Cardiovascular diseases (CVDs) continue to represent the leading cause of mortality worldwide, accounting for approximately 20.5 million deaths in 2021, which constitutes nearly one-third of all global fatalities [1]. In the European Union alone, CVDs were responsible for 32.4% of deaths during the same period [2]. Early detection, effective self-care, and continuous monitoring are critical strategies for mitigating complications and improving long-term patient outcomes [3, 4].

Electrocardiogram (ECG) continues to play a central role in first-line cardiac diagnostics, providing a non-invasive method to evaluate the heart’s electrical activity [5]. Standard 12-lead ECG recordings, typically performed during scheduled clinical check-ups [6], offer a comprehensive but transient snapshot of cardiac function. However, for patients with chronic heart conditions or those at elevated risk of ECG-detectable abnormalities, isolated measurements are insufficient to capture intermittent or evolving pathologies [7].

To address this limitation, portable ECG devices such as Holter monitors have been widely adopted to facilitate continuous cardiac monitoring over periods ranging from 24 h to several weeks [8]. Holter recordings provide valuable longitudinal data, capturing transient arrhythmias and anomalies that would otherwise go undetected during brief clinical assessments [9]. While long-duration ECG monitoring offers new opportunities, it also introduces specific challenges. Extended recordings generate massive volumes of data, which, while rich in diagnostic potential, are computationally demanding to store, process, and analyze [10, 11].

Historically, Holter monitors have operated as standalone systems, recording data locally for offline analysis [12]. Recent advances, however, have driven the evolution toward continuous wireless transmission of ECG data using wearable patches, smartwatches, and next-generation Holter devices [13, 14]. These cloud-connected systems enable near real-time analysis, often powered by artificial intelligence (AI) algorithms, providing scalable solutions for automated diagnostics and long-term health monitoring [15, 16].

Despite these benefits, cloud-based transmission introduces substantial trade-offs. Continuous data streaming requires significant bandwidth, increases power consumption, and raises concerns regarding data privacy and security. Moreover, reliance on network connectivity can lead to data transmission delays and intermittent loss of service, causing local storage bottlenecks during disruptions [16–18]. These constraints highlight the need for more efficient data handling strategies in wearable cardiac monitoring systems.

Optimizing the efficiency of continuous ECG monitoring requires the ability to preprocess and classify signals directly on the device. By identifying and discarding low-quality or clinically irrelevant segments at the source, it can help reduce transmission loads, potentially extend battery life, and enhance user privacy. Embedded artificial intelligence offers a promising approach to achieving this goal, enabling real-time, on-device ECG analysis that reduces reliance on external computational infrastructure.

Research in ECG biometrics has traditionally focused on security and identification applications, where ECG signals are used as unique physiological signatures to verify individual identity. In these approaches, distinctive features such as heartbeat intervals, waveform morphology, and rhythm patterns are extracted to establish identity with high confidence [39–45]. Studies have demonstrated the robustness of ECG-based biometrics across varying sensor types and acquisition conditions, supporting its potential for practical deployment.

Beyond security applications, the integration of ECG biometrics into healthcare diagnostics has recently emerged as a promising research direction. Shusterman and London [46] proposed a personalized ECG monitoring (pECG) framework that combines adaptive machine learning and distributed architectures to refine patient-specific baselines over time. Their system extracts minimal features locally while relying on server-based adaptation to handle evolving physiological changes, illustrating the benefits of personalization in continuous cardiac monitoring.

Similarly, Mangold et al. [47] conducted a large-scale study using over 970,000 ECG records from approximately 100,000 patients to evaluate the feasibility of ECG signals as biometric identifiers. Employing a Siamese neural network architecture, they achieved high matching accuracy and further demonstrated that single-lead ECGs can approach the performance of full 12-lead configurations. Their analysis of longitudinal data also highlighted the dynamic nature of ECG fingerprints, suggesting the importance of adaptive modeling for long-term applications.

Efforts to implement real-time ECG analysis on embedded hardware have also been reported. Raj and Ray [48] developed a microcontroller-based monitoring system capable of local feature extraction and arrhythmia detection. Their approach relied on classical signal processing methods, including the Discrete Cosine Stockwell Transform and support vector machines, to achieve efficient on-device classification with low power consumption.

Recent lightweight 1D CNNs for single-lead ECG have also been explored for arrhythmia detection. Rahman and Faezipour [49] proposed a compact time-domain CNN and argued that its computational footprint makes it suitable for embedded deployment. While their focus is population-level arrhythmia classification rather than biometric verification, the architectural theme—shallow temporal convolutions over short ECG windows—aligns with our motivation for compact models on constrained devices. In contrast, our study emphasizes a fully embedded, subject-specific pipeline with static int8 quantization, explicit memory/timing analysis on a Cortex-M7, and a two-stage flow with signal-quality gating.

