Hypertension Management via Photoplethysmography: An Ensemble Learning-Based Approach for Classification of Blood Pressure Using Fourier Synchrosqueezed Transform

2.1 PPG-based hypertension detection systems

Several studies have investigated the use of ML and DL algorithms for predicting the risk of hypertension. For instance, Yen et al. ‎[29] focused on enhancing hypertension classification accuracy using PPG signals from the PPG-BP figshare database ‎[30]. They developed DL models designed with varying parameters and different architectures, including long short-term memory (LSTM), bidirectional long short-term memory (BLSTM), deep residual network convolutional neural network (ResNetCNN) and Extreme Inception (Xception). The best-performing model was a combination of Xception and BLSTM, achieving 48% precision, 45% recall, and 76% accuracy. However, the results showed relatively low recall and precision values, implying a significant number of false negatives and false positives, which could lead to undetected hypertension cases and unnecessary interventions with higher healthcare costs.

On the other hand, Tjahjadi et al. ‎[31] achieved better results in hypertension classification using the same dataset. They used TF-based PPG features and a BLSTM architecture, obtaining accuracy, recall, and specificity of 94.64%, 88.09%, and 98.57%, respectively. Nevertheless, the BLSTM required considerable training time (30 min and 26 secs) despite using a relatively small dataset (900 observations). In a different approach, Tjahjadi and Ramli ‎[32] used the K-nearest neighbors (KNN) algorithm with 2100 PPG samples, obtaining accuracy, recall, and specificity of 86.77%, 74.28%, and 91.86%, respectively, for hypertension class level. While these studies [31, 32] outperformed Yen et al. ‎[29], their datasets lacked diversity in SBP and DBP values as presented in ‎Figure 2 and ‎Figure 3, respectively.

2.png

Figure 2. Histograms of SBP values taken from the figshare database

3.png

Figure 3. Histograms of DBP values taken from the figshare database

In another study, Liang et al. ‎[33] experimented with four ML models (AdaBoost, KNN, BT, and Logistic Regression) using various feature sets, including pulse arrival time (PAT), PPG morphology features, and a combination of PPG and PAT features. They conducted three classification trials: normotension vs. hypertension (NT vs. HT), normotension vs. prehypertension (NT vs. PHT), and normotension plus prehypertension vs. hypertension (NT+PHT vs. HT). The KNN model with the combined features showed the best performance, achieving F1 scores of 84.34%, 94.84%, and 88.49% for the respective trials. However, using PAT as a feature is not practical for hypertension management, as it requires an additional sensor for electrocardiogram (ECG) signals. Besides, the features extraction process demands high-quality PPG signals ‎[34], which is difficult due to their susceptibility to motion artefacts ‎[27].

In a different study, the same researchers explored utilizing TF analysis via the continuous wavelet transform (CWT) method for transforming the signals into scalogram representations with three color channels. The resultant representations were then fed as RGB images into a pretrained CNN model ‎[34]. The F1 score results for the classification trials were 92.55%, 80.52%, and 82.95% for NT vs. HT, NT vs. PHT, and NT+PHT vs. HT, respectively. While this approach improved the feature extraction process, the methodology was tested with a relatively small dataset, indicating the need for validation with a larger and more diverse dataset containing a wider range of SBP and DBP values.

2.2 FSST

PPG signals are non-stationary bio-signals characterized by time-varying properties resulting from dynamic physiological processes. Traditional frequency analysis techniques, such as the Fourier Transform (FT), face challenges when dealing with these signals due to their changing frequency components over time ‎[35]. The FT operates under the supposition that the signal displays stationary behavior through an unchanging frequency constitution over its temporal evolution. This assumption is embedded in the fundamental definition of the FT, which provides a frequency-domain representation of a signal without considering time variations. The continuous FT of a signal $s(t)$ is given by:

$S(f)=\int_{-\infty}^{+\infty} s(t) e^{-2 i \pi f t} d t$     (1)

where, $S(f)$ is the Fourier Transform of $s(t)$, and $f$ represents frequency. This integral transforms the entire signal $s(t)$ into the frequency domain, assuming that the signal's frequency components are constant over the entire duration of the signal.

