TransfoRhythm: A Transformer Architecture Conductive to Blood Pressure Estimation via Solo PPG Signal Capturing
Authors: Amir Arjomand, Amin Boudesh, Farnoush Bayatmakou, Kenneth B. Kent, Arash Mohammadi
Organizations: Faculty of Computer Science, University of New Brunswick, Fredericton, New Brunswick, Canada · Concordia Institute for Information Systems Engineering, Concordia University, Montreal, Quebec, Canada
Recent statistics indicate that approximately 1.3 billion individuals worldwide suffer from hypertension, a leading cause of premature death globally. Blood Pressure (BP) serves as a critical health indicator for accurate and timely diagnosis and/or treatment of hypertension. Traditional BP measurement methods rely on cuff-based approaches, which lack real-time, continuous, and reliable BP estimates, crucial for the timely diagnosis/treatment of hypertension. Driven by recent advancements in Artificial Intelligence (AI) and Deep Neural Networks (DNNs), there has been a surge of interest in developing data-driven and cuff-less BP estimation solutions. In this context, current literature predominantly focuses on coupling Electrocardiography (ECG) and Photoplethysmography (PPG) sensors, though this approach is constrained by reliance on multiple sensor types. An alternative, utilizing standalone PPG signals, presents challenges due to the absence of auxiliary sensors (ECG), requiring the use of morphological features while addressing motion artifacts and high-frequency noise. To address these issues, the paper introduces the TransfoRhythm framework, a Transformer-based DNN architecture built upon the recently released physiological database, MIMIC-IV. Leveraging the Multi-Head Attention (MHA) mechanism, TransfoRhythm identifies dependencies and similarities across data segments, forming a robust framework for cuff-less BP estimation solely using PPG signals. To our knowledge, this paper represents the first study to apply the MIMIC IV dataset for cuff-less BP estimation. TransfoRhythm achieves highly accurate results with a Root Mean Square Error (RMSE) of [2.21, 1.84] and a Mean Absolute Error (MAE) of [1.37, 1.06] for systolic and diastolic blood pressures, respectively.
Figures & tables
Figure 1: This block diagram depicts the streamlined process for estimating systolic and diastolic blood pressures. Features are extracted from the dataset (top left), combined with positional encoding, and form an embedded input matrix. This matrix is then inputted into the model block, which integrates a multi-head attention mechanism. The output of the model is further refined in the position-wise block. Finally, the Time compressor and flattening blocks are applied to reduce the time frame and yield the estimated systolic and diastolic blood pressure values.
Figure 2: Time distribution of records depicted in the graph. The x−axis represents the record’s length, ranging from a few minutes to a maximum of 17 hours. The darker shades of blue indicate a higher number of observations. The majority of patients have records with a time duration of less than 1.5 hours.
Figure 3: Distribution of patient data based on SBP and DBP values. The x−axis represents the blood pressure value, while the y−axis depicts the data frequency for systolic blood pressure (left) and diastolic blood pressure (right). Each bar corresponds to the number of records of corresponding pressure. The density of the bars reveals the concentration of observations, highlighting that most patients fall within the normal range.
Figure 4: A comparison of the signal before and after preprocessing. The x-axis represents time (s), while the y-axis denotes signal amplitude. The figure demonstrates the effectiveness of the preprocessing stage in suppressing noise and producing a cleaner and smoother signal, while accurately preserving the peaks and feet of the waveform.
ID
Feature
Description
F1
PPG_Integration
Area under one PPG cycle.
F2
PPG_Pulse_Amplitude
Amplitude difference between the PPG peak and cycle start.
F3
SDPPG_Foot_Peak
Duration between the SDPPG foot and SDPPG peak.
F4
PPG_Rise_Peak
Duration from the start of the PPG cycle to the systolic peak.
F5
FDPPG_Positive_Peak
Maximum positive amplitude of the first-derivative PPG signal.
F6
PPG_Fall_Half
Duration from the PPG peak to the half-amplitude point in the falling phase.
Table 2: Patient-independent fold composition, observed hypertension and hypotension counts, and fold-specific feature selection after collinearity filtering.
Parameter
Value / criterion
Minimum record duration
Records shorter than 15 min were excluded.
PPG normalization and valid range
Each retained PPG recording was independently normalized before segmentation using zi=(xi−μrecord)/σrecord , producing a signal with zero mean and unit standard deviation. The valid amplitude range was −5 to +5 standard-deviation units.
Out-of-range rejection
After normalization and before filtering, a 10-s window was rejected when more than 10% of its samples were outside the range −5 to +5 , indicating excessive large-amplitude deviations or severe artifacts.
Flat-line rejection
A window was rejected if any sliding interval of 125 samples (approximately 2.00 s at 62.4 Hz) had a standard deviation below 0.02 or a peak-to-peak amplitude below 0.05 normalized units.
