Deep Learning for Non-Invasive Blood Pressure Monitoring: Model Performance and Quantization Trade-Offs

Abstract

The development of non-invasive blood pressure monitoring systems remains a critical challenge, particularly in resource-constrained settings. This study proposes an efficient deep learning framework integrating Edge Artificial Intelligence for continuous blood pressure estimation using photoplethysmography (PPG) signals. We evaluate three architectures: a residual-enhanced convolutional neural network, a transformer-based model, and an attentive BPNet. Using the MIMIC-IV waveform database, we implement a signal processing pipeline with adaptive filtering, statistical normalization, and peak-to-peak alignment. Experiments assess varying temporal windows (10 s, 20 s, 30 s) to optimize predictive accuracy and computational efficiency. Attentive BPNet achieves the best performance, with systolic blood pressure (SBP) estimation yielding a mean absolute error (MAE) of 6.36 mmHg, diastolic blood pressure (DBP) an MAE of 4.09 mmHg, and mean arterial pressure (MBP) an MAE of 4.56 mmHg. Post-training quantization reduces the model size by 90.71% (to 0.13 MB), enabling deployment on Edge devices. These findings demonstrate the feasibility of deploying deep learning-based continuous blood pressure monitoring on edge devices. The proposed framework provides a scalable and computationally efficient solution, offering real-time, accessible monitoring that could enhance hypertension management and optimize healthcare resource utilization.

1. Introduction

Cardiovascular diseases remain a leading global health concern, with hypertension affecting approximately 1.28 billion adults worldwide [1]. Continuous blood pressure monitoring is essential for the early detection, management, and prevention of cardiovascular complications [2]. However, traditional oscillometric cuff-based measurements suffer from significant limitations, including low temporal resolution, patient discomfort, and susceptibility to artifacts, such as white coat hypertension, which can compromise diagnostic accuracy [3,4]. The development of cuffless blood pressure estimation offers a promising alternative, leveraging photoplethysmography (PPG) and electrocardiography (ECG) signals for continuous, non-invasive monitoring [3,5]. At the same time, advancements in Edge Artificial Intelligence have enabled the deployment of sophisticated machine learning algorithms on resource-constrained embedded systems. These technologies facilitate real-time, on-device inference while addressing key challenges associated with cloud-based solutions, such as latency and data privacy concerns [6,7].

Despite these advancements, wearable blood pressure monitoring systems still face challenges in achieving high accuracy and cross-subject generalizability [8]. Many existing methods require subject-specific calibration, which limits large-scale adoption in clinical settings [9]. Additionally, optimizing deep learning architectures for deployment on edge devices while maintaining clinical-grade accuracy remains a significant research challenge.

This study introduces a comprehensive Edge AI-based framework for continuous, non-invasive blood pressure estimation using PPG signals.

Our research objectives are to do as follows:

• Develop accurate models for estimating mean, systolic, and diastolic blood pressure.
• Optimize these models for deployment on resource-constrained devices.
• Evaluate model performance in terms of accuracy, computational efficiency, and inference time.

Our contributions include the following:

• Proposing a novel Edge AI-based blood pressure estimation framework.
• Demonstrating the feasibility of continuous monitoring on wearable devices without the need for subject-specific calibration.
• Providing insights into the trade-offs between accuracy and computational efficiency in Edge AI applications for healthcare.

Unlike prior studies that primarily focus on high-resource environments or calibration-dependent models, our work presents a scalable, calibration-free deep learning framework optimized for wearable edge devices. This approach advances the field by enabling accurate, real-time blood pressure monitoring without compromising computational efficiency or generalizability.

The remainder of this paper is structured as follows: Section 2 reviews contemporary literature on cuffless blood pressure estimation and healthcare-focused Edge AI applications. Section 3 details our methodological framework, including signal preprocessing, model development, and Edge AI optimization strategies. Section 4 presents experimental results and their implications. Section 5 concludes with key findings and future research directions.

2. Related Works

The advancement of continuous, non-invasive blood pressure monitoring is a critical area in cardiovascular diagnostics. This section reviews the evolution of these methodologies, emphasizing their theoretical foundations, practical limitations, and integration with edge artificial intelligence for real-time healthcare applications.

