Resource-Constrained On-Chip AI Classifier for Beat-by-Beat Real-Time Arrhythmia Detection with an ECG Wearable System

Abstract

The electrocardiogram (ECG) is one of the vital physiological signals for human health. Lightweight neural network (NN) models integrated into a low-resource wearable device can benefit the user with a low-power, real-time edge computing system for continuous and daily monitoring. This work introduces a novel edge-computing wearable device for real-time beat-by-beat ECG arrhythmia classification. The proposed wearable integrates the light AI model into a 32-bit ARM^® Cortex^™-based custom printed circuit board (PCB). The work analyzes the performance of artificial neural network (ANN), convolutional neural network (CNN), and long short-term memory (LSTM) models for real-time wearable implementation. The wearable is capable of real-time QRS detection and feature extraction from raw ECG data. The QRS detection algorithm offers high reliability with a 99.5% F1 score and R-peak position error (RPE) of 6.3 ms for R-peak-to-R-peak intervals. The proposed method implements a combination of top time series, spectral, and signal-specific features for model development. Lightweight, pretrained models are deployed on the custom wearable and evaluated in real time using mock data from the MIT-BIH dataset. We propose an LSTM model that provides efficient performance over accuracy, inference latency, and memory consumption. The proposed model offers 98.1% accuracy, with 98.2% sensitivity and 99.5% specificity while testing in real time on the wearable. Real-time inferencing takes 20 ms, and the device consumes as low as 5.9 mA of power. The proposed method achieves efficient performance in real-time testing, which indicates the wearable can be effectively used for real-time continuous arrhythmia detection.

1. Introduction

Cardiac signals are vital indicators of the heart’s electrical activity and rhythm, aiding in the diagnosis of cardiovascular diseases (CVDs). An accurate electrocardiogram (ECG) analysis allows the early detection of arrhythmias, myocardial infarction, and other abnormalities, improving patient outcomes through timely intervention [1,2]. The importance of cardiac health monitoring has grown exponentially due to the rising global prevalence of cardiovascular conditions, which remain the leading cause of death worldwide. Recent statistics underline the gravity of CVDs, revealing that nearly 18 million individuals globally succumb annually to cardiac-related ailments, accounting for approximately 32% of all global deaths [3]. In the United States alone, CVDs continue to impose a substantial health and economic burden. The American Heart Association reports nearly 655,000 annual deaths attributed directly to heart disease, making it the leading cause of mortality nationwide [4]. Furthermore, the direct medical costs of CVDs in the U.S. are projected to escalate to over USD 749 billion annually by 2035, underscoring the urgent need for effective preventative and diagnostic strategies [5]. The severity of cardiac diseases is often exacerbated by a late diagnosis, significantly impacting patients’ quality of life and healthcare expenditure. Conditions such as congestive heart failure, arrhythmias, and coronary artery diseases typically require extensive and costly medical interventions, including hospitalization, surgical procedures, and long-term medication [6,7]. Early detection through daily cardiac monitoring can substantially mitigate these costs and improve survival rates. Continuous and remote monitoring technologies facilitate the early identification of abnormal cardiac patterns, enabling proactive treatment and potentially preventing acute cardiac events [8]. Therefore, innovations in wearable and home-based cardiac monitoring systems are becoming indispensable tools in reducing disease burden and healthcare costs associated with cardiovascular conditions.

Cardiac signals reveal the heart’s electrical patterns and rhythm, crucial for diagnosing CVDs. Precise ECG interpretation enables the early detection of arrhythmias, infarctions, and other issues, improving outcomes via prompt intervention [9]. Untreated arrhythmia can lead to certain long-lasting health conditions. Traditional clinical approaches to diagnosing and monitoring arrhythmias include 12-lead ECGs, ambulatory Holter monitoring, and implantable loop recorders [10,11]. Though clinically validated, their infrequent use misses brief, asymptomatic, or sudden arrhythmias, often delaying diagnosis by weeks or months [12,13]. Moreover, traditional clinical monitors rely on bulky equipment that limits mobility and require offline data analysis. In contrast, continuous real-time ECG monitoring can uncover hidden events and reduce the risk of worsening heart conditions [14,15,16].

