Source-linked AI summary
ExpertoRhythm: Morphology-Aware Learning for Waveform Reconstruction and Cuffless Blood Pressure Estimation from Single-Channel PPG
Amir Arjomand, Kenneth B. Kent, Georgiy Krylov
TL;DR
Cuff-based BP measurements are inconvenient for frequent use and do not provide continuous trends, while PPG-to-BP estimation must preserve waveform morphology and pressure-range dynamics. ExpertoRhythm reconstructs ABP from a single PPG channel using a morphology-aware objective. It reports 2.46/1.46 mmHg SBP/DBP MAE and standards-style clinical agreement.
Problem
Cuff-based BP measurements are inconvenient to take frequently and do not provide continuous trends, while pointwise PPG-to-ABP losses may underemphasize local morphology that governs SBP/DBP.
Source: §1 · pp. 1–2Method
Results
2.46/1.46 mmHg SBP/DBP MAE is reported, with low bias and strong agreement in evaluation of predicted-waveform extrema.
Source: §1 · p. 2, §6.2 · p. 6Takeaways & Limitations
The authors conclude that a single PPG channel can support accurate waveform recovery when the training objective is designed to preserve morphology, with the reported evaluation meeting clinical-grade criteria.
Source: §1 · p. 2, §7 · p. 8Wearable deployment remains to be characterized with latency, memory, compute/energy metrics, and deployment-oriented optimization; cross-dataset and wearable-PPG validation is also future work.
Source: §7 · p. 9
Abstract
from arXiv · showhide
Continuous cuffless blood pressure (BP) monitoring from photoplethysmography (PPG) has strong potential for wearable health and telemonitoring, but accurate estimation remains difficult because PPG-to-BP mapping must preserve subtle waveform morphology and pressure-range-dependent dynamics. We introduce ExpertoRhythm, an attention-enhanced 1D U-Net that reconstructs the arterial blood pressure (ABP) waveform from a single-channel PPG signal and derives systolic and diastolic BP directly from the reconstructed waveform. The central contribution is a composite morphology-aware learning objective that integrates range-weighted SmoothL1 reconstruction with a window-range regularizer to emphasize high-dynamic BP segments and reduce amplitude under/over-shoot. On the UCI cuff-less BP dataset with 942 subjects, ExpertoRhythm achieves 2.46/1.46 mmHg MAE for systolic/diastolic BP (SBP/DBP), while obtaining a 30.4% average relative error reduction over pure MSE across waveform reconstruction and BP estimation metrics. Clinical-style evaluation further demonstrates low bias and strong agreement across the BP range, including high-pressure windows up to 200 mmHg, satisfying AAMI criteria and achieving BHS Grade A. These results suggest that morphology-aware waveform reconstruction from a single PPG channel can provide an accurate and practical pathway toward continuous cuffless BP monitoring in wearable and remote-care settings.
1 Introduction
ExpertoRhythm addresses the mismatch between pointwise waveform training and BP readouts determined by local peak and trough morphology. It combines range-weighted SmoothL1 with a window-range regularizer in a CPCA-enhanced 1D U-Net and reports improved BP estimation with standards-style evaluation.
1 Introduction
- Localized peak and trough amplitudes and their immediate shape govern SBP/DBP, so small errors there can matter despite low overall waveform error.The passage contrasts these morphology-dependent readouts with pointwise waveform losses that prioritize average sample fidelity.Source: §1 · p. 2
- 2.46/1.46 mmHg SBP/DBP MAE is reported for ExpertoRhythm on the UCI cuff-less BP dataset, with evaluation meeting AAMI and BHS criteria.The dataset contains 942 subjects, and the reported windows include high-pressure cases up to SBP 200 mmHg.Source: §1 · p. 2
- Range-weighted SmoothL1 and a window-range amplitude regularizer form the proposed morphology-aware objective for waveform-derived SBP/DBP.The objective emphasizes high-dynamic windows and penalizes amplitude under- or over-shoot.Source: §1 · p. 2
- The paper presents a four-layer CPCA 1D U-Net and standards-style evaluation of its agreement under AAMI/BHS criteria.Source: §1 · p. 2
2 Background and Related Work
The paper compares ExpertoRhythm with representative PPG-to-ABP reconstruction approaches and describes a protocol intended to make reported results comparable. The selected methods use single-channel PPG and target continuous waveform reconstruction for clinical SBP/DBP extraction.
2 Background and Related Work
- The comparison set spans convolutional encoder–decoders, a transformer, and morphology-aware or hybrid-objective methods.The cited examples are PPG2ABP [14], ABP-Net, TransfoRhythm [2], and De Palma et al..Source: §2 · p. 2
- Baselines were selected for single-channel PPG input, continuous waveform reconstruction for clinical SBP/DBP extraction, and direct comparability under resource-constrained deployment scenarios.Source: §2 · p. 2
- The comparison uses performance numbers as originally reported and follows authors' evaluation protocols and test-set definitions where available.The stated approach is intended to reduce confounding from data partitions or metric variations.Source: §2 · pp. 2–3
- Figure 1 depicts a four-level CPCA-enhanced 1D U-Net with skip connections for PPG-to-ABP regression.Source: Figure 1 · p. 3
3 Method
The method uses a four-level 1D U-Net with skip connections and CPCA attention to map a single-channel PPG window to a time-aligned ABP estimate. Window-level SBP and DBP are obtained from the predicted waveform's extrema.