Non-Invasive Blood Pressure Monitoring Techniques: Traditional oscillometric cuff-based methods, despite being the clinical standard, present significant limitations for continuous monitoring, including measurement intermittency, patient discomfort, and susceptibility to artifacts like white coat hypertension. These limitations have driven extensive research into cuffless monitoring approaches [1]. Several non-invasive techniques have been explored. Pulse Transit Time (PTT) estimates blood pressure by measuring the time a pulse wave travels between two arterial sites. While Mukkamala et al. highlight its potential, PTT is affected by arterial stiffness variations and requires frequent calibration [10,11]. Pulse Wave Analysis (PWA) examines arterial waveform characteristics to estimate blood pressure. Avolio et al. discuss its ability to assess central blood pressure and arterial stiffness, often relying on applanation tonometry or photoplethysmography [10]. This review further explores recent advancements and ongoing challenges in noninvasive blood pressure estimation, focusing on cuffless techniques such as Pulse Transit Time (PTT), along with emerging pressure-based, ultrasound-based, and deep learning methods, emphasizing their potential for accurate and continuous monitoring in both clinical and home settings [11,12]. Bioimpedance measures electrical impedance variations in body tissues to estimate blood volume and indirectly blood pressure, though it remains in early development [13,14,15]. The vascular unloading technique, or volume-clamp method, maintains constant blood volume in finger arteries, enabling continuous measurements, as described by Fortin et al. [16,17]. Ultrasound-based methods utilize arterial diameter changes to infer blood pressure, with Wang et al. demonstrating a wearable ultrasonic device for continuous monitoring [18,19]. Optical techniques, such as imaging photoplethysmography and smartphone-based methods, leverage skin color changes for blood pressure estimation, though motion artifacts and skin tone variations remain challenges [5,20].

Each of these approaches offers unique advantages and limitations. PWA provides detailed cardiovascular insights but requires precise sensor placement. PTT is simple yet lacks long-term stability. The vascular unloading technique delivers accurate continuous measurements but requires specialized hardware. Ultrasound-based methods directly measure arterial diameter changes but face miniaturization and power constraints. Optical techniques show promise for integration into consumer devices but require further refinement for reliability.

Photoplethysmography (PPG) for Blood Pressure Estimation: The shift away from mercury sphygmomanometers has accelerated the adoption of noninvasive methods like PPG, which offer simple, cost-effective, and wearable solutions for continuous blood pressure monitoring [21]. Kurylyak et al. developed an artificial neural network (ANN) using 21 PPG features to estimate systolic and diastolic blood pressure [22]. Xing and Sun employed a multilayer feed-forward neural network to improve accuracy [5]. Schlesinger et al. initially used an AlexNet-inspired CNN on PPG spectrograms and later introduced a Siamese network for refined blood pressure estimation [23]. Casadei et al. applied an ANN regression model to PPG data from the MIMIC II database [24]. Mukkamala et al. explored multiple PPG-based blood pressure estimation methods, including volume clamping, oscillometry, and waveform feature extraction [25].

Edge Artificial Intelligence and Computing in Healthcare: Edge Artificial Intelligence, which enables machine learning on resource-constrained devices, is revolutionizing healthcare by facilitating real-time monitoring while reducing latency and privacy concerns associated with cloud-based solutions [26]. It offers several advantages: real-time processing, enabling immediate physiological analysis [27]; privacy preservation, by processing sensitive health data locally [28]; power efficiency, through optimized ML models on low-power microcontrollers [27]. However, challenges remain. Model compression is essential for maintaining accuracy while reducing size, with pruning, quantization, and knowledge distillation being key techniques [29]. Hardware constraints necessitate careful optimization of ML models for limited-memory devices [29]. Calibration and adaptation require efficient on-device learning to ensure reliable blood pressure estimations [29].

Recent advancements have addressed these challenges. Hymel et al. introduced Edge Impulse, a cloud-based platform that streamlines Edge AI development, integrating data collection, digital signal processing, and ML deployment optimizations [30]. Tsoukas et al. reviewed Edge AI applications in healthcare, detailing model optimization techniques like quantization and pruning for embedded deployment [31]. Anbukarasu et al. demonstrated TinyHR, a TinyML-based heart rate estimation model deployed on an ESP32 device, showcasing low-power solutions for biomedical applications [32].

Signal Processing Techniques for Physiological Signals: Effective signal processing is crucial for extracting meaningful insights from physiological data. Preprocessing typically begins with detrending techniques to remove low-frequency artifacts, followed by bandpass filtering (0.5–10 Hz) to preserve cardiac cycle components [33,34]. Feature extraction involves peak detection algorithms to identify systolic and diastolic points in PPG and ABP signals, along with waveform segmentation for cycle-specific analysis [34,35]. Morphological analysis extracts pulse amplitude, width, and reflection indices for further cardiovascular assessment [36]. Wavelet transforms enable time-frequency decomposition for feature extraction and noise reduction [37]. Singular Spectrum Analysis (SSA) improves PPG signal quality by separating key components from noise [35]. Derivative analysis, including velocity and acceleration photoplethysmograms, provides deeper vascular insights [33].