ECG arrhythmia detection has greatly benefited from deep neural networks, including classical artificial neural networks (ANNs), convolutional neural networks (CNNs), and recurrent models like long short-term memory (LSTM) deployed on various platforms such as mobile devices, cloud servers, and edge hardware [17,18]. Smartphone-based applications can instantly detect abnormal heartbeats as they occur, alerting patients and doctors immediately, which is crucial for conditions that are episodic or asymptomatic [19,20]. This real-time inter-patient and intra-patient computation dramatically improves continuous long-term monitoring and the likelihood of capturing sporadic arrhythmias compared to a brief in-clinic ECG. Studies have validated that deep learning approaches can achieve high accuracy in detecting arrhythmias from single-lead wearables over large patient cohorts [21,22]. In general, deploying CNN/LSTM models on personal devices allows unobtrusive 24/7 monitoring, timely intervention, and improved management of cardiac conditions outside the hospital setting, which marks a significant advantage over intermittent clinical exams. However, most of the deep neural network (DNN) models are complex and require more computational resources. As a result, both mobile and cloud-based continuous ECG signal monitoring suffer from higher power consumption and latency [23,24]. Compared to lightweight models, these heavy models tend to run the processing unit for a longer period, resulting in increased power usage. Moreover, edge computing within wearables offers privacy, better synchronization, and reliability advantages as data are processed locally on the wearable, preserving confidentiality and reducing network dependency [25,26,27].

Deploying TinyML models directly on low-resource wearable SoCs enables the real-time detection of cardiac arrhythmias with several distinct advantages over smartphone or cloud-based AI. By performing inference on-device at the sensor source, network communication delays are eliminated. Unlike cloud systems that incur transmission latency, the TinyML model results in near-instant detection and prompt arrhythmia alerts [28,29]. With true edge computing, the system is no longer dependent on continuous connectivity, allowing continuous monitoring to continue offline without requiring a phone or wireless connection. Additionally, running the AI locally on an ultra-low-power MCU removes the need for constant Bluetooth/Wi-Fi data streaming to an external device, which significantly cuts down wireless energy overhead and preserves the wearable’s battery life [30,31]. These help the system keep sensitive health data on-device to safeguard patient privacy [32]. Achieving these benefits on a low-resource microcontroller ($\mu$C) unit is challenging, so model optimization techniques such as pruning, quantization, and knowledge distillation have become vital to compress and accelerate deep learning models for ultra-constrained hardware. These techniques convert floating-point parameters to lower-precision integers and eliminate unnecessary parameters with minimal impact on performance. As a result, these can substantially reduce memory and computing requirements, thereby enabling complex arrhythmia detection networks to run within the few kilobytes of RAM available on devices like wearable SoCs [33,34,35]. These benefits are achieved with low inference time and reduced model size, often with only a negligible impact on model accuracy [35,36,37,38]. However, most of the implementations have integrated the AI models into high-performance chips such as the Raspberry Pi 3 Model B+ (Raspberry Pi Foundation, Cambridge, UK), Jetson Nano Developer Kit (Nvidia, Santa Clara, CA, USA), and a field programmable gate array (FPGA)-based chip for better accuracy. These consume greater power and provide reduced battery life for a wearable device [39,40,41]. Moreover, the wearable can be heavy, which often reduces the user’s flexibility. Therefore, a practical wearable made with a low-resource $\mu$C unit of a few hundred kilobytes of memory can overcome these limitations.

In this study, three NN models, ANN, CNN, and LSTM, are integrated and tested on a custom wearable. The proposed wearable integrates the pretrained NN models with the device firmware to detect five different ECG beats: normal (N), supraventricular (SV), ventricular (V), fusion (F), and unknown (Q). The model is trained in Edge Impulse Studio^® and deployed as a C++ library. The custom wearable uses a 32-bit 64 MHz ARM^® Cortex^™ CPU and 2.4 GHz Bluetooth. The firmware is designed to capture raw ECG data for real-time processing. The device firmware can detect the QRS complex, extract features, and classify the cardiac beats in real time. The system can reliably transmit the classification result and raw data to a mobile application. The MIT-BIH dataset is used for training and testing. The study shows that an LSTM-based model integrated into the wearable can infer the cardiac beats with 98.1% accuracy. The proposed model also shows improvement in SV and V beat detection. The proposed LSTM model offers a 20 ms inferencing latency, peak RAM usage of 6.7 kB, and low power consumption at 5.9 mA.

2. Related Work

ECG morphological features play an important role in detecting abnormal cardiac beats. Several methods have been utilized to detect the QRS complex of an ECG wave. A Pan–Tompkins-based QRS detection is widely popular for its real-time implementation. The method analyzes the amplitude and slope of the cardiac signal and updates the threshold periodically [42]. Felix et al. proposed a low-complexity QRS-analyzing model with 99.6% sensitivity using a least squares parabola [43]. The use of machine learning algorithms is also available for QRS analysis. Rizwan et al. provide a technique with a discrete wavelength transform (DWT)-based support vector machine (SVM) for QRS detection [44]. Nguyen et al. offer a low-noise approach with a triangle template to match with R-peak slopes [45]. Another efficient real-time R-peak detection is proposed by Bae et al., where a pair of derivative filters is integrated into the process [46]. The work reaches a sensitivity and positivity of 99.8%. An implementation of QRS detection is shown in the mobile application platform by Neri et al. [47]. The study uses a modified Pan–Tompkins approach to detect the R peaks. Most of the studies use the MIT-BIH Arrhythmia database for validation of R peak detection [48]. Habibi et al. propose another real-time QRS detection that provides 99.69% and 99.60% of sensitivity and positive predictivity, respectively, upon validation with the MIT-BIH dataset [49]. The real-time QRS detection integrated into low-resource wearables can be used for further cardiac beat classification and analysis [50,51,52].

