TransfoRhythm: A Transformer Architecture Conductive to Blood Pressure Estimation via Solo PPG Signal Capturing
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: I'm Rosa.
Dev: And I'm Dev, here with Taro.
Rosa: We're looking at the paper: "TransfoRhythm: A Transformer Architecture Conductive to Blood Pressure Estimation via Solo PPG Signal Capturing".
Dev: Here's the gist of the paper. TransfoRhythm introduces a Transformer-based Deep Neural Network architecture to estimate blood pressure using only solo PPG signals from the MIMIC-IV dataset, representing a novel approach to cuff-less BP estimation The paper demonstrates that this framework achieves highly accurate results with an RMSE of
2.21, 1.84: and an MAE of
1.37, 1.06: for systolic and diastolic blood pressures, respectively <ref:2404.15352#pg3>.
Rosa: Problem Context Blood pressure serves as a critical health indicator for accurate and timely diagnosis and/or treatment of hypertension <ref:2404.15352#pg2> 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 <ref:2404.15352#pg2> Alternative methods of BP measurement, such as cuff-less monitoring, have been explored to provide continuous monitoring and reduce the potential for measurement errors caused by cuff-based methods <ref:2404.15352#pg3> Cuffless BP measurement offers the potential for continuous monitoring which enables long-term hypertension management and facilitates the precise diagnosis of cardiovascular conditions <ref:2404.15352#pg3> Nocturnal BP provides critical information for the risk assessment of cardiovascular diseases CVD, as abnormally high surges or deep dips in BP during the night are strongly linked to severe health outcomes including an increased possibility of stroke, heart failure, and even mortality <ref:2404.15352#pg3> Data Preparation and Feature Engineering The study utilizes the Medical Information Mart for Intensive Care (MIMIC)-IV Waveform Dataset available on Physio-Net <ref:2404.15352#pg7> The recording process was conducted at a sampling rate of 62.4 Hz <ref:2404.15352#pg7> In the preprocessing phase, data cleaning and signal filtering were performed, including removing records with frequent outrange amplitudes or those not meeting minimum time duration requirements <ref:2404.15352#pg8> A 5th order Butterworth filter in the frequency range of 0.7 to 10 Hz and a 5th order Moving Average Filter (MAF) were applied to remove overshoots and fluctuations <ref:2404.15352#pg8> The feature extraction process begins with identifying and isolating cycles within each frame of the PPG signal by locating peak and trough points in both the PPG signal and its second derivative <ref:2404.15352#pg9> From the extracted cycles, a set of 12 features was selected based on their relationships derived from correlation analysis <ref:2404.15352#pg10> These 12 features include metrics such as PPG Cycle Duration TD1, PPG Rise Half Peak Trhp, and PPG Integration PPGI <ref:2404.15352#pg11> Model Architecture Design The proposed TransfoRhythm framework is built on the regressive time series transformer network designed to learn from time series data while maintaining sequence order and enabling parallel computation <ref:2404.15352#pg10> To expand the dimensionality of the feature vectors, an embedding feature similar to those used in Natural Language Processing (NLP) tasks is employed <ref:2404.15352#pg10> Positional encoding is then added to the extracted features to create the input data matrix which is fed into a model block via the Multi-Head Attention MHA mechanism <ref:2404.15352#pg12> The architecture incorporates a multi-head attention mechanism with 14 heads for BP estimation, and a linear layer in the last section of multi-head attention to introduce non-linearity <ref:2404.15352#pg13> Finally, a Time Frame Compressor unit is employed to reduce dimensionality while preserving temporal signals before the final flattened layer transforms the signals to systolic and diastolic blood pressure values <ref:2404.15352#pg13> Performance Evaluation and Comparison The model achieved high predictive accuracy with mean R2 values of 0.993 for SBP and 0.994 for DBP <ref:2404.15352#pg14> The Mean Absolute Error (MAE) was found to be 1.50 mmHg for SBP and 1.17 mmHg for DBP <ref:2404.15352#pg14> These results demonstrate that the proposed TransfoRhythm framework can reliably estimate blood pressure across different data subsets, as validated through cross-validation <ref:2404.15352#pg14> When compared against benchmark models on the MIMIC IV dataset, the TransfoRhythm outperforms all benchmark models, a result largely attributed to its internal MHA mechanism <ref:2404.15352#pg19> The model successfully met the AAMI standard and achieved “Grade A” in the BHS standard <ref:2404.15352#pg16> The Bland-Altman plot for SBP demonstrated a mean difference of 0.26 mmHg with 95% limits of agreement ranging from −3.60 mmHg to 4.12 mmHg <ref:2404.15352#pg18> The results demonstrate the superiority of the proposed TransfoRhythm framework for BP estimation, achieving MAE and STD values in the range of
1.50, 1.84: mmHg for DBP and
1.17, 1.42: mmHg for SBP <ref:2404.15352#pg21> The consistently low MAE and RMSE values highlight the model’s potential for accurate and reliable BP estimation in real-world applications <ref:2404.15352#pg21> Future work could investigate the model’s performance on additional datasets or incorporate diverse data sources to improve generalization <ref:2404.15352#pg24> Exploring more advanced attention-based network architectures could further enhance model accuracy and robustness <ref:2404.15352#pg24> The study represents the first attempt to develop a BP estimation model solely based on PPG signals from MIMICIV <ref:2404.