Wang et al. [50] proposed an ECG biometric authentication framework based on self-supervised representation learning for IoT edge devices. Their approach employed a convolutional encoder trained via contrastive pre-training on largely unlabeled ECG data, followed by correlation-based identity verification using per-subject templates. Evaluated on the PTB Diagnostic ECG Database and cross-validated on external datasets such as MIT-BIH Arrhythmia and ECG-ID, their method achieved high authentication accuracy (≈ 99.1% on PTB-DB and ≤ 98.5% on unseen datasets), demonstrating strong generalization capabilities. To support embedded operation, they further implemented model pruning and 8-bit quantization on a Cortex-M4F microcontroller, reporting only marginal accuracy loss (0.5%) while reducing computational cost by approximately 37%. This work highlighted the potential of self-supervised feature extraction for low-power biometric systems and emphasized the scalability of global ECG encoders across multiple users. In contrast, the present work pursues a complementary objective: instead of learning a population-wide ECG representation for universal authentication, it investigates a personalized verification paradigm in which the model is trained to recognize a single individual’s cardiac signature directly on the device. The proposed framework integrates a compact, statically quantized CNN with an initial signal-quality assessment stage, enabling selective and fully autonomous operation on the embedded hardware.

Collectively, these studies demonstrate significant progress toward personalized, automated ECG analysis. They highlight the potential of ECG biometrics for continuous health monitoring, the importance of adapting to individual physiological baselines, and the feasibility of deploying classification systems on embedded platforms. However, challenges remain in integrating deep learning-based personalization into fully embedded systems with significantly reduced reliance on cloud infrastructure, particularly under the resource constraints typical of wearable Holter devices.

Real-time ECG biometric classification on embedded hardware was achieved through a structured workflow, addressing the specific challenges of operating within the computational and memory constraints of a resource-limited microcontroller.

This study utilized the PTB Diagnostic ECG Database [51]. A summary of key database attributes is provided in Table 1. The database contains high-resolution, multi-channel ECG recordings from 290 subjects, sampled at 1,000 Hz with 16-bit resolution, and includes conventional 12-lead recordings alongside Frank leads. Basic demographic and clinical annotations are also provided.

For model development and evaluation, we trained and evaluated subject-specific models for five subjects (IDs 180, 093, 233, 007, and 011). Subject 180, a 37-year-old healthy male with seven recordings (each approximately 10 min), served as a primary example due to the larger number of recordings, but all five subjects were included for training, validation, and benchmarking as outlined in the Methodology. This multi-subject setup ensured that training, validation, and testing were conducted across distinct recordings on a per-subject basis.

Only the aVL lead was used to simulate a Holter setup, reflecting the single-lead configuration of our target wearable device. Although data acquisition from our custom Holter system is planned for future work, the PTB data was used to establish initial feasibility. Signals were downsampled from 1,000 Hz to 500 Hz to reduce computational overhead while preserving critical features such as QRS complexes and ST-segment morphology. Each recording was segmented into non-overlapping 5-s windows and standardized to zero mean and unit variance prior to training.

The target platform for this study is a custom-developed wearable Holter device designed for single-channel ECG acquisition [54]. The device is depicted in Figure 1. It uses a three-electrode configuration (right arm, left arm, and ground) suitable for continuous ambulatory monitoring. The device integrates a high-performance STM32H743VI6 microcontroller, enabling real-time on-device neural network inference.

Key hardware specifications relevant to embedded AI deployment are summarized in Table 2. The Cortex-M7 processor, combined with 1 MB RAM and efficient DMA support, provides the necessary computational resources for running quantized convolutional models within tight power and memory constraints.

While the device supports modular upgrades such as multi-lead ECG acquisition, auxiliary sensing (IMU, barometer, magnetometer), and 2G/4G cellular communication, these features were not utilized in the present study. The current setup focuses solely on the future single-lead ECG acquisition and processing for proof-of-concept validation.

The proposed convolutional neural network (CNN) processes 5-s, single-channel ECG segments sampled at 500 Hz, corresponding to 2,500 time steps per input. The architecture consists of four convolutional layers followed by a fully connected classifier, progressively transforming raw ECG waveforms into an abstract representation suitable for classification. Although convolutional networks are often treated as ”black boxes,” their hierarchical structure allows them to learn increasingly complex patterns [55], as demonstrated by Figure 2, which visualizes the output of selected filters at different layers.