Quantization Techniques for Edge Device Optimization: Quantization is a crucial optimization for deploying deep learning models on edge devices. By reducing model weight and activation precision—typically from 32-bit floating point to lower bit-width representations—quantization significantly lowers memory usage and computational costs. Post-training quantization (PTQ) and quantization-aware training (QAT) are two primary approaches. PTQ includes dynamic quantization, where weights are pre-quantized while activations remain dynamic, and static quantization, which applies fixed calibration for both weights and activations. Key parameters include the scale factor (S) and zero point (Z), ensuring accurate mapping between quantized and floating-point values. Per-channel quantization enhances precision at the cost of increased complexity [38].

The edge deployment workflow involves identifying computationally intensive operations for quantization, applying post-training dynamic quantization, and refining with static quantization if needed. Calibration on a representative dataset ensures optimal performance. Quantization typically reduces the model size by 3–4×, accelerates inference by 1.2–3×, and minimizes power consumption, though aggressive quantization may impact accuracy. Quantization-aware training helps mitigate accuracy loss [37,38,39,40].

Despite notable advancements in non-invasive blood pressure monitoring, several key challenges remain. First, the integration of advanced signal processing techniques with Edge AI for blood pressure estimation is still underexplored. Second, the development of robust, calibration-free models that generalize well across subjects—without the need for personalized calibration—remains a significant hurdle, particularly for edge-based deployments. Third, there is a lack of comprehensive comparative studies evaluating Edge AI solutions across different hardware platforms and real-world use cases. In response to these gaps, our research presents a novel Edge AI-based framework for continuous, non-invasive blood pressure estimation using PPG signals. While several existing studies have employed deep learning models with wearable sensors, many rely on resource-intensive architectures or require subject-specific calibration, limiting their scalability and practicality. In contrast, our approach focuses on developing a calibration-free, cross-subject generalizable model optimized for real-time performance on resource-constrained edge devices. By combining advanced signal processing, efficient deep learning architectures, and post-training quantization, our work addresses the critical challenges of scalability, efficiency, and clinical applicability. This framework not only enhances the feasibility of real-time, continuous blood pressure monitoring but also contributes to improved cardiovascular health management and better patient outcomes in diverse settings.

3. Dataset Description

This study uses the MIMIC-IV Waveform Database (version 0.1.0), which includes photoplethysmography (PPG), electrocardiography (ECG), and arterial blood pressure (ABP) measurements from 198 subjects, totaling 200 records. The dataset follows a hierarchical structure based on subject-specific identifiers (subject_id) and hospital admission designations (hadm_id), enabling efficient retrieval via the WFDB (version 0.1.0), software framework [41]. To prevent data leakage, we stratified the dataset at the patient level, allocating 60% for training, 20% for validation, and 20% for testing. We retained only records containing both PPG and ABP signals, resulting in 60% of the initial dataset. The dataset maintains a uniform sampling frequency of 62.47 Hz and includes from 5 to 13 concurrent physiological signals per segment. Temporal duration varies from 0.2561 to 112,817.64 s, with segment counts ranging from 2 to 65 per patient.

We implemented a multi-stage preprocessing pipeline to ensure reliable cardiovascular parameter extraction. To eliminate transient artifacts, we removed the first 20 s of each signal before converting them into numerical arrays. After evaluating multiple segmentation strategies, we selected a 30 s window, yielding 1874 samples per segment, as the optimal balance between resolution and computational efficiency. We implement this initial 20 s removal step to systematically eliminate transient noise, such as movement-induced artifacts and sensor stabilization fluctuations, ensuring that our dataset consists solely of stable and physiologically meaningful signals.

To maintain physiological validity, we applied clinically accepted blood pressure thresholds, retaining systolic blood pressure (SBP) values between 75 and 180 mmHg and diastolic blood pressure (DBP) values between 40 and 110 mmHg. We ensured temporal alignment between PPG and ABP signals to preserve data integrity. These measures enhanced signal quality and ensured the dataset remained clinically relevant for model development and evaluation. Figure 1 confirms the effectiveness of the filtering process.

We applied advanced digital filtering techniques to enhance signal quality. A fourth-order Butterworth bandpass filter (0.5–8.0 Hz) refined PPG signals by preserving fundamental cardiac frequencies while suppressing low-frequency drift and high-frequency noise. ABP signals underwent high-pass filtering (0.05 Hz cutoff) to eliminate baseline wandering while maintaining key morphological features. We used zero-phase filtering with bidirectional coefficient application to preserve phase relationships essential for cardiovascular timing analysis. Figure 2 shows the before and after filtering of ABP and PPG signals, respectively. To standardize signal amplitude, we implemented min–max normalization with a stability factor (ε = 1 × 10⁻⁸) to prevent singular transformations (Equation (1)). This normalization was applied independently to each temporal window, ensuring consistent amplitude scaling without distorting local signal characteristics. Figure 3 illustrates the PPG and ABP signals before and after normalization. For precise waveform alignment, we developed a peak-to-peak alignment algorithm incorporating advanced signal processing techniques for PPG and ABP waveforms. This approach included systematic peak detection using physiological constraints and adaptive temporal thresholding (minimum inter-peak distance: one-tenth of the window duration), precise central peak localization through computational geometry (minimizing Euclidean distance from the temporal midpoint), and waveform translation using circular convolution for optimal peak alignment. Figure 4 shows a sample of the peak-to-peak alignment of ABP and PPG signals.