Title and authors: Rosa: We’re now looking at the paper titled TransfoRhythm: A Transformer Architecture Conductive to Blood Pressure Estimation via Solo PPG Signal Capturing by Amir Arjomanda, Amin Boudeshb, Farnoush Bayatmakoub, Georgiy Krylova, Kenneth B. Kenta, and Arash Mohammadib.
Dev: That’s a solid team of researchers working on this problem in the field of biomedical signal processing and deep learning applications. The title clearly points to using a Transformer architecture for blood pressure estimation from solo pulse oximetry data.
Taro: I always like seeing papers that focus on cuff-less measurement because that's where the real clinical impact is, moving away from those cumbersome, traditional methods.
Rosa: It’s about taking what we already have—the PPG signal—and using advanced AI to unlock a new way to get blood pressure estimates in real time or continuously.
Dev: The authors are tackling a massive problem because blood pressure is a critical indicator for hypertension, and they’re trying to find a non-invasive solution that doesn't require constant cuff application.
Taro: They are moving away from relying on just one type of signal like ECG or other complex inputs and focusing purely on the optical technique of PPG.
Rosa: That shift in focus is significant because it means you only need a single sensor placed on the skin to get these useful measurements, which makes deployment much more flexible.
Dev: It’s about making continuous monitoring possible, which allows for long-term management of hypertension and better tracking of cardiovascular conditions like stroke or heart failure.
Taro: So they are aiming to provide a tool that is less intrusive for patients, which is a key consideration in clinical adoption.
The paper's summary: Rosa: The core idea here is that the TransfoRhythm framework uses a regressive time series transformer network to learn from the sequence of solo PPG signals to estimate blood pressure.
Dev: To unpack that, they take the raw signal, clean it with filters, extract twelve specific features from those cycles, and then use those features as input for an embedding feature similar to NLP tasks.
Taro: And they add positional encoding to create an input data matrix before feeding it into a model block utilizing a multi-head attention mechanism.
Rosa: That multi-head attention mechanism with fourteen heads is the core engine that allows the network to weigh different temporal parts of the signal against each other effectively.
Dev: They then introduce a linear layer right in the last section of that multi-head attention structure, which introduces some non-linearity before everything gets compressed.
Taro: After that, they use a Time Frame Compressor unit to reduce the dimensionality while trying to preserve those important temporal signals before transforming them into systolic and diastolic blood pressure values.
Rosa: So in essence, it’s a complex pipeline: signal processing, feature extraction, embedding with positional encoding, attention across time steps, dimensionality reduction by the compressor unit.
Dev: And they claim this entire system achieves high accuracy on the MIMIC-IV dataset with an RMSE of two point two one for systolic and one point eight four for diastolic blood pressure.
Taro: That accuracy metric is what we need to focus on, because it shows that even with all those complex layers, the final output is still clinically relevant.
The paper's improvements: Rosa: The authors suggest a few key enhancements, starting with leveraging the Multi-Head Attention mechanism more effectively for better BP estimation accuracy.