Bandpass filtering
A fifth-order Butterworth bandpass filter with cutoff frequencies of 0.7 and 10 Hz was applied using zero-phase forward–backward filtering.
Moving-average smoothing
A three-sample moving average filter, corresponding to approximately 0.048 s at 62.4 Hz, was applied after bandpass filtering.
Table 3: Preprocessing, signal-quality, cycle-detection, and label-extraction thresholds.
T
MAE SBP
MAE DBP
RMSE SBP
RMSE DBP
GFLOPs
Lat. (ms)
Params (M)
Size (MB)
24
1.61
1.19
2.49
1.98
1.78
17.6
26.84
102.39
32
1.46
1.11
2.34
1.91
2.17
20.9
26.85
102.43
48
1.37
1.06
2.21
1.84
2.43
24.3
26.87
102.51
64
1.42
1.10
2.29
1.88
3.06
30.4
26.88
102.56
Table 4: Ablation study on transformer input sequence length T .
Figure 5: Characterization of features on the signal. In the PPG signal, the peak is defined as the maximum value of a PPG wave, while d.notch is defined as the minimum value between two consecutive peaks. In the SD PPG signal, amp peak refers to the maximum amplitude, while amp foot refers to the minimum amplitude between two consecutive waves.
Fold
Test patients
Test samples
SBP
DBP
MAE
RMSE
ME
R-square
MAE
RMSE
ME
R-square
Fold 1
40
3,850
1.73
3.44
+0.11
0.874
1.64
2.55
−0.21
0.881
Fold 2
40
4,120
1.90
2.58
+0.86
0.918
1.40
2.44
−0.38
0.887
Fold 3
40
3,940
2.08
3.32
−0.77
0.871
1.30
2.80
−0.36
0.872
Fold 4
39
4,050
1.86
3.16
−0.38
0.890
1.40
2.71
−0.23
0.879
Fold 5
39
4,040
2.08
3.07
+0.14
0.887
1.72
2.46
−0.12
0.936
Table 5: Patient-independent 5-fold cross-validation performance of TransfoRhythm.
Hyperparameters
Type and Value
Dropout
0.15
Optimization Algorithm
Adam
Learning Rate
1E-4 (with dynamic wight decay)
Activation Function
ReLU
Loss Function
MSE
Batch Size
128
Table 6: Values of the considered hyperparameters
Figure 6: Distributions of prediction errors for systolic (SBP) and diastolic (DBP) blood pressure measurements. The error histograms are centered around zero, indicating balanced model performance for SBP and DBP.
Figure 7: True versus predicted blood pressure values for systolic blood pressure (left) and diastolic blood pressure (right). The data points (navy color) indicate model predictions, while the golden line represents the ideal correspondence between true and predicted values.
Figure 8: Residual plots for systolic (SBP) and diastolic (DBP) blood pressure predictions. The dashed red line represents zero residuals, emphasizing a balanced error distribution around this reference.
Records
ME (mmHg)
STD (mmHg)
Cumulative error percentage
<5 mmHg
<10 mmHg
<15 mmHg
TransfoRhythm Architecture
SBP
360
-0.01
3.1
93.28%
96.83%
98.63%
DBP
360
-0.26
2.57
94.20%
96.14%
97.18%
AAMI Standard Requirement
>85
<5
<8
-
-
-
BHS Standard Requirement
Grade A
-
-
-
60%
85%
95%
Grade B
-
-
-
50%
75%
90%
Table 7: Evaluation of the Proposed Model with AAMI and BHS Standards.
Figure 9: Bland-Altman plots for assessing agreement between ground truth and predicted values of SBP and DBP.
Arch.
Model
MAE
MAE
RMSE
RMSE
R2
R2
Params
GFLOPs
Latency
Size
SBP
DBP
SBP
DBP
SBP
DBP
(M)
(ms)
(MB)
CNN
ResNet_1D
2.28
1.06
5.55
3.21
0.962
0.985
3.85
0.86
11.00
14.68
CNN
AlexNet_1D
3.09
1.59
7.98
4.60
0.922
0.969
2.35
0.13
7.30
8.99
CNN
U-Net
2.76
1.56
7.47
4.70
0.931
0.967
68.07
1.532
48.50
259.66
RNN
Bi-LSTM
3.86
1.90
7.98
4.19
0.922
0.974
2.56
0.231
17.80
9.76
Hybrid
CNN-RNN Network
2.79
1.96
5.62
3.91
0.961
0.977
13.28
0.319
19.36
50.65
Table 8: Comparison with benchmark models on the MIMIC-IV dataset.