$$x_{\mathrm{normalized}}={\displaystyle \frac{x-\min \left(x\right)}{\max \left(x\right)-\min \left(x\right)+\mathsf{\epsilon }}}$$

(1)

This alignment strategy established a precise temporal correspondence between cardiovascular waveform features, improving phase consistency and preserving key morphological characteristics. By enhancing temporal coherence, this methodology significantly strengthened subsequent analytical procedures, enabling reliable cross-subject comparisons and facilitating the detection of subtle physiological variations in cardiovascular dynamics.

In selecting our waveform alignment strategy, we considered the trade-off between algorithm complexity and computational efficiency, particularly for edge computing deployment. While simpler peak searching methods are less computationally demanding, we chose a more precise peak-to-peak alignment algorithm incorporating advanced signal processing techniques driven by the potential for enhanced blood pressure prediction accuracy through improved waveform alignment. To mitigate the increased computational cost associated with this more complex algorithm, we employed post-training quantization, which effectively reduces the model size and computational demands, making our framework suitable for resource-constrained edge devices.

Finally, we developed a feature extraction framework to analyze photoplethysmographic waveforms, capturing essential temporal, statistical, and morphological characteristics for blood pressure estimation. Temporal analysis included first through fourth-order statistical moments, incorporating measures of central tendency, dispersion, and higher-order moments, such as skewness and kurtosis. Amplitude-based features, including global extrema and peak-to-peak variations, provided insight into signal fluctuations across multiple time windows. Morphological analysis employed an optimized peak detection algorithm tailored to PPG waveform characteristics, using physiologically constrained temporal thresholds. Extracted features included peak frequency for heart rate variability, inter-peak interval statistics for cardiac rhythm assessment, and the temporal distribution of successive peaks. Amplitude-based features captured statistical moments of peak amplitudes, variations in peak heights, and systematic characterization of peak prominence.

4. Methodologies

This investigation presents a comprehensive evaluation of novel deep learning architectures optimized for edge computing environments, specifically designed for photoplethysmography-based blood pressure estimation. To comprehensively assess the landscape of relevant deep learning approaches, we selected four distinct architectural paradigms for rigorous comparison: a residual-enhanced convolutional network, an attentive BPNet, a transformer-based model, and a feature-engineered neural network. This selection was driven by several key considerations. First, these architectures represent a diverse spectrum of deep learning techniques widely applied in time-series analysis and signal processing, ranging from efficient convolutional networks to state-of-the-art attention-based models. Second, given this study’s focus on edge deployment, we deliberately included architectures known for varying degrees of computational complexity, allowing us to investigate the critical trade-offs between model accuracy and suitability for resource-constrained edge environments. Third, we aimed to explore both established and novel approaches, incorporating recent advancements like Transformers and Attention mechanisms alongside more traditional convolutional methods and a feature-engineered baseline.

Residual-Enhanced Convolutional Architecture: The residual-enhanced architecture implements a sophisticated hierarchical feature extraction paradigm, incorporating skip connections to mitigate gradient degradation while maintaining computational efficiency.

Key Components:

• Initial Convolution: A convolution layer (64 filters, kernel size 7) extracts low-level features from raw PPG signals. Batch Normalization and ReLU activation stabilizes training dynamics.
• Residual Blocks: Three residual blocks progressively increase filter depths (64→128→256) [Equation (2)].

$$F\left(x\right)=H\left(x\right)+x$$

(2)

where $H\left(x\right)$ represents two convolutional layers within the block. Skip connections ensure gradient flow during back propagation.

3.

Global Pooling: Global average pooling aggregates temporal features into a fixed size representation.

4.

Dense Layer: Fully connected layer (128→64→3) refine extracted features before predicting SBP, DBP, and MBP.

Attentive BP Net: Attentive BPNet integrates convolutional layers with an attention mechanism to focus on critical regions of the PPG signal, enhancing interpretability and computational efficiency. The architecture is optimized to capture both local and global temporal dependencies in the signal.