Dev: They specifically highlight that improving the performance of that multi-head attention mechanism can provide highly accurate results, achieving a Root Mean Square Error of one point eight four for diastolic blood pressure, meeting AAMI and BHS standards as shown in Table three.
Taro: That’s good because it suggests there’s room for further tuning, which means the model isn't just a static solution but something you can keep improving.
Rosa: They also suggest that the framework can be superior when compared against benchmark models on the MIMIC IV dataset, largely attributed to its internal multi-head attention mechanism.
Dev: They claim this superiority over models like ResNet1D, Hybrid CNN-RNN Network, and U-Net by achieving a lower MAE of one point zero six for systolic blood pressure and two point two one for diastolic blood pressure on Table four.
Taro: So the claim is that their specific attention mechanism is what gives them that edge over those other established architectures in this comparison.
Rosa: They also mention that their ability to handle variable-length inputs without needing reduction methods allows it to learn more accurate representations of the input compared to conventional deep learning methods that often require padding or truncation.
Dev: That’s a good point because if you can handle variable lengths naturally, you avoid losing information upfront, which is a common pitfall in sequence modeling.
Conclusion: Rosa: So to wrap up, the paper on TransfoRhythm shows that this Transformer architecture for solo PPG signals can achieve highly accurate blood pressure estimation.
Dev: They showed that the model can reliably estimate both systolic and diastolic blood pressure with an RMSE values of two point two one and one point eight four on the MIMIC-IV dataset.
Taro: The system is designed to provide continuous monitoring capabilities, which is crucial for detecting those nocturnal BP patterns linked to severe health outcomes like stroke or heart failure.
Rosa: So, in simple terms, this means we have a framework that can provide reliable blood pressure estimates using only solo PPG data.
Dev: It’s a significant step forward because it shows the power of sequence models when applied to this type of signal for clinical use without needing external hardware.
Taro: And the next step is figuring out how to make this robust enough to handle real-world noise and misbehaving world scenarios gracefully.
Rosa: So, that’s our look at TransfoRhythm for today, moving on to whatever new arXiv papers we have next.
Faculty of Computer Science, University of New Brunswick, Fredericton, New Brunswick, Canada · Concordia Institute for Information Systems Engineering, Concordia University, Montreal, Quebec, Canada
eess.SP, cs.LG
Submitted: 2024-04-15
Updated: 2026-11-02
Journal ref: A. Arjomand, et al., "TransfoRhythm: A Transformer Architecture Conductive to Blood Pressure Estimation via Solo PPG Signal Capturing," Applied Soft Computing, 203, 116316, 2026
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 80/100
The gist: TransfoRhythm introduces a Transformer-based Deep Neural Network architecture to estimate blood pressure using only solo PPG signals from the MIMIC-IV dataset…
Key concepts
- Solo PPG Signals
- This refers to using only the Photoplethysmography (PPG) signal from a single sensor instead of traditional cuff measurements. The paper focuses on extracting meaningful blood pressure information solely from these raw, continuous heart rate signals.
- Transformer Network
- A type of deep learning architecture designed for sequential data like time series. It uses a mechanism called Multi-Head Attention to weigh the importance of different parts of the input signal over time, allowing it to learn complex temporal patterns in the blood pressure data.
- Multi-Head Attention (MHA)
- A core component of the Transformer model that allows the network to simultaneously focus on multiple aspects (or 'heads') of the input data. This helps capture various relationships and dependencies within the PPG signal, leading to more accurate blood pressure predictions.
- Regressive Time Series Transformer
- A specific design for a transformer where the model learns from past data points to predict future values. This structure is ideal for blood pressure monitoring because it preserves the crucial sequence order of the signals while enabling parallel computation during training.