The first convolutional layer applies 32 filters with a kernel size of 125 and a stride of 8, effectively downsampling the signal while preserving its key morphological characteristics. This layer’s receptive field spans 250 ms, allowing it to process an entire QRS complex along with adjacent waveform components. Batch normalization is used to stabilize training, followed by a ReLU activation. A max-pooling operation with a kernel size of 4 and stride 4 further reduces the temporal resolution, outputting a feature map of shape (32, 78), where each feature encodes localized signal variations. As illustrated in Figure 3b, this stage predominantly highlights sharp transitions, which are characteristic of QRS complexes and other high-frequency elements.

The second convolutional layer refines these initial representations using 32 filters with a smaller kernel size of 3, a stride of 1, and padding of 1, ensuring that local signal morphology is preserved while allowing the network to detect subtle waveform variations. Another max-pooling operation (kernel size 2, stride 2) reduces the feature map to (32, 39). At this stage, as seen in Figure 3c, the extracted features still retain noticeable ECG morphology but become more invariant to small perturbations.

The third convolutional layer expands the network’s capacity by increasing the number of filters to 64 while maintaining a kernel size of 3 and a stride of 1. A subsequent max-pooling operation (kernel size 2, stride 2) further reduces the temporal dimension to (64, 19), distilling the most salient inter-beat patterns. Figure 3d demonstrates that this layer begins to focus more on rhythm-based variations rather than fine waveform details, aiding in subject-specific identification.

The final convolutional layer (Figure 3e) preserves the structural integrity of the extracted features while condensing them into a compact representation. Using 64 filters with a kernel size of 3 and an adaptive average pooling operation, it outputs a feature map of size (64,1), which serves as an abstract signature of the input ECG segment. Unlike earlier layers, which retain interpretable waveform-like structures, this stage generates highly distilled feature vectors that no longer resemble the raw input but encapsulate key distinguishing characteristics.

Following convolutional feature extraction, the output is flattened and passed through a fully connected classifier. The first dense layer maps the 64-element feature vector to 32 neurons with ReLU activation, followed by a final layer that applies a sigmoid activation function to produce a probability score in the range [0,1], representing the likelihood that the input belongs to a specific individual. The fully connected layers refine the high-level representations, reinforcing subject-specific characteristics while discarding redundant information. Visualization of the network's layers is depicted in Figure 4.

We trained and evaluated subject-specific models for five enrolled subjects (180, 093, 233, 007, 011). For each subject, training, validation, and testing sets were drawn from distinct recordings to avoid leakage. For Subject 180—which had seven recordings—we used four recordings for training (≈ 40 min), two for validation, and one for testing. For the remaining subjects, analogous per-subject splits were used; the segment-level Train/Test/Eval counts are summarized in Table 3. In all cases, evaluation used recordings unseen during training and validation.

Across subjects, the time-domain CNN consistently provided the strongest verification performance. Averaged over all five enrolled subjects, the time-domain CNN achieved 0.91 ± 0.05 PC-F1 and 0.96 ± 0.03 PR-AUC (Table 4). The best result was observed for Subject 180 with PC-F1 = 0.95 and PR-AUC = 0.99, indicating highly separable per-subject embeddings and a robust precision–recall trade-off. The comparatively lower performance on Subject 011 (PC-F1 = 0.83) likely reflects a combination of smaller training set size (29 training segments) and greater intra-recording variability. Notably, the narrow standard deviations for the time-domain CNN contrast with the larger variability of FFT-domain models (e.g., CNN FFT PC-F1 SD = 0.33), suggesting that spectral features are more sensitive to subject- and session-specific conditions.

In four out of five subjects, time-domain models clearly outperformed their FFT-based counterparts, consistent with the aggregate statistics in Table 4. An interesting exception is Subject 233, where the FFT-domain CNN reached PC-F1 = 0.91 and PR-AUC = 0.96, closely trailing the time-domain CNN (0.94/0.97). This suggests that, for certain individuals, frequency-encoded morphology may capture subject-specific traits that are nearly as discriminative as time-local features. Nevertheless, FFT-domain performance was generally weaker (e.g., Subjects 007 and 011), indicating that phase and temporal context remain important for stable verification across diverse recording conditions.

The lightweight time-domain CNN consistently surpassed MobileNet variants on this task. Notably, Subject 093 showed a pronounced gap for MobileNet in the time domain (PC-F1 = 0.42), while the same architecture performed substantially better in the FFT domain (0.75). On average (Table 4), MobileNet Time reached 0.70 ± 0.18 PC-F1, reflecting higher across-subject variability than the CNN Time, whereas MobileNet FFT averaged 0.46 ± 0.32 PC-F1. We hypothesize that depthwise separable convolutions may be more sensitive to per-subject normalization and limited data in the time domain, whereas their inductive bias transfers more gracefully to spectral representations. Classical SVMs in the FFT domain were occasionally competitive (e.g., Subject 180, PC-F1 = 0.68 vs. MobileNet FFT at 0.44), underscoring that simpler decision boundaries can work reasonably well when features emphasize stable, subject-specific spectral bands.