For non-stationary signals, such as PPG, where the frequency content changes over time, this assumption is invalid. The FT fails to capture these time-varying characteristics, as it provides only a global frequency representation, losing all time-related information. To address this issue, TF analysis has been employed to gain insights into signal behavior, facilitating the effective analysis of non-stationary signals by capturing the evolving frequency content throughout time. One of the widely used methods for TF analysis is the STFT. STFT attempts to address the challenge of non-stationarity by applying a window function to the signal, allowing the analysis of frequency content within small time segments‎ [36]. The local frequency content of the signal is calculated using the equation:

$V_S^D(t, \gamma)=\int_{-\infty}^{+\infty} S(\alpha) D(\alpha-t) e^{-2 i \pi \gamma(\alpha-t)} d \alpha$       (2)

where, $S(\alpha)$ is the input signal, $D(t)$ represents the window function, $t$ denotes the time index, $\gamma$ corresponds to frequency, and $\alpha$ serves as a time-variable of integration.

However, STFT suffers from the classic trade-off between time and frequency resolution due to fixed window sizes ‎[37]. A narrow window provides good time resolution but poor frequency resolution, while a wide window offers good frequency resolution but poor time resolution. This restricts STFT's ability to effectively capture rapid changes in both time and frequency that often appear in non-stationary signals like PPG.

To overcome these limitations, the FSST offers an innovative solution by applying a nonlinear post-processing mapping to the conventional STFT ‎[38]. FSST aims to enhance the time-frequency representation by reassigning the signal's energy more accurately, resulting in better localization of frequency components [39]. The concept of reassignment in FSST facilitates signal interpretation through the redistribution of energy, and synchrosqueezing provides robust visualization and manipulation capabilities ‎[28]. The FSST formula can be expressed as:

$T_S^D(t, \omega)=\int_{-\infty}^{+\infty} V_S^D(t, \gamma) \delta\left(\omega-I_S(t, \gamma)\right) d \gamma$      (3)

where, $\delta(\omega)$ is the Dirac distribution, while $I_S(t, \gamma)$ denotes the instantaneous frequency estimation at time $t$ and frequency $\gamma$ obtained using the following expression:

$\begin{gathered}I_S(t, \gamma)=\frac{\partial \arg V_S^D(t, \gamma)}{\partial t}=\mathbb{R}\left\{\frac{1}{2 i \pi} \frac{\partial t V_S^D(t, \gamma)}{V_S^D(t, \gamma)}\right\} \\ \text { such that } V_S^D(t, \gamma) \neq 0\end{gathered}$     (4)

In FSST, the synchrosqueezing process significantly enhances the precision of frequency localization in the TF plane. This is achieved by reassigning the time-frequency representation $V_S^D(t, \gamma)$ obtained from the STFT to concentrate the signal's energy more precisely around the true instantaneous frequency trajectories $I_S(t, \gamma)$. By dynamically adjusting the representation of the signal $s(t)$, FSST mitigates the trade-off between time $t$ and frequency $\gamma$ resolution. The signal's energy, initially spread over a broader area in the TF plane due to the fixed window function $D(t)$ in STFT, is more accurately focused around $\omega$ using the Dirac distribution $\delta\left(\omega-I_S(t, \gamma)\right)$. This process allows $T_S^D(t, \omega)$ to better capture the nonstationary characteristics of PPG signals by following the rapid changes in frequency components $\gamma$ over time $t$. Consequently, FSST provides improved time $t$ and frequency $\gamma$ resolution, facilitating more detailed and accurate feature extraction from PPG signals for applications such as BP classification. A visual comparison between the resolution of FSST and STFT for a PPG signal is presented in Figure 4.

4a.png

(a) Normalized PPG signal

4b.png

(b) STFT representation

4c.png

(c) FSST representation

Figure 4. TF representations of a PPG signal

4.1 Classification results

We designed an ensemble learning-based classification system using MATLAB (version R2020a) to non-invasively determine BP levels from PPG signals. To validate the effectiveness of the proposed BT model, we conducted empirical comparisons against several state-of-the-art ensemble classifiers for this task. Specifically, we evaluated the BT model against BOT, SKNN, RUSBT, and SD methods. All classifiers were implemented as integrated in the MATLAB's Classification Learner application interface, where each model consisted of 30 base learners. To evaluate the models' sensitivity to variations in data size, we trained them on three different proportions of the full dataset: 10%, 30%, and 50%. Tables 3-5 report the classification results achieved by each algorithm over three trials for the 10%, 30%, and 50% data subsamples, respectively. This systematic validation framework allowed for an impartial assessment of the BT model's performance relative to other prominent ensemble techniques, under varying data availability.