Arrhythmia classification using machine learning (ML) models has been an effective method [53]. Acharya et al. developed a nine-layer CNN model to classify five arrhythmia beats [54]. The work offered an overall accuracy of 94.3% while tested offline on a PC. To overcome the issues with the MIT-BIH imbalance dataset, a dynamic class weighting can be implemented. Huang et al. proposed a dynamic weighting process that updated the loss weights within each training cycle [55]. The model improved the accuracy for minority classes, providing an overall accuracy of ~98%. Although offline methods can offer in-depth classification and enhanced accuracy, such models can be bulky and complex for low-resource $\mu$C devices, limiting their ability to be integrated into real-time wearable devices. These indicate the importance of lightweight models that can easily fit within resource-constrained platforms.

Model quantization, pruning, and distillation techniques can offer an optimized model for low-resource $\mu$C devices. Ribeiro et al. implement an eight-bit quantized CNN model that performs the arrhythmia classification on a mobile application platform [56]. The quantization process reduces model size to only 93 kB and inference latency to 7.65 ms. However, high-power chips from mobile devices also suffer from more energy consumption, which affects the long-term signal monitoring. Kim et al. propose a wearable device that incorporates a lightweight CNN model for real-time beat classification with an overall accuracy of 87.1% [57]. Sabor et al. propose another $\mu$C-based compact CNN model that classifies 8 s chunks of ECG signals [52]. The model consumes ~50 kB of memory and feeds a raw ECG signal with the previous R-R interval as the input features. More complex models, e.g., recurrent LSTM models, are becoming popular in beat classification on resource-constrained chips. Katsaouni et al. propose a spiking neural network (SNN) architecture implemented on an ARM^® A53 chip [58]. Another example of complex model integration into constraint environments is a transformer model run on a Gap9 processor for arrhythmia classification [59]. Falaschetti et al. show a recursive neural network (RNN)-based LSTM model deployed on an ARM^® M4 chip for real-time classification [60]. Jeong et al. present a squeeze-and-excitation-block-based ResNet (SE-RESNET) adversarial learning on inter-patient ECG variability for arrhythmia classification [61]. Huang et al. propose an energy-aware neural architecture search (NAS) framework for arrhythmia classification by a distributed IoT [62]. The work offers an energy-efficient search and partitioning approach on ARM Cortex-M series chips for multiclass arrhythmia detection. Although complex models yield better accuracy, they often require more power or computational latency. Moreover, most of the works lack real-time morphological feature extraction and integration into the DNN input layer.

This work proposes a novel LSTM model integrated into a custom wearable for arrhythmia classification. The method incorporates a novel feature set combining time series, spectra, and morphological features of ECG. The wearable uses a combination of two mutex buffers, each storing 10 s of ECG chunk for feature extraction and classification. The wearable is capable of preprocessing and feature extraction in real time. Finally, the tiny LSTM model is run on-device to provide real-time classification. The hardware platform is analyzed for ANN, CNN, and LSTM-based DNN models. The system performance is measured in terms of latency, power, and memory consumption and compared with existing low-resource edge-computing-enabled wearables.

3. Materials and Methods

The wearable system incorporates an on-chip LSTM model for ECG beat classification. Figure 1 presents the overall system for the proposed method. The custom wearable contains an ECG analog front end (AFE) and is capable of classifying beats in real time. The firmware of the wearable includes preprocessing, feature extraction, and beat classification stages. In the preprocessing stage, each peak is detected for beat segmentation. The corresponding beats are further processed for feature extraction. Extracted features are fed into the NN model for beat-by-beat classification. The wearable can also transmit raw ECG and beat classification results via Bluetooth Low Energy (BLE) to a mobile application.