Performance correlated with the amount and diversity of training material per subject. Subject 180 benefited from the largest pool of recordings, allowing us to reserve multiple unseen records for validation and testing while maintaining a sizable training set. In contrast, Subject 011 exhibited the lowest CNN time-domain PC-F1 (0.83), which we attribute to both fewer training segments and potentially higher within-subject variability. This observation aligns with a broader principle in personalized biometrics: subject-specific models improve with increased coverage of day-to-day variation, electrode placement drift, and noise conditions.

Reported PC-F1 values reflect thresholds selected on per-subject validation sets to optimize F1. The high PR-AUC values (often ≥0.96) indicate that scores are well-calibrated for ranking, offering flexibility to tune operating points post-deployment—for example, targeting low false accept rates in clinical workflows while retaining acceptable true accept rates. In practice, we recommend setting subject-specific thresholds based on a short enrollment calibration to match task requirements (e.g., stricter verification vs. more permissive monitoring).

From an embedded deployment perspective, we find it encouraging that a compact time-domain CNN reliably outperforms heavier or more generic architectures while remaining deployable on a microcontroller. The consistent edge of time-domain features suggests that preserving temporal morphology—even in short, 5-s windows—provides strong identity cues. At the same time, the near-parity of FFT performance for Subject 233 is a useful reminder that some users may present more stable frequency-domain signatures, which could be exploited in hybrid or adaptive pipelines. For the few lower-performing cases, we expect that modest additions—beat-synchronous alignment, rhythm-aware pooling, or subject-tailored data augmentation (e.g., controlled baseline wander, mild heart-rate shifts)—would close much of the gap without sacrificing on-device efficiency.

Figure 4 illustrates the evolution of training and validation loss across 50 epochs. The validation loss exhibited a smooth and steady decline, reaching its lowest value of 0.0289 at epoch 35, indicating strong generalization ability. The training loss followed a similar downward trajectory.

Table 6 summarizes representative studies on ECG-based biometric identification and embedded ECG classification, focusing on reported performance, model size, and computational complexity. Deep learning approaches such as those by Agrawal et al. [39] and Mangold et al. [47] demonstrated the strong discriminative potential of ECG morphology for user identification, achieving accuracies exceeding 98%. However, these works relied on high-capacity neural networks trained and evaluated in GPU or cloud environments, without addressing the constraints of wearable or resource-limited devices.

Recent progress has begun to shift toward edge-oriented and lightweight implementations. Wang et al. [50] presented a self-supervised ECG authentication framework optimized for IoT edge sensors, employing model pruning and 8-bit quantization to reduce computational cost by approximately 37% on a Cortex-M4F microcontroller, while maintaining near 99% accuracy on the PTB database. Similarly, Raj and Ray [48] demonstrated real-time ECG processing on a low-power microcontroller using classical DSP and machine learning techniques.

In contrast, the present work represents a fully embedded and personalized ECG biometric system capable of performing both signal-quality assessment and identity verification in real time on an STM32H7 microcontroller. Despite its compact memory footprint of only 48 kB after 8-bit quantization, the proposed convolutional neural network achieved 94.68% classification accuracy and an F1-score of 94.51%, with a total inference time of approximately 1.35 s per 5 s ECG segment.

This study demonstrated the feasibility of deploying real-time neural network inference for biometric ECG classification directly on a resource-constrained Holter device. By integrating a quantized convolutional neural network onto an STM32H7 microcontroller and employing a two-stage inference pipeline with signal-quality assessment, the system achieves real-time performance with efficient computational resource usage. The implementation confirms that low-power wearable devices can support continuous ECG biometrics with reduced reliance on cloud infrastructure.

While the achieved results validate the technical viability of personalized on-device ECG classification, the findings also highlight the importance of broader data diversity for further refinement. Recordings acquired under similar conditions (e.g., resting or supine posture) may exhibit high morphological consistency, potentially limiting model generalization. Future work will therefore focus on longitudinal data collection across multiple postures, daily activities, and physiological states. This will enable the model to learn invariant, subject-specific cardiac representations and adapt to gradual physiological changes, reducing overfitting while expanding applicability toward early detection of pathological trends. Ultimately, the proposed framework represents a stepping stone toward fully personalized, self-contained, and adaptive cardiac monitoring in everyday life.