Model Architecture
Signal
Datasets
Input Type
MAE SBP
MAE DBP
RMSE SBP
RMSE DBP
STD SBP
STD DBP
ME SBP
ME DBP
MLPlstm-BP [ Huang et al., 2022 ]
ECG PPG
MIMIC II
Raw Waveform
3.52
2.13
5.1
3.13
5.09
3.07
0.071
-0.495
Temporal Convolution Network (TCN) [ Zabihi et al., 2022 ]
ECG PPG
MIMIC I MIMIC III
Feature Extraction
2.59
1.33
3.03
1.58
3.58
1.97
-2.98
-3.65
CNN + LSTM [ Baker et al., 2021 ]
ECG PPG
MIMIC III
Raw Waveform
4.41
2.91
NR
NR
6.11
4.23
NR
NR
NABNet (BiConvLSTM) [ Mahmud et al., 2023 ]
ECG PPG
MIMIC III
Raw Waveform (multi-channel)
2.63
1.09
NR
NR
2.96
1.05
-0.678
-0.342
Bi-LSTM [ Chen et al., 2023 ]
ECG PPG
MIMIC II
Feature Extraction
2.51
1.38
3.22
1.78
NR
NR
-0.09
-0.07
LSTM network [ Guo et al., 2024 ]
ECG PPG
MIMIC III
Raw Waveform
8.34
3.47
NA
NA
10.68
4.76
NA
NA
Table 9: Comparison of studies in BP estimation in terms of methodology, database, features, and metrics (NR means not reported).
Figure 10: Comparison of SBP and DBP R2 across methods.
Blood pressure (BP) is a key marker for cardiovascular risk assessment and therapeutic decision-making, and Photoplethysmography (PPG) enables low-cost, wearable-friendly cuffless BP estimation. However, even with recent progress, many PPG-based models are trained with BP regression alone and may rely on amplitude-dominated shortcuts. In addition, demographic covariates that systematically modulate vascular compliance are often incorporated only via late fusion, limiting subject-specific representation learning. We propose a Transformer-based network for cuffless BP estimation from PPG signal, leveraging self-attention to capture long-range dependencies across multiple cardiac cycles. To account for subject-specific vascular differences, the model is conditioned on demographics via FiLM-style feature modulation applied through the attention and feed-forward sublayers of Transformer blocks. In addition, we add an auxiliary morphology head to guide the model to attend to BP-relevant waveform morphology associated with arterial stiffness and wave reflection. Under calibration-based evaluation protocols on the large-scale PulseDB dataset, the proposed method achieves MAE of 4.56 mmHg for systolic BP and 2.62 mmHg for diastolic BP, reducing errors by 47% and 50% compared with prior demographic-enhanced PPG baselines. The resulting lightweight, single-sensor model supports scalable and clinically grounded cuffless BP estimation in calibration-enabled deployment settings.
Yidan Shen, Neville Mathew, Maham Rahimi +4
Dept. of Electrical and Computer Engineering University of Houston Houston, USA · Dept. of Engineering Technology University of Houston Houston, USA · Dept. of Cardiovascular Surgery Houston Methodist Hospital Houston, USA +1
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±0.21~mmHg for systolic BP and 1.79±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.
Kindeep K. Dhatt, Tengyue Wu, Hanbang Hua +1
1. Vanderbilt Institute for Surgery and Engineering, Vanderbilt University, Nashville, TN, 37232 · 6. Department of Biomedical Engineering, Vanderbilt University, Nashville, TN, 37235 · 2. Department of Electrical and Computer Engineering, Vanderbilt University, Nashville, TN, 37235 +3
Objective. To develop and evaluate a cuffless continuous blood pressure (BP) estimator using temporal physiological and demographic features. We propose a hybrid Transformer framework to estimate diastolic and systolic BP from ECG/PPG-derived feature sequences. Approach. Rather than raw waveforms, the framework models 10-step sequences of six physiological descriptors and two demographic covariates. A Multi-Source Temporal Encoder Module combines Transformer, Kolmogorov-Arnold Network, and XGBoost branches to capture complementary temporal, nonlinear, and tabular information. A Dynamic Conditional Fusion-Decoder applies differential multi-head attention, token-weighted aggregation, and gated residual correction. A robust composite objective jointly optimizes DBP and SBP. Main results. Using the MIMIC-III Waveform and Clinical Databases, the source pool comprised 28,486 waveform segments from 203 subjects, and feature generation retained 53,621 observations from 166 subjects. On 2,431 segment-level held-out test windows, mean error +/- standard deviation was 0.41 +/- 3.74 mmHg for diastolic BP and -1.60 +/- 5.95 mmHg for systolic BP, with 95% limits of agreement of [-6.93, 7.74] and [-13.25, 10.06] mmHg, respectively. The proportions within 10 mmHg were 98.48% and 94.36%. The framework achieved the lowest standard deviations and narrowest limits of agreement among the locally retrained baselines. Significance. The feature-sequence fusion framework improved agreement with reference BP and fell within numerical AAMI and BHS Grade A thresholds on this split. This retrospective analysis is not formal device validation; subject-disjoint and external evaluation remain necessary before clinical use.
Yuexin Ma, Jingqi Hou, Yuxuan Kang +1
College of Computer Science, Beijing University of Technology, Beijing, China