The hardware part contains an AFE circuit for ECG. The main component of the AFE is an AD8232-based (Analog Devices, Wilmington, MA, USA) instrumentation amplifier. The lead II ECG can be captured by placing electrodes around the chest area. The AFE circuit provides ECG filtering (0.5 Hz–41 Hz) and an amplification of 1100 (V/V) [63]. The wearable can capture cardiac data in practical scenarios from diverse participant groups [64]. The processing unit of the wearable contains a 32-bit 64 MHz ARM^® Cortex^™ M4 CPU and a 2.4 GHz Bluetooth radio. The system also contains a Lithium Polyimide (Lipo) rechargeable battery chip, MCP 73831 IC (Microchip Technology, Chandler, AZ, USA). The wearable is powered by a Lipo battery (Pkcell LP552530) of 350 mAhr and 3.7 V capacity. The BLE data speed is set to 1 Mbps for efficient and reliable data transmission.

Figure 2 depicts the overall framework of the proposed classification system. The dataset was divided into 60%/20%/20% for training, validation, and test sets, respectively. The model was trained under Edge Impulse Studio^®. For training, each NN frame size was set to 0.5 s of ECG data, corresponding to 180 samples for a 360 Hz sampling rate. The features were extracted for each beat and then passed into the classifier network. Three types of DNN models, ANN, CNN, and LSTM models, were tested for the proposed framework. At the end of training, the model was deployed as a C++ library. The C++ library contained all the source code and headers for the classification model. The library was integrated into the BLE chip. Each model training consisted of 50 epochs of training, along with a batch size of 32 and a learning rate of 0.005. An Adam optimizer was used with a learning rate of 0.005 and 1st- and 2nd-moment decay rate of 0.9 and 0.999, respectively. For model testing, the MIT-BIH test dataset was fed as mock data into the wearable device and tested in real time. The pretrained model was evaluated on a set of 21,892 previously unseen heartbeats. The test dataset was partitioned into arrays of 600 beats. Each of the 600 beat arrays was stored in the firmware to test the model’s performance. In total, the test dataset was segmented into 37 arrays to complete the model testing. An ECG block of 10 s was taken for the R peak detection. Upon detecting R peaks, a beat segmentation window of 0.5 s was selected for feature extraction and classification. The 0.5 s window centered on the detected R peak point and traversed 0.25 s forward and backward from the R peak.

3.1. Database

The MIT-BIH arrhythmia dataset was used to develop the pretrained classification models [48]. The dataset contains 48 records of ECG from 47 subjects. The dataset comes with beat-by-beat annotation for 15 types of arrhythmias.

Moreover, standard five-class labels are provided by the Association of the Advancement of Medical Instrumentation (AAMI). The dataset offers a substantial quantity of time series data that is suitable for DNN models. Our work incorporated the lead II ECG from the MIT-BIH dataset. The dataset was segmented into a total of 109,494 beats, and the data processing followed the 360 Hz sampling rate.

Table 1. Description of AAMI standard beat classes.

AAMI Class	Code	MIT-BIH Beat Description	Symbol
Non-ectopic beat	N	Normal Left bundle branch block Right bundle branch block Atrial escape Nodal premature	$•,\mathrm{N}$ L R e j
Supra-ventricular ectopic beat	SV	Atrial premature Aberrated atrial premature Nodal premature Supraventricular premature	A a J S
Ventricular ectopic beat	V	Premature ventricular Ventricular escape	V E
Fusion beat	F	Fusion of ventricular and normal beat	F
Unknown beat	Q	Paced Unclassifiable Paced and normal beat fusion	/ Q f

provides the mapping of the AAMI standard over the MIT-BIH dataset. The pretrained model classified the beats into one of the five classes: normal (N), supraventricular (SV), ventricular (V), fusion (F), and unknown (Q) beats.

3.2. Firmware Processing

The firmware design mainly has three parts: preprocessing, feature extraction, and classification. The mutex buffer technique is implemented in the firmware to ensure continuous data storage in real time. Each buffer can hold 10 s of ECG data. Once the buffer is full, it moves forward to the data processing stages. In the meantime, the alternative buffer keeps storing the incoming data. A buffer becomes free once the corresponding BLE transmission finishes. Figure 3 provides the flowchart for the overall firmware processing.

3.2.1. Preprocessing

The data preprocessing stage starts with a band-pass filter applied to the raw ECG signal. The filter has a cutoff frequency range from 2 Hz to 26 Hz, as most of the ECG power lies within that spectrum. The band-pass-filtered signal is further smoothed by a moving window filter of window size 21. The cleaned ECG data are then processed for beat detection.