Key Components:

• Initial Feature Extraction: The input layer processes raw PPG signals through a convolutional layer (3232 filters, kernel size 1515), followed by batch normalization and ReLU activation. Max-pooling reduces the temporal resolution while retaining salient features.
• Hierarchical Feature Extraction: Two additional convolutional layers (6464 and 128,128 filters, kernel size 77) extract higher-order features. Each convolutional layer is followed by batch normalization, ReLU activation, max-pooling, and spatial dropout (p = 0.1) to prevent overfitting in time-series data.
• Attention Mechanism: An attention layer computes importance weights for each time step using a dense layer with a tanh activation function. A softmax operation normalizes the weights across the time axis. The weighted feature map is computed through element-wise multiplication of the attention weights with the feature map.
• Global Pooling: Global average pooling (mean trends) and global max pooling (extreme values) are applied to the weighted feature map. The pooled outputs are concatenated to form a compact representation capturing both average and extreme signal characteristics.
• Dense Layers: Two fully connected layers (128,128 and 6464 neurons with ReLU activation) refine the features before outputting predictions for systolic (SBP), diastolic (DBP), and mean arterial pressure (MBP).
• Loss Function: A custom loss function was designed to incorporate physiological constraints (Equation (3)):

$$L_{\mathrm{BP}}=\mathrm{MSE}\left(y_{\mathrm{true}},y_{\mathrm{pred}}\right)+\mathsf{\lambda }\cdot \max \left(0.15-\left(y_{\mathrm{pred},\mathrm{SBP}}-y_{\mathrm{pred},\mathrm{DBP}}\right)\right)$$

(3)

This penalizes physiologically implausible differences between SBP and DBP predictions.

Transformer-Based Model: The transformer-based model leverages self-attention mechanisms to capture long-range dependencies in PPG signals, which are critical for understanding temporal variations in cardiovascular dynamics [42].

Key Components:

• Positional Encoding: Since transformers lack inherent temporal awareness, sinusoidal positional encodings are added to the input embeddings (Equations (4) and (5)):

$$PE\left(\mathrm{pos},2i\right)=\sin \left({\displaystyle \frac{\mathrm{pos}}{{\mathrm{10,000}}^{\frac{2i}{d_{\mathrm{model}}}}}}\right)$$

(4)

$$PE\left(\mathrm{pos},2i+1\right)=\cos \left({\displaystyle \frac{\mathrm{pos}}{{\mathrm{10,000}}^{\frac{2i+1}{d_{\mathrm{model}}}}}}\right)$$

(5)

2.

Embedding Layer: A dense layer maps the input signal into a higher-dimensional space ($d_{\mathrm{model}}=256$).

3.

Multi-Head Self-Attention: Four transformer layers are stacked, each consisting of multi-head self-attention (88 heads) followed by feed-forward networks. Residual connections and layer normalization stabilize training.

$$\mathrm{Attention}\left(Q,K,V\right)=\mathrm{softmax}\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$

(6)

4.

Global Pooling: The final transformer output is aggregated using global average pooling to reduce dimensionality while retaining salient features.

5.

Output Layers: Dense layers (6464 neurons with ReLU activation) refine the features before predicting SBP, DBP, and MBP.

4.1. Experimental Design

We optimized the training process to ensure robust convergence and generalization across model architectures. The models minimized Mean Squared Error (MSE) to predict systolic (SBP), diastolic (DBP), and mean blood pressure (MBP). Attentive BPNet enhanced accuracy by integrating physiological constraints into its custom loss function. We used Adam optimizer with initial learning rates of ($1\times {10}^{-3}$). A batch size exploration from 4 to 256 revealed that the Attentive BPNet performed best with smaller batches, benefiting from improved gradient estimation and lower memory overhead. To prevent overfitting, early stopping halted training after 10 epochs of no improvement. ReduceLROnPlateau decreased the learning rate by a factor of 0.2 when validation loss plateaued within a 5-epoch window. Model checkpointing preserved the best-performing model based on validation metrics. This training strategy ensured efficient learning, stability, and strong generalization while maintaining computational efficiency. In addition, for the feature engineering neural network, two dense layers were included, each followed by batch normalization and dropout (p = 0.3). The final dense layer outputs a single regression value for SBP/DBP/MBP.

We implemented a various post-training quantization, including dynamic range quantization, full integer quantization, float 16 quantization, and integer-only quantization on a TensorFlow Lite optimization framework to address computational and memory constraints in edge environments. This framework applies multiple quantization techniques, balancing model size, inference speed, and accuracy. A key component of the quantization process involves a calibration dataset generator, which samples 100 representative training data points to determine optimal activation quantization parameters. This systematic approach ensures accurate input distribution representation while minimizing quantization errors.