To validate and compare the classification performance of models, we utilized a standardized evaluation framework. Specifically, we employed an 80:20 train-test split on the 3 subsets of the dataset created using the $C_n^3$ variable, which encapsulates both real and imaginary components of the FSST features. To ensure robust and unbiased estimates, we adopted a k-fold cross-validation approach. Specifically, we randomly partitioned each $C_n^3$ sub-datasets into 5 equal mutually exclusive subsets (folds). Then, each model was trained on 4 folds and directly evaluated on the held-out fold, repeating this process 10 times such that each fold was used once for validation. The average classification accuracy across the 10 iterations was then computed to provide a reliable aggregate performance metric for each model, minimizing variability that could arise from a single train-test split.

Table 3. Models performance using 10% of the data

Notes: 1. obs: observation. 2. sec: second.

Table 4. Models performance using 30% of the data

Table 5. Models performance using 50% of the data

The results across Tables 3-5 indicate that increasing the proportion of data generally improved performance of the BT and SKNN models. Specifically, the SKNN algorithm achieved training accuracies ranging from 94.5% to 97.1% and test accuracies ranging from 87.4% to 91.6% as the set size increased from 10% to 50% of the full dataset. The BT model demonstrated relatively superior and more consistent classification ability, with training accuracies varying between 96.7-98.5% and test accuracies from 91.5-96.2% over the different data sizes. However, the BOT, RUSBT and SD algorithms did not exhibit substantial gains in accuracy despite being provided with larger subsets. This suggests they may have been less capable of effectively learning the underlying patterns from the FSST-based feature space, likely attributable to inherent limitations in their respective architectures for this particular classification task and dataset.

For instance, the BOT model achieved a peak training accuracy of 82.3% and test accuracy of 79.2%. This suboptimal performance can be attributed to the fact that boosting algorithms are highly sensitive to noise and overfitting, especially in datasets with complex, non-linear relationships. The sequential nature of BOT means that any noise in the training data can be amplified [56], leading to poor generalization on the test set. Similarly, the RUSBT model attained a maximum training accuracy of 76.8% and validation accuracy of 76.7%.  RUSBT combines random undersampling with boosting to handle class imbalance. However, the random undersampling process can lead to the loss of important information, especially in small or moderately sized datasets. This loss of information, coupled with the boosting process's sensitivity to noise, likely hindered the RUSBT model's ability to learn effectively.

The SD model achieved a maximum training accuracy of 74.9% and test accuracy of 74.2%. The SD approach employs LDA on random subspaces of the feature space. However, LDA rests on the stringent assumption that data is linearly separable between classes [57]. PPG signals exhibit inherent nonlinearity and non-stationarity due to their complex physiological origin. Thus, the linear decision boundaries used by LDA are insufficiently expressive to accurately capture the intricate, nonlinear relationships embedded within the PPG feature space mapped by the FSST. Ultimately, while the BT and SKNN models demonstrated strong performance and robustness across different dataset sizes, the BOT, RUSBT, and SD models struggled due to their sensitivity to noise, loss of information during undersampling, and assumptions of linear separability, respectively. This analysis underscores the importance of selecting appropriate models that align with the data's characteristics to achieve optimal classification performance.

In terms of computational efficiency, the BT and SD models exhibited the most favorable training time scaling as the proportion of data was increased. Specifically, BT required 16, 46, and 89 seconds to train on 10%, 30%, and 50% of the full dataset, respectively. Comparatively, SD took 14, 25, and 43 seconds over the same training sizes. However, during model prediction the BT approach significantly outperformed SD in terms of speed, with observed throughput of up to 21,000 obs/sec versus 9,400 obs/sec for SD. This discrepancy can be attributed to inherent differences in their algorithmic architectures. BT leverages the parallelizable and efficient tree-based structure to rapidly classify new samples. In contrast, SD's classification mechanism involves more computationally intensive Linear Discriminant Analysis calculations during each prediction.