Terminology
Summary
TransfoRhythm introduces a Transformer-based Deep Neural Network architecture to estimate blood pressure using only solo PPG signals from the MIMIC-IV dataset, representing a novel approach to cuff-less BP estimation The paper demonstrates that this framework achieves highly accurate results with an RMSE of [2. p. 0 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. Specifically, TransfoRhythm achieves highly
accurate results with a Root Mean Square Error (RMSE) of [2. 84] and
a Mean Absolute Error (MAE) of [1. 84] and
a Mean Absolute Error (MAE) of [1. 06] for systolic and diastolic blood
pressures, respectively.Where the paper says it
Where the paper says it
Where the paper says it
Problem Context
Blood pressure serves as a critical health indicator for accurate and timely diagnosis and/or treatment of hypertension <ref:2404.15352#pg2> 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 <ref:2404. p. 0 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. Such an approach, however, carries the following drawbacks: (i) Cuff
devices are not readily accessible in resource-limited settings, leading to the
lack of awareness and control of hypertensive conditions; (ii) The inflation
and deflation of cuffs can disrupt gold-standard hypertension diagnosis and
hypotension surveillance, making these devices less suitable for frequent or
prolonged usage, and; (iii) Widely used cuff-based devices fail to provide continuous BP measurement, limiting immediate detection of hypotension and
real-time therapy titration [4]. Alternative methods of BP measurement,
such as cuff-less monitoring, have been explored to provide continuous monitoring and reduce the potential for measurement errors caused by cuff-based
methods. Alternative methods of BP measurement,
such as cuff-less monitoring, have been explored to provide continuous monitoring and reduce the potential for measurement errors caused by cuff-based
methods. Cuffless BP measurement offers the potential for continuous monitoring, which enables long-term hypertension management and facilitates
the precise diagnosis of cardiovascular conditions. Nocturnal BP provides critical information for the risk assessment
of cardiovascular diseases (CVD). For instance, abnormally high surges or
deep dips in BP during the night (more than 10 − 20%) are strongly linked
to severe health outcomes, including an increased possibility of stroke, heart
failure, and even mortality [5].Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Data Preparation and Feature Engineering
The study utilizes the Medical Information Mart for Intensive Care (MIMIC)-IV Waveform Dataset available on Physio-Net <ref:2404.15352#pg7> The recording process was conducted at a sampling rate of 62. p. 5 MIMIC-IV includes ECG, PPG, and
invasive ABP signals and is the latest and most recently released Dataset of
its kind following a joint effort from Beth Israel Deaconess Medical Center
(BIDMC) and Massachusetts Institute of Technology (MIT). The recording
process was conducted at a sampling rate of 62. To
our knowledge, this paper represents the first study to apply the MIMIC IV
dataset for cuff-less BP estimation. Performance evaluation through comPreprint submitted to Applied Soft Computing February 19, 2025
arXiv:2404. In this study, in the preprocessing phase, data cleaning and signal filtering were performed. In the
data cleaning step, signals with frequent outrange amplitudes and records
that did not meet the minimum time duration (less than 15 minutes) as well
as certain frames of long records that contained disqualified signals, such as
flattened-line or extremely noisy segments were detected, and removed from
the records. In the second step of the preprocessing phase, i. , the signal
filtering step, a 5
th order Butterworth filter in the frequency range of 0. , the signal
filtering step, a 5
th order Butterworth filter in the frequency range of 0. 7
to 10 Hz and a 5
th order Moving Average Filter (MAF) were applied to remove overshoots and fluctuations. The feature extraction process begins with identifying and isolating the cycles within each frame of the PPG signal. This is accomplished through a
robust algorithm that simultaneously locates the peak and trough (valley)
points in both the PPG signal and its second derivative (SDPPG). frames are evaluated through amplitude thresholding to ensure quality by filtering out noise or artifacts caused by poor sensor contact or motion; frames
with average amplitudes outside the typical PPG range are skipped. From
the extracted cycles, a set of 12 features was selected based on their relationships derived from correlation analysis, with detailed descriptions provided