3.2.2. R-Peak Detection

An adaptive threshold-based R-peak detection algorithm is applied to the smoothed ECG signal. The method follows the Pan–Tompkins approach. Figure 4 shows the signal processing blocks for QRS complex detection. The process involves a differentiator filter, the square of the signal, and a moving integrator. The R-peak threshold changes adaptively as shown by (1)–(3).

$$\begin{array}{cc}\hfill S{P}_k& =\alpha p_k+(1-\alpha )S{P}_{k-1}\hfill \end{array}$$

(1)

$$\begin{array}{cc}\hfill N{P}_k& =\alpha p_k+(1-\alpha )N{P}_{k-1}\hfill \end{array}$$

(2)

$$\begin{array}{cc}\hfill Th{r}_k& =N{P}_k+\beta (S{P}_k-N{P}_k)\hfill \end{array}$$

(3)

where $S{P}_k$ and $N{P}_k$ are the running signal and noise peaks, respectively. $p_k$ is the kth candidate peak in the moving window integrated signal. Both $S{P}_k$ and $N{P}_k$ are updated by blending a fraction, $\alpha$(0.12), of the new peak. The threshold is updated by (3), based on the signal and noise levels and by using the factor $\beta$(0.2), which controls how far above the noise floor we require a peak to be before calling it an R peak. Any peak is taken as an R peak if it is above the running threshold, followed by (4) and (5),

$$\begin{array}{c}\hfill n\in R\mathrm{if}m\left[n\right]\ge {\mathrm{Thr}}_k\end{array}$$

(4)

$$\begin{array}{c}\hfill n-n_{\mathrm{last}}\ge \mathrm{Refractory}\mathrm{Period}\end{array}$$

(5)

where $m\left[n\right]$ is the moving-window-integrated signal. A refractory period of 200 ms is maintained while searching for the R peaks. The P and T peaks are captured in a similar manner. The P peak is searched a minimum of 60 ms before the R peak index. The T-wave peak is identified by searching for the maximum at least 350 ms after the R-wave. The corresponding equations for P and T-wave detection is given by (6) and (7), respectively:

$$\begin{array}{cc}\hfill P_{\mathrm{mean}}& =\alpha y_{\mathrm{SM}}\left[i\right]+(1-\alpha )P_{\mathrm{mean}},\mathrm{if}{y}_{\mathrm{SM}}\left[i\right]\ge P_{\mathrm{mean}}\hfill \end{array}$$

(6)

$$\begin{array}{cc}\hfill T_{\mathrm{mean}}& =\alpha y_{\mathrm{SM}}\left[i\right]+(1-\alpha )T_{\mathrm{mean}},\mathrm{if}{y}_{\mathrm{SM}}\left[i\right]\ge T_{\mathrm{mean}}\hfill \end{array}$$

(7)

where $y_{\mathrm{SM}}\left[i\right]$ is the smoothed ECG signal, and $P_{\mathrm{mean}}$ and $T_{\mathrm{mean}}$ are only updated if there exists a peak greater than the mean value. Otherwise, there is no P and T peak detected. The process changes dynamically with respect to the changes in signal level.

3.2.3. Feature Extraction

To extract morphological features, the R peak detection algorithm is utilized to capture the ECG signal-specific data. A list of feature sets is provided in

Table 2. List of extracted features.

Time Series Features	Mean	RMS	SD	Kurtosis	Skewness
Spectral Features	PS 2–4 Hz	PS 4–6 Hz	PS 6–8 Hz	Spectral kurtosis	Spectral skewness
Morphological Features	R peak	T peak	P peak	RR interval	QRS duration
	PR interval	ST interval	QT interval	Minimum	Maximum

.

Ten morphological features include the ECG peaks and corresponding intervals. These signal-specific features provide better knowledge to the deep learning models. In addition, five separate time series and spectral features are also fed into the models. The time series features include the mean, root mean square (RMS), standard deviation (SD), kurtosis, and skewness, given by (8)–(12).

$$\begin{array}{cc}\hfill \mu & ={\displaystyle \frac{1}{N}}\sum _{n=0}^{N-1}x\left[n\right]\hfill \end{array}$$

(8)

$$\begin{array}{cc}\hfill \mathrm{RMS}& =\sqrt{{\displaystyle \frac{1}{N}}\sum _{n}x{\left[n\right]}^2}\hfill \end{array}$$

(9)

$$\begin{array}{cc}\hfill \mathrm{SD}& =\sqrt{{\displaystyle \frac{1}{N}}\sum _n{(x\left[n\right]-\mu )}^2}\hfill \end{array}$$

(10)

$$\begin{array}{cc}\hfill \mathrm{Skewness}& ={\displaystyle \frac{1}{N}}\sum _n{\left(\frac{x\left[n\right]-\mu }{SD}\right)}^3\hfill \end{array}$$

(11)

$$\begin{array}{cc}\hfill \mathrm{Kurtosis}& ={\displaystyle \frac{1}{N}}\sum _n{\left(\frac{x\left[n\right]-\mu }{SD}\right)}^4\hfill \end{array}$$

(12)