The remaining models exhibited less favorable scaling of training time with increasing data proportions compared to BT and SD. Notably, despite using a modest ensemble of only 30 base learners, the SKNN algorithm was significantly slower during model fitting. Specifically, SKNN required 3342 seconds to train on 50% of the full dataset, two orders of magnitude longer than BT or SD, with a correspondingly low prediction throughput of 42 obs/sec. This inefficiency stems from KNN's underlying classification mechanism, which involves calculating the distance between each new sample and all examples in the training set [58]. As data volume grows, this distance computation becomes prohibitively expensive from a complexity standpoint. While SKNN achieved good prediction accuracy, its quadratic computational complexity severely limits the algorithm's potential for further tuning by increasing hyperparameters like ensemble size or training data proportions. Any such adjustments would excessively prolong training time at the detriment of runtime performance required for real-time applications.

Training time also scaled less favorably for the BOT and RUSBT models as dataset proportions increased. BOT's sequential approach, whereby each additional tree is fit to emphasize mistaken examples from previous trees, intrinsically demands greater computational burden relative to a purely parallel ensemble method like BT. Moreover, RUSBT compounds this complexity by incorporating random undersampling during training. The added steps of subsampling and sequential boosting accumulation elongate its training relative to simpler bootstrapping-based techniques.

10a.png

(a) Model 1

10b.png

(b) Model 2

10c.png

(c) Model 3

Figure 10. BT tuning process

Ultimately, BT emerged as the preferable algorithmic choice for this real-time BP classification application owing to its balanced optimization of accuracy, efficiency, and scalability. Its parallelizable tree-based framework facilitates rapid model fitting and prediction even on larger datasets through sample-wise decomposition and averaging. This makes BT well-suited for applications requiring timely inference, such as continuous hemodynamic monitoring using PPG sensors. In contrast, approaches like BOT and RUSBT that impose sequential dependencies exhibited poorer calibration of computational demand to growing data volumes. Therefore, BT's combination of robust performance and favorable runtime properties support its suitability for the proposed real-time blood pressure estimation system.

To gain further insights into the BT model, we conducted parameter optimization by systematically increasing the number of base learner trees from 10 to 200. This tuning process employed a held-out validation subset comprising 20% of the original data. Figure 10 depicts the classification accuracy of BT models trained on datasets constructed from different FSST features, as the number of constitutive decision trees is incremented.

Across all feature representations, classification accuracy increases rapidly with additional trees, suggesting the BT ensembles converge near 200 trees. The tuning process thereby provides further empirical evidence confirming BT as an appropriate modeling approach for these nonlinear challenges involving large, complex biomedical datasets. Table 6 displays the results of the tuned model using the 48,072 size dataset, with the training set comprising 15,524 subjects for each class and the test set comprising 500 subjects for each class. Training required approximately 4 minutes per model, a reasonable duration given the large data volume. Performance was evaluated based on accuracy, confusion matrices, and F1 score using specific classification tasks as detailed below.

Table 6. The BT models’ performances

As depicted in Table 6, the proposed methodology exhibited good performance, achieving 100% training accuracy for each experiment. Testing accuracies were 96.9%, 95.7% and 96.1% for Experiments 1, 2 and 3, respectively. It is evident that model 1 performed best among the three, although all three models proved to be effective predictors of BP levels.

The training and testing confusion matrices (Figures 11 and 12) provide insight into model performance. The matrices outline relationships between true versus predicted classes, with diagonal cells (blue) representing TPs/TNs and off-diagonal cells (red) representing FPs/FNs. Columns indicate predicted classes (output axis) while rows indicate actual classes (target axis). Grey cells denote positive predictive values (column-wise) and true positive rates (row-wise).

The training matrices confirmed 100% accuracy distinguishing actual cases within each class. It indicates also that of all predicted cases 100% were correctly classified. However, evaluating the reliability of the experimental models relies on analyzing their test performance. Specifically, experiment 1 correctly classified 97.4%, 94.8%, and 98.6% of actual HT, PHT, and NT cases, respectively. Further, experiment 1 indicates also that of all HT, PHT and NT predictions, 98.4%, 97.7% and 94.8% were correctly classified, respectively.

Similarly, experiment 2 correctly classified 95.2%, 94%, and 97.8% of actual HT, PHT, and NT cases, respectively, while indicating also that of all HT, PHT and NT predictions, 97.7%, 96.1% and 93.3% were correctly classified, respectively. Furthermore, experiment 3 correctly classified 97%, 93%, and 98.2% of actual HT, PHT, and NT cases, respectively, while indicating also that of all HT, PHT and NT predictions, 98.2%, 97.1% and 93.2% were correctly classified, respectively.