Dynamic range quantization applies asymmetric quantization for weights while keeping activations in floating-point. It transforms weights using $W_{\mathrm{quantized}}=\mathrm{round}\left(\mathrm{scale}\cdot W_{\mathrm{float}}+\mathrm{zero}\_\mathrm{point}\right)$, where scale and zero_point are determined per-tensor or per-channel to minimize quantization error. On the other hand, full integer quantization converts both weights and activations to 8-bit precision using $q=\mathrm{round}\left({\displaystyle \frac{r}{s}}\right)+z$, where r is the real value, s is the scaling factor, z is the zero point, and q is the quantized value. Statistical analysis of representative data ensures minimal information loss. Float 16 Precision Optimization reduces storage requirements while maintaining computational precision. It represents weights in half-precision binary format as $\mathrm{float} 16={\left(-1\right)}^{\mathrm{sign}}\times 2^{\mathrm{exponent}-15}\times \left(1+\mathrm{fraction}\right)$ with 1-bit sign, 5-bit exponent, and 10-bit fraction. In addition, integer-only quantization aggressively optimizes models for edge deployment by fully converting inputs, weights, activations, and outputs to 8-bit integer format. This approach significantly reduces computational overhead while preserving model performance.

These optimizations enhance efficiency while maintaining accuracy, making deep learning models viable for real-time edge deployment. The table below compares the performance of original and quantized models.

To ensure reproducibility and transparency, we detail here the hardware configuration employed for both training and testing the deep learning models in this study. All experiments were conducted using the Google Colaboratory (Colab) platform, which provides access to cloud-based computational resources. The primary hardware components utilized were an NVIDIA A100 GPU for accelerated deep learning computations, a multi-core x86-64 architecture CPU (Intel Xeon), and approximately 12–16 GB of RAM. Data and models were stored using the cloud storage infrastructure provided by Google Colaboratory.

4.2. Experimental Results

We evaluate non-invasive blood pressure estimation methods, focusing on model performance, quantization efficacy, and deployment feasibility in resource-constrained environments. The evaluation follows clinical standards and incorporates statistical validation essential for medical device assessment. The assessment framework applies key statistical metrics to ensure accuracy and reliability. Root Mean Square Error (RMSE) detects large estimation errors that could impact clinical decisions. Mean Bias identifies systematic over- or under-estimation tendencies, preventing consistent prediction errors. Standard deviation measures prediction stability, ensuring reliability in clinical applications. Mean absolute error (MAE) quantifies average deviation from true values, providing a clear measure of accuracy.

Table 1. Performance comparison of deep learning models for blood pressure estimation across different batch sizes, highlighting the impact on MAE, RMSE, and model stability. The residual-enhanced convolutional network demonstrates consistent accuracy.

Model	SBP MAE	SBP RMSE	SBP ME	SBP SD	SBP R2	DBP MAE	DBP RMSE	DBP ME	DBP SD	DBP R2	MBP MAE	MBP RMSE	MBP ME	MBP SD	MBP R2
Attentive BPNet	11.55	13.78	−10.65	8.75	−0.44	4.09	5.36	1.76	5.06	0.15	4.56	6.09	−2.33	5.63	0.28
Attentive BPNet—batch size 64	13.27	16.59	−11.21	12.24	−1.08	5.22	6.86	1.99	6.56	−0.40	6.25	8.30	−2.37	7.95	−0.34
Attentive BPNetl—Batch Size 8	7.77	10.85	−5.00	9.63	0.11	6.67	8.15	5.62	5.90	−0.97	5.13	6.91	2.04	6.60	0.07
Attentive BPNet—Batch Size 4	7.55	10.44	−5.74	8.75	0.18	7.88	8.96	7.31	5.19	−1.39	5.04	6.47	2.95	5.76	0.19
Attentive BPNet—Batch Size 16	10.76	14.96	−3.15	14.63	−0.69	8.36	9.67	5.68	7.82	−1.77	7.85	9.82	2.05	9.61	−0.88
Attentive BPNet- Wide and Deep batch 32	12.57	16.72	−5.10	15.93	−1.12	14.17	16.02	12.92	9.47	−6.62	10.86	12.91	6.34	11.24	−2.24
Attentive BPNet—Wide and Deep batch 4	14.10	18.33	−4.18	17.85	−1.54	13.10	15.24	11.50	10.01	−5.90	11.49	13.81	6.27	12.30	−2.71
Attentive BPNet—Wide and Deep batch 256	17.46	21.88	4.92	21.31	−2.62	19.98	22.89	18.98	12.80	−14.56	18.35	21.92	15.54	15.46	−8.35
TransfoRhythm	9.39	12.11	0.37	12.11	−0.11	9.46	10.64	8.71	6.11	−2.36	8.05	9.77	5.79	7.87	−0.86
Feature Engineering—Separate Models	10.22	12.68	5.45	11.45	−0.22	19.22	20.65	18.99	8.11	−11.66	12.08	13.46	10.75	8.09	−2.52
Feature Engineering—Separate Models—iteration 2	10.16	13.97	−6.49	12.37	−0.48	11.05	12.69	10.07	7.72	−3.78	10.40	11.75	8.87	7.70	−1.69
Feature Engineering—Separate Models—iteration 3	10.52	13.06	5.59	11.81	−0.29	10.89	12.12	9.23	7.86	−3.36	12.44	13.86	10.51	9.03	−2.74
Feature Engineering—Separate Models—iteration 4	13.38	15.69	9.63	12.39	−0.86	9.07	10.62	6.26	8.58	−2.35	11.10	12.66	8.61	9.28	−2.12
Residual Enhaced Convolutional Network	6.36	8.93	−1.11	8.86	0.40	8.35	9.78	7.81	5.88	−1.84	6.90	8.40	5.58	6.28	−0.37
Residual Enhaced Convolutional Network—deeper network	12.73	15.63	10.52	11.55	−0.85	13.07	14.96	12.67	7.95	−5.64	12.82	15.03	12.19	8.79	−3.40

