pressure — anchored training and testing, with all signals resampled to 125 Hz. Beats with PTT outside 0.05–0.5 s were discarded.

Why This Matters

Impact on research. The paper challenges the common assumption that BP estimation requires multi-second windows or deep sequence models. It shows that a 94-dimensional fused representation built from a single beat can achieve clinically interpretable accuracy, and that physiology-grounded features (PTT, HRV) and learned morphology are complementary rather than redundant — the CNN-only baseline's large error makes that point sharply. It also contributes a subject-disjoint evaluation across 30 random splits, with the combined cohort spanning 109 subjects.

Real-world applications:

• Ambulatory and at-home monitoring of hypertension, which affects approximately one in three adults and contributes to roughly 12.5% of worldwide deaths, with over 1.28 billion adults aged 30–79 living with the condition and fewer than half achieving adequate control.
• Continuous monitoring during exercise, thermal stress, and other dynamic states where conventional cuff measurement is impractical — the paper notes cuff measurements were omitted during cycling and that consumer-grade home cuffs have shown biases exceeding 10 mmHg.
• Low-power wearable deployment, since the model footprint is below 2 MB and the Core ML extractor occupies 0.58 MB with 0.153 ms mean latency per beat.
• Long-term adherence support, motivated by the paper's observation that nearly 50% of patients fail to reach BP targets due to poor adherence.

Industry relevance. The pipeline is compatible with execution on a mobile gateway such as an iOS tablet for near-real-time inference, with the Core ML conversion and Apple silicon benchmarking making the deployment path concrete rather than hypothetical. The architecture deliberately avoids the power and latency costs of long-sequence deep models, which the authors identify as conflicting with wearable energy and memory constraints.

Future Directions

• Multi-wavelength fusion. The ear-PPG hardware supports four simultaneous optical channels, but this study used a single optimal channel to minimize model complexity. Fusing multiple wavelengths could improve robustness to motion artifacts and wavelength-dependent perfusion changes.
• Larger and more diverse validation. The combined dataset spans 109 subjects, and the authors state that additional validation on larger, more diverse populations is needed. The PulseDB cohort represents clinically monitored hospital patients rather than healthy individuals, and disease annotations such as arrhythmia status were not used for subject stratification, so demographic and clinical differences may affect generalizability.
• Higher sampling rates. All signals were resampled to 125 Hz, which may limit the temporal precision of features such as PTT compared to higher sampling rates.
• Lightweight temporal extensions. The current beat-wise framework does not explicitly model long-term temporal dependencies; the authors suggest lightweight temporal extensions could improve performance during abrupt physiological transitions. They also note the current ear-PPG hardware was adapted from a head-mounted design and is larger than needed for earlobe-specific integration, with future miniaturization and exploration of reflective versus transmissive ear-based PPG.

An additional open question concerns beat windowing: beat-level segments were extracted using a fixed temporal window selected by empirical tuning to balance morphological fidelity and computational efficiency, and deployment-oriented work may refine windowing based on target hardware, latency, and application-specific conditions.

Target Audience

This paper is most useful to biomedical engineers and wearable-hardware developers working on cuffless cardiovascular monitoring, machine learning researchers interested in hybrid feature-engineering-plus-learning architectures under tight resource budgets, and clinical or translational researchers evaluating whether beat-level PPG and ECG signals can support continuous blood pressure tracking outside the clinic. Readers without a background in PTT, HRV, or gradient-boosted trees will need supporting material, since the paper moves quickly between physiological mechanism and model architecture.

Authors’ abstract

Continuous cuffless blood pressure (BP) monitoring remains challenging due to motion artifacts, physiological variability, and the limited robustness of conventional pulse transit time (PTT) models under dynamic conditions. Many prior approaches rely on multi-second windows to stabilize estimation, an assumption that is frequently violated during real-world monitoring with intermittent signal corruption. Here, we show that discriminative BP-related information is preserved at the single-beat level and present a lightweight multi-modal wearable framework for continuous BP estimation. The system integrates synchronized chest electrocardiography (ECG) and ear-clip reflectance photoplethysmography, each co-located with a 6-axis inertial measurement unit to provide motion context. We introduce a hybrid learning architecture in which a one-dimensional convolutional neural network extracts a 64-dimensional embedding from individual PPG beats and fuses it with 30 physiology-grounded features, including PTT statistics and heart rate variability, followed by LightGBM regression. The method was evaluated using a multi-phase stress protocol ($n=10$) and the PulseDB public dataset with subject-disjoint validation. Across 30 independent runs, the model achieved mean absolute errors of $4.02 \pm 0.21$~mmHg for systolic BP and $1.79 \pm 0.05$~mmHg for diastolic BP, corresponding to a 28.2\% reduction in combined MAE relative to baseline models. By enabling beat-wise estimation without long temporal context, this framework supports computationally efficient cuffless BP monitoring suitable for wearable deployment under practical resource constraints. The source code for this work is available at https://github.com/SYMBIOX-Lab/BP-wireless.