The models underwent three classification trials: non-HT (NT + PHT) vs HT, NT vs HT, and NT vs PHT, in line with previous literature [31-34]. Tables 7-9 overview performance metrics - F1 score, precision, specificity, sensitivity, and accuracy - for model 1-3 across the trials. Model 3 attained the best F1 score (99%) for NT vs HT, while Models 1-2 scored 97.5-98%. Model 1 outperformed for NT+PHT vs HT (99%) and NT vs PHT (99.1%) tasks, while models 2-3 scored 98.9-96.6% and 98.8-98.8%, respectively. Overall, the models demonstrated comparable and consistently high performance across trials, indicating capabilities for BP level detection.

4.2 Comparative analysis

This section evaluates the system’s classification performance relative to prior studies. Table 10 compares F1 scores from our work and others [31-34] for the (NT + PHT) vs HT, NT vs PHT, and NT vs HT classification tasks. According to the comparison results, our proposed models demonstrate superior robustness and efficiency. A major reason is that our models were trained on a much larger and more diverse dataset comprising a wider BP value range, an aspect not considered in earlier work.

11.png

Figure 11. Confusion matrices describing the training performance of the three experimental models

12.png

Figure 12. Confusion matrices describing the testing performance of the three experimental models

Table 7. The models’ performances using the first experimental dataset (real variable)

Table 8. The models’ performances using the second experimental dataset (imaginary variable)

Table 9. The models’ performances using the third experimental dataset (absolute variable)

Table 10. Results comparison relative to prior studies

Note: TS: Test Subjects.

The proposed BT models achieved F1 scores of 99%, 99.1%, and 99% for the NT vs. HT, NT vs. PHT, and non-HT vs. HT classification tasks, respectively. This exceeded prior published method performance, aside from study [32] which reported F1 scores of 100% for NT vs. HT and NT vs. PHT using a KNN model. However, their approach attained a lower F1 of 90.8% for the non-HT vs. HT task compared to our BT models' consistent F1 scores exceeding 98.8%. Meanwhile, Tjahjadi et al. [31] achieved competitive yet slightly lower F1 scores of 97.29%, 97.39%, and 93.93% across the same tasks using a smaller 786 sample dataset. However, both studies [31, 32] relied on more limited data lacking the diversity in BP ranges observed in our sample collection (Figures 2, 3, 6). Additionally, our BT models were evaluated on a larger test set of 500 subject-segments per class, providing a more robust evaluation. Moreover, training each BT model in our study required only approximately 4 minutes to calibrate, significantly faster training times than previous works - for example over 30 minutes for the BLSTM model in reference [31] using a smaller 786 sample set, and over 350 minutes for the CNN approach in reference [34] using 2323 samples.  

For a fair comparison with previous studies that employed similar modeling techniques, Table 10 presents results from Liang et al. [33] who used various feature extraction methods as inputs for a BT classifier. In their study, a combination of PAT and 10 PPG morphological features delivered the highest F1 scores of 94.13%, 83.88%, and 88.22% for NT vs. HT, NT vs. PHT, and non-HT vs. HT classifications, respectively. In contrast, when only FSST features of the PPG signal were used as input, our proposed BT model demonstrated significantly enhanced performance across all tasks, with F1 scores exceeding 96%. Although Liang et al. [33] achieved competitive results using a KNN model with PPG+PAT features, their method required an additional sensor beyond just PPG. Moreover, morphological feature extraction is sensitive to signal quality [34] and susceptible to motion artifacts, which may limit its practical applicability [27].

When solely using PPG signals without additional sensors, Liang et al. [33] reported markedly lower performance from their BT model, achieving F1 scores of 84.98%, 78.48%, and 75.32% for NT vs. HT, NT vs. PHT, and non-HT vs. HT classifications, respectively. Their KNN classifier yielded nearly identical results. In contrast, our BT models attained significantly higher F1 scores exceeding 96% across all tasks using only 44 features directly extracted from the raw PPG waveform via FSST, without requiring special pre-processing steps to isolate morphological characteristics. Notably, our proposed framework maintained this consistent, superior classification ability while overcoming limitations of prior work such as smaller, less diverse datasets [31, 32], computationally expensive training procedures [31, 34], and reliance on additional sensors or susceptibility to noise inherent in complex feature extraction [33, 34]. Overall, these comparisons demonstrate the proposed models establish new state-of-the-art performance for non-invasive BP monitoring using exclusively PPG data, represented by efficiently extracted Fourier features, without pre-processing vulnerabilities or constraints of earlier efforts.