in Table 1 and visual representations in Figure. Table 1: Definition of the selected features. Feature Symbol Description
PPG_Cycle_Duration TD1 = Tend - Tstart Total Duration
PPG_Rise_Half_Peak Trhp The duration between the start to the
half peak amplitude (rise mode)
PPG_Peak_To_Notch TD2 = Tdn-Tp The duration between peak to dicrotic
notch
PPG_Rise_Peak Tp The duration between the start to the
peak
PPG_Fall_Half Tfh The duration between peak to half of
its amplitude (fall mode)
PPG_Fall_Peak TD3 = Tend - Tp The duration between peak to foot
(end)
PPG_BPM_Frame PBF Cycles number per frame
PPG_Integration PPGI = (Green Area) Under Curve (cycle) Area
SDPPG_Fall_Half_Foot SDPPG_Tfhf The duration between the start to the
half of its amplitude (fall mode)
SDPPG_Extremum_Amp SDPPG_AMP SDPPG Maximum Amplitude
SDPPG_Integration SDPPGI = (Blue Area) SDPPG Upper Curve Area
SDPPG_Foot_Peak TD4 = Tpeak - Tfoot SDPPG Duration between Foot to
Peak
Furthermore, T refers to the time frame sequence and equals 48.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Model Architecture Design
The proposed TransfoRhythm framework is built on the regressive time series transformer network designed to learn from time series data while maintaining sequence order and enabling parallel computation <ref:2404.15352#pg10> To expand the dimensionality of the feature vectors, an embedding feature similar to those used in Natural Language Processing (NLP) tasks is employed <ref:2404. p. 3 The TransfoRhythm framework
is built on the regressive time series transformer network, which is crafted
to learn from time series data while preserving sequence order and allowing
parallel computation during training and inference. To expand the feature
vectors’ dimensionality, an embedding feature similar to those used in Natural Language Processing (NLP) tasks is employed. To expand the feature
vectors’ dimensionality, an embedding feature similar to those used in Natural Language Processing (NLP) tasks is employed. Positional encoding is
then added to the extracted features to create the input data matrix, which
is fed into a model block via the MHA mechanism. To capture additional information about in-frame sequence dependency, multiple learnable sets (Heads) are stacked and concatenated, In the
current study, we utilize a multi-head attention mechanism with 14 heads
for BP estimation. The architecture parameters of the networks such as
heads-number are set regarding the experiments. 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.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance Evaluation and Comparison
The model achieved high predictive accuracy with mean R2 values of 0.993 for SBP and 0. p. 2 Recent advancements in machine learning (ML)
have sparked a growing interest in using ML-based methods for cuff-less BP
measurement from ECG and PPG signals. These ML models, particularly
deep learning (DL) architectures, have demonstrated significant potential in
extracting discriminative features from ECG and PPG signals, estimating
blood flow volumetric changes, and predicting systolic (SBP) and diastolic
(DBP) BP values. To
our knowledge, this paper represents the first study to apply the MIMIC IV
dataset for cuff-less BP estimation. Performance evaluation through comPreprint submitted to Applied Soft Computing February 19, 2025
arXiv:2404. Specifically, TransfoRhythm achieves highly
accurate results with a Root Mean Square Error (RMSE) of [2. 84] and
a Mean Absolute Error (MAE) of [1. al [12] developed a hybrid CNNLSTM model using PPG and ECG signals, achieving MAEs of 4. 41 mmHg
for SBP and 2. 41 mmHg
for SBP and 2. 91 mmHg for DBP on data from 84 subjects. 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. 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. There are various approaches to measuring
Blood Pressure, among which, cuff-based measurement [3] (where a cuff is
employed to temporarily stop the blood flow to obtain the extremum BP values) is the common traditional method used both in hospitals and in-home
care. Such an approach, however, carries the following drawbacks: (i) Cuff
devices are not readily accessible in resource-limited settings, leading to the
lack of awareness and control of hypertensive conditions; (ii) The inflation
and deflation of cuffs can disrupt gold-standard hypertension diagnosis and
hypotension surveillance, making these devices less suitable for frequent or
prolonged usage, and; (iii) Widely used cuff-based devices fail to provide continuous BP measurement, limiting immediate detection of hypotension and
real-time therapy titration [4]. Recent advancements in machine learning (ML)
have sparked a growing interest in using ML-based methods for cuff-less BP
measurement from ECG and PPG signals. These ML models, particularly