The spectral features are calculated by taking a discrete Fourier transform (DFT) on the time series data as shown in (13). The corresponding spectral power (PS) is calculated for three different frequency bands: 2–4 Hz, 4–6 Hz, and 6–8 Hz. The spectral power in each band is calculated by utilizing Equations (14) and (15). Moreover, skewness and kurtosis are calculated in the frequency domain by (18) and (19).

$$\begin{array}{cc}\hfill X\left[k\right]& =\sum _{n=0}^{L-1}x\left[n\right]e^{-j2\pi kn/L}\hfill \end{array}$$

(13)

$$\begin{array}{cc}\hfill P\left[k\right]& ={\left|X\left[k\right]\right|}^2\hfill \end{array}$$

(14)

$$\begin{array}{cc}\hfill P_{f_1,f_2}& =\sum _{\begin{array}{c}k:\\ f_1\le \frac{k{F}_s/D}{L}<f_2\end{array}}P\left[k\right]\hfill \end{array}$$

(15)

$$\begin{array}{cc}\hfill {\mu }_P& ={\displaystyle \frac{1}{K}}\sum _{k=0}^{K-1}P\left[k\right]\hfill \end{array}$$

(16)

$$\begin{array}{cc}\hfill {\sigma }_P^2& ={\displaystyle \frac{1}{K}}\sum _k{(P\left[k\right]-{\mu }_P)}^2\hfill \end{array}$$

(17)

$$\begin{array}{cc}\hfill {\mathrm{Skewness}}_{\mathrm{spec}}& ={\displaystyle \frac{1}{K}}\sum _k{\left(\frac{P\left[k\right]-{\mu }_P}{{\sigma }_P}\right)}^3\hfill \end{array}$$

(18)

$$\begin{array}{cc}\hfill {\mathrm{Kurtosis}}_{\mathrm{spec}}& ={\displaystyle \frac{1}{K}}\sum _k{\left(\frac{P\left[k\right]-{\mu }_P}{{\sigma }_P}\right)}^4\hfill \end{array}$$

(19)

Figure 5 presents the signal processing steps for time series and spectral feature extraction methods. The smoothed signal is decimated by a factor of 3 before applying the DFT. A 128-point DFT analysis is conducted over the decimated signal to calculate spectral powers for separate frequency bands. The process follows similar steps to those used in Edge Impulse Studio^® during training.

3.2.4. DNN Architecture

Three deep learning models were used for training and testing: ANN, CNN, and LSTM. Figure 6 shows the NN architecture for each model. The ANN model includes five successive fully connected dense layers (D), each with a ReLU activation, progressively learning higher-level combinations of the inputs. The input layer takes 20 features as input. The softmax layer transforms the network’s final activation values into a probability distribution across the target classes. The output layer provides the classification results into the five AAMI classes.

The CNN model adds three 1D convolutional layers (CL), each with a different filter size and width. The corresponding 1D maxpool layer downsamples by a factor of 2. A flatten layer follows to merge the pooled feature maps into a single vector. Two more dense layers are also added before the final classification. In the LSTM architecture, two stacked LSTM layers (LS) are employed, with dense layers placed both before and after them. LSTM layer 1 (LS1) captures forward dependencies and exposes the full sequence for the next layer. LSTM layer 2 (LS2) aggregates the sequence into a single vector. Both of the layers add a dropout layer for regularization.

Table 3. ANN, CNN, and LSTM model architectures.

ANN (D1, D2, D3, D4, D5)		CNN (CL1, CL2, CL3, D1, D2)		LSTM (D1, LS1, LS2, D2, D3)
Model	Units	Model	Filters/Units	Model	Units
A1	32, 32, 32, 32, 32	C1	32, 32, 32, 32, 32	L1	32, 32, 32, 32, 32
A2	32, 64, 64, 64, 64	C2	32, 64, 32, 32, 32	L2	64, 32, 32, 64, 64
A3	32, 32, 64, 64, 128	C3	64, 32, 32, 32, 32	L3	64, 128, 128, 64, 64
A4	32, 64, 64, 64, 128	C4	64, 64, 32, 32, 32	L4	64, 64, 64, 64, 64
A5	32, 64, 64, 128, 128	C5	128, 64, 32, 48, 24	L5	128, 32, 32, 128, 128
A6	64, 64, 64, 64, 64	C6	128, 64, 32, 32, 32	L6	128, 64, 64, 128, 128
A7	64, 64, 64, 128, 128	C7	128, 64, 32, 64, 32	L7	64, 128, 128, 64, 32
A8	64, 64, 128, 128, 128	C8	128, 64, 32, 128, 128	L8	128, 128, 128, 64, 32
A9	64, 128, 128, 128, 128	C9	128, 96, 64, 128, 128	L9	128, 128, 128, 64, 64
A10	128, 128, 128, 128, 128	C10	128, 128, 128, 128, 128	L10	128, 128, 128, 128, 128

presents how each model differs in layer depth and the number of units per layer. The abbreviation terms A, C, and L represent the different ANN, CNN, and LSTM model architectures, respectively.