4.3 System design and clinical perspectives

PPG provides a convenient and noninvasive means of measuring pulsatile blood volume changes, with the periodic waveform conveying insights into cardiovascular health [17]. However, deriving meaningful features from PPG signals is confounded by the nonstationary nature of the signals, whereby the statistical properties evolve dynamically over time. To address this, time-frequency decomposition was performed using FSST to generate high-resolution spectrograms revealing transient cardiovascular variabilities related to BP regulation. Spectrographic analysis enabled engineering statistical features including mean, variance, skewness and kurtosis, all together represent pressure-dependent physiological changes missed by basic morphological features. This set of TF signal features formed the input vectors for the BP classifier, providing a robust representation of the non-stationary dynamics within the PPG data.

For the classification architecture, an ensemble approach was pursued using 200 bagged decision trees, improving accuracy and robustness over individual models by reducing variance and overfitting [50]. Bagging provides an efficient means to leverage large training datasets as the models can be fit in parallel on data subsets. Specifically, the model was trained on a sizeable dataset spanning normotensive, prehypertensive, and hypertensive BP ranges collected from the MIMIC-III database [41]. This diversity is expected to improve real-world generalizability and reduce biases compared to models trained on a narrow BP distribution. Low computation complexity enabled training in just minutes per experiment on a dataset of this scale. Additionally, testing on a balanced dataset better assesses performance across the spectrum of the population.

The approach demonstrated potential as inexpensive assistant tool in clinical setting, as the experimental models achieved F1 scores ranging from 96.6% to 99.1% for non-HT vs. HT, NT vs. HT, and NT vs. PHT classifications. Trials that play a significant role for routine prehypertension and hypertension detection, which in turn contribute for the early hypertension diagnosis and management [34]. Additionally, these results justify employing the proposed system in real-world scenarios where various sources of disturbance could interfere, owing to synchrosqueezing stability in handling perturbated signals [59]. Which is reasonable as the PPG signals were captured from patients in ICU wards [40].

Furthermore, by requiring only raw PPG input, the technique provides a convenient means for wearable-based hypertension monitoring without reliance on cuff devices. The lightweight tree models allow implementation without cloud computing, enabled by low memory footprint compared to deep neural networks. The key advantages in using bagged trees over DL models in wearable monitoring devices are summarized as follows:

-Bagged decision tree models are highly suitable for implementation on wearable devices due to their lower complexity and minimal computational requirements. Storing the parameters for decision tree models requires very little memory on the wearable hardware. This allows the entire bagged ensemble to be housed on the device, rather than relying on cloud connectivity for model storage and predictions. This reduces latency, costs, and privacy concerns by keeping data localized. In contrast, deep neural networks have substantial storage needs due to the large number of weight parameters distributed across many layers, making cloud storage more practical.

-For prediction, bagged trees run efficiently on cheaper, lower-power CPUs by avoiding expensive high-RAM GPUs required by deep nets to process their computationally intensive activation layers in parallel. By avoiding such specialized hardware, tree ensembles enable cost savings on the device processors.

-Bagged trees support continuous incremental learning by continuously adding new models to the ensemble on-device as more training data becomes available [48]. Retraining deep nets to update their weights is much more computationally expensive. This makes regular model updating prohibitive on wearable processors.

-For clinical applications, model explanations matter for regulatory approval, accountability, and transparency around AI assisted medical decision making. Accordingly, decision trees have clear logical structures that are easy to understand and visualize making them more interpretable. In contrast, deep neural networks operate like a black box, thus, obscuring the basis for their predictions.

By enabling affordable BP monitoring directly through consumer wearables, this approach promotes accessible screening without expensive medical devices. The lightweight system allows self-tracking globally, even in resource-constrained settings, supporting prevention through timely personalized notifications. This addresses many patients’ lack of knowledge regarding acceptable BP levels [25, 26], which currently relies on infrequent clinic measurements. Widespread deployment can enhance early diagnosis, convenient management, and improved outcomes by averting target organ damage and associated healthcare costs from untreated hypertension.