deep learning (DL) architectures, have demonstrated significant potential in
extracting discriminative features from ECG and PPG signals, estimating
blood flow volumetric changes, and predicting systolic (SBP) and diastolic
(DBP) BP values. These representations can
be extracted from the original PPG signal, multiple-order signal derivative,
and Pulse Wave Decomposition (PWD) of the PPG. In the modeling step,
to achieve accurate BP estimation, recent studies have employed various
methodologies ranging from traditional ML models, such as Support Vector
Regression (SVR) and Random Forest [29, 30, 31], to Deep Neural Network
(DNN) architectures [1, 32, 21, 24] such as LSTM networks, and Gated Recurrent Unit (GRU) models with attention mechanisms [19, 28, 33, 22]. al [12] developed a hybrid CNNLSTM model using PPG and ECG signals, achieving MAEs of 4. 41 mmHg
for SBP and 2. To
our knowledge, this paper represents the first study to apply the MIMIC IV
dataset for cuff-less BP estimation. Performance evaluation through comPreprint submitted to Applied Soft Computing February 19, 2025
arXiv:2404.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Future Work
The results demonstrate the superiority of the proposed TransfoRhythm framework for BP estimation, achieving MAE and STD values in the range of [1.50, 1.84] mmHg for DBP and [1. p. 2 41 mmHg
for SBP and 2. 91 mmHg for DBP on data from 84 subjects. al [12] developed a hybrid CNNLSTM model using PPG and ECG signals, achieving MAEs of 4. 41 mmHg
for SBP and 2. Specifically, TransfoRhythm achieves highly
accurate results with a Root Mean Square Error (RMSE) of [2. 84] and
a Mean Absolute Error (MAE) of [1. In the modeling step,
to achieve accurate BP estimation, recent studies have employed various
methodologies ranging from traditional ML models, such as Support Vector
Regression (SVR) and Random Forest [29, 30, 31], to Deep Neural Network
(DNN) architectures [1, 32, 21, 24] such as LSTM networks, and Gated Recurrent Unit (GRU) models with attention mechanisms [19, 28, 33, 22]. Lately,
transformer-based networks have emerged as major players in this field and
there are a handful of recent studies [20, 24, 34, 35, 36, 37] that investigated the potential of the attention mechanism for BP estimation through
PPG signals. 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.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved BP estimation accuracy using TransfoRhythm's Multi-Head Attention mechanism can provide highly accurate results with a Root Mean Square Error of 1.84 mmHg for diastolic blood pressure, meeting AAMI and BHS standards as shown in Table 3.
-
The system can achieve superior performance compared to benchmark models on the MIMIC IV dataset, outperforming ResNet1D, Hybrid CNN-RNN Network, and U-Net by achieving a lower MAE of 1.06 mmHg for systolic blood pressure and 2.21 mmHg for diastolic blood pressure (Table 4).
-
The model can provide continuous BP monitoring crucial for detecting nocturnal BP patterns, which are strongly linked to severe health outcomes like stroke and heart failure, as discussed in the Introduction regarding
abnormally high surges or deep dips in BP during the night.
-
By utilizing a regressive time series transformer network, the system can learn from time series data while
preserving sequence order and allowing for parallel computation during training and inference,
which enhances its ability to capture long-term dependencies in blood pressure signals. -
The framework's ability to handle variable-length inputs without using reduction methods allows it to
learn more accurate representations of the input
compared to conventional DL methods that often require padding or truncation.
Abstract
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.
Sources
- Cuff-less Arterial Blood Pressure Waveform Synthesis from Single-site PPG using Transformer & Frequency-domain Learning
- Pay Less Attention with Lightweight and Dynamic Convolutions
- Comparing Attention-based Convolutional and Recurrent Neural Networks: Success and Limitations in Machine Reading Comprehension
- Attention as Robust Representation for Time Series Forecasting
- Transformers in Time-series Analysis: A Tutorial
Related papers
- Open-H-RF: A Large-Scale Ultrasound Raw Channel Dataset Initiative
- Scalable Attention for 5G NR Channel Estimation
- Low-Altitude Wireless Networks: A Comprehensive Survey
- Sensing-Assisted OFDM with Index Modulation: A Waveform for Optical Wireless Integrated Communication and Localisation
- Iteration-Free Candidate Pruning and an Index-Domain VLSI Architecture for Sort-Free MIMO Box Decoding up to 1024-QAM
- A Foundation Model for Large-Scale Wireless Network Planning, Operation and Optimization