shows the experimental results of applying different deep-learning architectures for blood pressure estimation, highlighting the impact of different model selections and batch sizes. The residual-enhanced convolutional network consistently outperformed other models, achieving the lowest errors across all metrics. It maintained an SBP MAE of 6.36 mmHg, DBP MAE of 8.35 mmHg, and MBP MAE of 6.90 mmHg, demonstrating superior predictive accuracy. In contrast, the transformer-based model exhibited significantly higher errors, with SBP MAE ranging from 9.39 to 13.69 mmHg, and further deteriorated post-quantization, making it less suitable for real-time applications.

Batch size variations significantly affected performance. Attentive BPNet performed best with a batch size of 4 (SBP MAE 7.55 mmHg) but degraded with larger batches, reaching 17.46 mmHg at batch size 256. Similarly, the transformer-based model’s SBP MAE increased from 9.39 mmHg to 13.69 mmHg as batch size increased. In contrast, the residual-enhanced convolutional network remained stable across batch sizes, reinforcing its robustness and suitability for real-world deployment.

Table 2. Model size reduction after quantization, highlighting compression efficiency across architectures.

Architecture	Original Size	Post-Quantization	Reduction
Attentive BPNet	5.64 MB	0.50 MB	91.13%
Transformer-Based	9.28 MB	1.10 MB	88.15%
Residual-Enhanced	1.40 MB	0.13 MB	90.71%

shows the impact of quantization on model size reduction across different architectures. Attentive BPNet achieved the highest compression, reducing from 5.64 MB to 0.50 MB (91.13%), followed by the residual-enhanced network, which shrank from 1.40 MB to 0.13 MB (90.71%). The transformer-based model, despite its initial large size of 9.28 MB, compressed to 1.10 MB, achieving the lowest reduction rate (88.15%). While all models benefited from quantization, the transformer-based model remained the largest post-quantization, indicating inefficiency for edge deployment. Attentive BPNet achieved the most significant reduction but still struggled with stability in accuracy. The residual-enhanced network balanced both compression and performance, making it the most efficient model for resource-constrained environments.

Table 3. Impact of quantization methods on model size, inference time, and accuracy across architectures. The residual-enhanced network maintains stability, while the transformer-based model suffers from extreme latency and accuracy degradation.

Attentive BP Net
Model	Orginal Model	Dynamic Range Quantized Model	Float 16	Full Integer Quantization	Integer-Only Quantization
Model Size (MB)	5.64	0.5	0.94	0.52	0.52
Size reduction (%)	-	−91.13	−83.3333	−90.78	−90.78
Inference Time	30.46	24.78	25.02	36.63	37.78
MAE	7.84	7.84	7.84	7.78	7.78
RMSE	10.42	10.41	10.42	10.36	10,365
ME (Bias)	−0.07	0	−0.07	−0.19	−0.19
SD	10.42	10.41	10.42	10.36	10.36
Transformer-Based Model
Model Size (MB)	9.28	1.77	2.01	1.1	1.1
Size reduction (%)		−80.92	−78.34	−88.14	−88.14
Inference Time	32.61	1793.35	1249.69	3087.91	3034.05
MAE	10.87	10.86	10.87	8.25	60.65
RMSE	12.85	12.84	12.85	11.25	62.33
ME (Bias)	7.79	7.79	7.8	−3.45	−60.67
SD	10.21	10.21	10.21	10.71	14.3
Residual-Enhanced Convolutional Network
Model Size (MB)	1.4	0.23	0.52	0.13	0.13
Size reduction (%)		−83.57	−62.85	−90.71	−90.71
Inference Time	6.23	4.66	1.2	2.6	2.6
MAE	7.77	7.75	7.77	7.65	7.65
RMSE	10.85	10.84	10.85	10.83	10.83
ME (Bias)	−5	−4.9	−5	−5	−5.2
SD	9.63	9.64	9.63	9.69	9.64