3.1 Regression Architecture and I/O
- A four-level 1D U-Net maps channel-first PPG input to a time-aligned ABP estimate, with encoder widths {C, 2C, 4C, 8C}, bottleneck 16C, and a mirrored decoder.The stated base width is C=72; skip concatenation connects encoder and decoder.Source: §3.1 · pp. 3–4
- The encoder downsamples temporal resolution, the bottleneck aggregates context, and the decoder upsamples while concatenating corresponding encoder features to recover detail.Source: §3.1 · p. 3
- Each encoder and decoder stage uses two kernel-size-3 Conv1D layers with ReLU and batch normalization, followed by CPCA attention that reweights channels.The passage describes the channel reweighting as squeeze-and-excitation style.Source: §3.1 · pp. 3–4
- Window-level SBP and DBP are computed as the maximum and minimum, respectively, of the predicted ABP waveform.Source: §3.1 · p. 4
4 Morphology-Aware Loss
ExpertoRhythm combines range-weighted SmoothL1 reconstruction with a window-range regularizer to emphasize high-dynamic windows and limit amplitude under- or over-shoot.
4 Morphology-Aware Loss
- The composite objective combines range-weighted SmoothL1 reconstruction with a window-range regularizer because pointwise MAE/MSE can under-emphasize high-dynamic windows important for waveform amplitude and extrema-derived SBP/DBP.The range-weighted SmoothL1 term averages over time samples within each window; ε stabilizes the weighting, and wmax caps it.Source: §4 · p. 4
- The range regularizer compares predicted and reference amplitudes over a trimmed interval, then penalizes under- or over-shoot beyond margin δ using a smooth gate.The interval trims τ samples from both ends; the gate steepness is controlled by s.Source: §4 · pp. 4–5
- The stated default settings are SmoothL1 β = 1.0, ε = 10^-6, wmax = 5, τ = 10 samples per side, δ = 0.01, and s = 10.Losses use min–max normalized ABP, so δ is in normalized units.Source: §4 · pp. 4–5
5 Data and Training Protocol
The experiments use synchronized PPG and invasive ABP from the UCI dataset, split by subject to prevent leakage, with a fixed held-out test set and cross-validation for model selection.
5 Data and Training Protocol
- The UCI dataset contains 12,000 records from 942 subjects, with synchronized PPG and invasive ABP sampled at 125 Hz and segmented into 10-second windows.Each window contains T = 1250 samples.Source: §5 · p. 5
- Subjects’ record blocks remain indivisible to prevent leakage; the protocol holds out 25% for testing and uses five-fold cross-validation on the remaining 75% for model selection.After hyperparameter selection, the model is retrained on the full 75% pool and evaluated once on the held-out test set.Source: §5 · p. 5
- Preprocessing trims edges, applies a zero-phase 0.1–30 Hz Butterworth bandpass, and separately min–max normalizes PPG and ABP using training-split extrema.Training uses PyTorch, Adam, batch size 256, and early stopping on validation SBP/DBP MAE; λAR = 0.6, wmax = 5, and ε = 10^-6.Source: §5 · p. 5
6 Results
ExpertoRhythm’s results show strong BP accuracy and clinical agreement, while ablations identify a moderate channel width and morphology-aware loss as effective choices. Compared with representative methods, it achieves competitive SBP/DBP MAE using a single raw PPG channel.
6.1 Ablation Study
- C=72 reduces SBP/DBP MAE versus C=32, while C=86 yields only marginal gains at much higher cost.The study selects C=72 as the best accuracy and size trade-off.Source: §6.1 · p. 6, Table 1 · p. 6
- On fixed C=72, range-weighted SmoothL1 outperforms MSE, and adding the range regularizer improves results further, peaking at λAR = 0.6.Higher regularizer values cause slight over-regularization; the reported improvement averages relative error reduction across waveform MAE/RMSE and SBP/DBP MAE.Source: §6.1 · p. 6, Table 2 · p. 6
6.2 Metrics and Clinical Result
- 2.46 mmHg SBP MAE and 1.46 mmHg DBP MAE accompany near-zero bias and strong agreement across 27,260 test windows.Predicted-versus-reference correlations are R = 0.965 for SBP and R = 0.949 for DBP; Bland–Altman analysis reports no evident proportional bias.Source: §6.2 · p. 6, Figure 2 · p. 7, Figure 3 · p. 7
- AAMI mean/SD requirements are met and BHS Grade A is achieved for both SBP and DBP.The reported errors are near-unbiased and within AAMI limits, and BHS thresholds are satisfied across standard tolerance bands.Source: §6.2 · pp. 6–7, Table 4 · p. 8
6.3 Comparison to State-of-the-Art
- 2.46/1.46 mmHg SBP/DBP MAE with single raw PPG improves over UCI baselines and is competitive with recent sequence models.U-Net+BiLSTM reports close errors on 150 records versus ExpertoRhythm’s 12,000; some other methods use richer or derived inputs.Source: §6.3 · p. 7, Table 3 · p. 8
7 Discussion and Conclusion
ExpertoRhythm’s results point to both model capacity and objective alignment as drivers of performance. The authors conclude that single-channel PPG can recover the waveform when the training objective preserves it, while wearable deployment and cross-dataset validation remain future work.
7 Discussion and Conclusion
- At fixed architecture, the morphology-aware objective consistently reduces SBP/DBP error, while width gains saturate beyond a moderate setting.The loss ablation attributes the error reduction to more amplitude-precise waveform reconstruction rather than architectural scaling alone.Source: §7 · p. 8
- The authors conclude that a waveform can be recovered from single-channel PPG when the training objective is designed to preserve it.They report strong accuracy and meeting clinical-grade criteria.Source: §7 · p. 9
- Wearable use remains constrained by the need for low-latency, energy-efficient edge inference.The authors identify deployability as a remaining limitation.Source: §7 · p. 9
- Future work will report efficiency metrics, study deployment-oriented optimization, and validate generalizability on additional datasets.Planned validation includes MIMIC-III, MIMIC-IV, and wearable-acquired PPG datasets using consistent subject-level protocols.Source: §7 · p. 9