4. Results

The pretrained DNN models were integrated into the firmware of the wearable device. The processing of the QRS complex from raw ECG is shown in Figure 7. Raw ECG signals were band-pass-filtered (2–26 Hz) and smoothed. Figure 7e denotes the detected R, P, and T peaks from the cardiac signal. The band-pass filter removed the baseline wander and improved the signal-to-noise ratio (SNR).

The performance of the R-peak detection process was validated by comparing the MIT-BIH dataset over 48 records. The MIT-BIH dataset provides the annotation for each R peak. For every record, the detected peaks are listed with true positive (TP), false positive (FP), and false negative (FN) values, respectively.

Table 4. MIT-BIH QRS detection and timing performance for each record.

Record		Actual		TP		FP		FN		Sn, %		Pr, %		F1, %		RPERMS (ms)		SDRR (ms)
100	201	2273	1963	2273	1959	3	3	0	4	100.00	99.80	99.87	99.85	99.93	99.82	1.51	9.21	1.86	15.19
101	202	1862	2136	1862	2103	7	8	3	33	99.84	98.46	99.63	99.62	99.73	99.03	1.13	1.80	5.00	6.91
102	203	2187	2980	2187	2945	3	30	0	35	100.00	98.83	99.86	98.99	99.93	98.91	4.62	12.40	14.37	18.34
103	205	2080	2656	2080	2651	7	7	4	5	99.81	99.81	99.66	99.74	99.74	99.77	1.27	1.76	2.86	5.39
104	207	2213	1860	2213	1853	32	5	16	7	99.28	99.62	98.57	99.73	98.93	99.68	9.16	5.52	18.32	14.19
105	208	2542	2955	2542	2930	37	7	26	25	98.99	99.15	98.57	99.76	98.78	99.46	4.02	39.82	16.30	35.82
106	209	1971	3005	1971	3004	18	6	56	1	97.24	99.97	99.10	99.80	98.16	99.88	10.53	0.81	15.82	3.24
107	210	2132	2650	2132	2641	2	8	5	9	99.77	99.66	99.91	99.70	99.84	99.68	2.64	3.34	7.82	9.59
108	212	1763	2748	1763	2946	79	0	0	2	100.00	99.93	95.71	100.00	97.81	99.97	34.99	1.55	56.35	4.42
109	213	2531	3251	2531	3249	9	5	1	2	99.96	99.94	99.65	99.85	99.80	99.89	3.97	6.87	6.05	13.25
111	214	2123	2262	2123	2257	39	11	1	5	99.95	99.78	98.20	99.51	99.07	99.65	4.89	5.97	9.09	13.11
112	215	2538	3363	2538	3362	4	4	1	1	99.96	99.97	99.84	99.88	99.90	99.93	1.54	2.61	5.29	6.09
113	217	1794	2208	1794	2201	5	5	1	7	99.94	99.68	99.72	99.77	99.83	99.73	1.29	9.46	3.32	15.57
114	219	1877	2154	1877	2146	20	4	2	8	99.89	99.63	98.95	99.81	99.42	99.72	18.02	2.92	15.51	7.36
115	220	1953	2048	1953	2046	4	0	0	2	100.00	99.90	99.80	100.00	99.90	99.95	0.88	1.11	3.08	2.45
116	221	2382	2427	2382	2422	9	2	30	5	98.76	99.79	99.62	99.92	99.19	99.86	3.97	8.16	9.40	12.45
117	222	1534	2483	1534	2461	18	6	1	22	99.93	99.11	98.84	99.76	99.38	99.43	6.50	1.99	5.92	6.24
118	223	2272	2605	2272	2604	18	8	6	1	99.74	99.96	99.21	99.69	99.47	99.83	15.52	3.59	13.33	6.18
119	228	1987	2053	1987	2037	6	25	0	16	100.00	99.22	99.70	98.79	99.85	99.00	5.38	11.94	5.81	17.22
121	230	1861	2256	1861	2255	9	6	2	1	99.89	99.96	99.52	99.73	99.71	99.85	2.60	1.40	7.35	5.99
122	231	2475	1571	2475	1570	2	10	1	1	99.96	99.94	99.92	99.37	99.94	99.65	1.61	2.06	1.90	11.00
123	232	1515	1780	1515	1763	23	12	3	7	99.80	99.60	98.50	99.32	99.15	99.46	1.58	1.48	8.27	5.18
124	233	1619	3079	1619	3072	8	4	0	7	100.00	99.77	99.51	99.87	99.75	99.82	3.27	9.00	9.06	14.90
200	234	2597	2753	2597	2751	7	5	4	2	99.85	99.93	99.73	99.82	99.79	99.87	16.68	0.78	16.90	4.39
Total: 48		109,494		109,123		550		371		99.7		99.5		99.6		Overall: 6.3		Overall: 10.7