compares the effects of different quantization methods on model size, inference time, and performance across three architectures. The residual-enhanced convolutional network maintained the best balance between accuracy and efficiency, with the smallest post-quantization size (0.13 MB) and the lowest inference time (2.6 ms). Despite an over 90% reduction in size, its MAE (7.65) and RMSE (10.83) remained stable across quantization types, confirming its robustness for edge deployment. The attentive BPNet achieved the highest compression (91.13%), reducing from 5.64 MB to 0.5 MB, but inference time varied significantly. While full and integer-only quantization preserved accuracy, they increased latency to 36.63 ms and 37.79 ms, making real-time deployment less feasible. Dynamic range and float 16 quantization provided slightly better efficiency without significant accuracy loss. The transformer-based model performed the worst in efficiency. Despite reducing from 9.28 MB to 1.1 MB, inference time exploded to over 3000 ms in full and integer-only quantization, making it impractical for edge applications. Additionally, its MAE fluctuated drastically, reaching 60.65 in integer-only quantization, rendering the model unreliable. Overall, the residual-enhanced network consistently achieved low latency, minimal accuracy degradation, and high compression, making it the most viable option for edge deployment. The attentive BPNet remained a reasonable alternative with trade-offs in latency, while the transformer-based model proved inefficient for real-time applications.

To evaluate the calibration-free performance of the residual-enhanced network, we assessed its accuracy on a held-out evaluation dataset (20% of patients) comprised of subjects not seen during training or validation. On this independent dataset, the model achieved a mean absolute error (MAE) of 6.36 ± 8.86 mmHg for SBP, 8.35 ± 5.88 mmHg for DBP, and 6.90 ± 6.28 mmHg for MBP, demonstrating its ability to generalize to unseen individuals without subject-specific calibration.

5. Discussion

Regulatory bodies, such as the Association for the Advancement of Medical Instrumentation (AAMI), establish strict validation protocols for blood pressure monitoring devices. These standards require a mean difference of ±5 mmHg and a standard deviation below 8 mmHg, while the British Hypertension Society (BHS) mandates cumulative percentage thresholds for different accuracy grades. Evaluation of deep learning models showed varying compliance with these standards. Attentive BPNet initially failed to meet AAMI limits, with an SBP MAE of 13.27 mmHg. Architectural refinements improved performance, reducing SBP MAE to 7.55 mmHg, DBP MAE to 7.88 mmHg, and MBP MAE to 5.04 mmHg. While within clinical limits for mean blood pressure, systolic predictions remained near AAMI’s upper boundary. The transformer-based model exhibited marginal compliance, with SBP MAE at 9.39 mmHg, DBP MAE at 9.46 mmHg, and MBP MAE at 8.05 mmHg. However, post-quantization, its performance degraded, with MAE increasing to 10.87 mmHg, exceeding clinical thresholds.

The residual-enhanced network consistently met clinical standards, achieving an SBP MAE of 6.36 mmHg, DBP MAE of 8.35 mmHg, and MBP MAE of 6.90 mmHg. Post-quantization, it maintained accuracy, with MAE variation limited to ±0.12 mmHg and bias deviation below 0.2 mmHg. Attentive BPNet showed moderate stability but remained sensitive to quantization. The transformer-based model failed to meet AAMI standards post-quantization, with increased MAE and inference latency. The residual-enhanced network emerged as the most viable candidate for clinical deployment, balancing regulatory compliance, efficiency, and statistical reliability. It consistently met AAMI standards, maintained minimal deviation in key measurements, and demonstrated optimal compression (0.13 MB) with an inference time of 2.6 ms. Its stable performance across quantization schemes makes it the most practical choice for real-world blood pressure monitoring, particularly in resource-limited settings.

While our evaluation primarily aligns with AAMI standards, calculating cumulative errors in accordance with BHS guidelines represents an important future direction. Given the stable performance metrics, we expect our residual-enhanced convolutional network would meet the cumulative error thresholds defined by BHS. We acknowledge that the AAMI standard specifies compliance in terms of mean difference (bias) within ±5 mmHg and standard deviation under 8 mmHg, whereas our manuscript primarily reported mean absolute error (MAE). To ensure clarity, we explicitly calculated and confirmed that the residual-enhanced convolutional network’s mean difference (bias) is within the ±5 mmHg AAMI criterion. Thus, the model meets the defined regulatory standard.

6. Conclusions

This study advances non-invasive blood pressure estimation by systematically evaluating deep learning architectures, quantization strategies, and clinical validation protocols. The residual-enhanced network outperformed alternatives, achieving MAE values of 6.36, 8.35, and 6.90 mmHg for SBP, DBP, and MBP, respectively. It met and exceeded AAMI standards while maintaining stability under aggressive quantization. The model’s ability to preserve accuracy with a 90.71% reduction in size marks a significant advancement for embedded medical devices. The proposed quantization framework confirms that sophisticated deep learning models can operate efficiently in resource-constrained environments without compromising clinical precision. Achieving a 2.6 ms inference time while maintaining regulatory compliance represents a breakthrough for real-time monitoring applications.