provides the F1 score, precision (Pr), and sensitivity (Sn) for each of the 48 records. To measure the overall timing performance, the R-peak position error (RPE) and the standard deviation of RR intervals (SD_RR) were calculated for each record. As given by (20), the RPE_RMS denotes the root mean square of the RPE, calculated by the timing difference between the annotated peak (AP) and the detected peak (DP). The SD_RR is the standard deviation of all RR intervals for each record, given by (21). As depicted in Table 4, the proposed QRS detection algorithm detected 109,123 R peaks correctly out of 109,494 total actual beats. The obtained sensitivity was 99.7%. To measure the timing performance, RPE_RMS and SD_RR were determined, yielding 6.31 ms and 10.69 ms, respectively.

$${\mathrm{RPE}}_{\mathrm{RMS}}=\sqrt{{\displaystyle \frac{1}{\mathrm{TP}}}\sum _{i=1}^{\mathrm{TP}}{\left(A{P}_i-D{P}_i\right)}^2}$$

(20)

$${\mathrm{SD}}_{\mathrm{RR}}=\sqrt{{\displaystyle \frac{1}{N_{\mathrm{RR}}}}\sum _{i=1}^{N_{\mathrm{RR}}}{(R{R}_i-\overline{RR})}^2}$$

(21)

Figure 8 presents the performance of the real-time time series feature extraction block. The raw ECG data were preprocessed with the filtering stages and passed to the feature extraction block. To measure the effectiveness of the feature extraction performance, the extracted features from the proposed method were compared to the reference values. The reference values were the extracted features from Edge Impulse Studio^®. The studio provided the beat-by-beat extracted feature data for both training and testing sets. In total, the feature sets were compared for 109,494 ECG beats. As shown in Figure 8, the proposed extracted time series features correlated strongly with the reference values. The correlation coefficient (R²) of the mean, root mean square (RMS), and standard deviation (SD) was 0.99, whereas the R² value of 0.98 slightly differed for kurtosis and skewness. Overall, the time series feature extraction block showed effective and reliable performance and could be integrated into the input layer of the DNN model.

The spectral features were also extracted in real time, and the proposed method was validated against the reference values from Edge Impulse Studio^®. The R² values varied from 0.92 to 0.94 for spectral power values over 2–4 Hz, 4–6 Hz, and 6–8 Hz as presented in Figure 9. These frequency ranges were chosen as most of the ECG signal power lies within those ranges. The proposed algorithm offered good correlation with the power spectral features. The ECG spectral kurtosis provided the lowest correlation of 0.83. The lower correlation could be caused by the differences in the implementation of the filter functions.

Figure 10 presents the performance comparison of the pretrained DNN models for the MIT-BIH test dataset. The test data were embedded as mock data into the wearable device for that performance analysis. The test performance values were analyzed for different feature combinations: time series, spectral, and combined features. Four performance metrics, F1 score, precision, sensitivity, and accuracy, were compared to determine the optimized model. The model numbers correspond to the description provided in Table 3. Figure 10a depicts the ANN performance for time series, spectral, and combined features. As the ANN model size increased, the model test performance improved for spectral feature sets compared to time series and combined feature sets. Similarly, for the CNN model, both spectral and combined feature sets tended to offer better performance over time-series-based features as shown in Figure 10b. The CNN model outperformed the ANN model in terms of classifying the ECG beats. Overall, the LSTM models gave the best performance as the model size increased. As shown in Figure 10c, the LSTM models with combined features provided the best performance metrics.

The performance of the real-time wearable system was also validated over two major metrics: latency and memory consumption. Figure 11 presents these performance metrics for different feature-based DNN models. The metrics were calculated using a real-time analysis of the mock test dataset from the firmware. Figure 11a depicts the ANN model performance for different feature sets. The ANN model offered an inference latency between 3 and 5 ms when tested. Among all three DNN models, the ANN offered the lowest memory consumption. Similarly, Figure 11b presents the CNN model performance. The CNN models offered a range of latencies between 3 and 10 ms for all types of feature sets. The CNN models tended to consume between 2 and 11 kB of memory as the model size increased. Figure 11c shows the LSTM model performance where the model’s inference latency varied in a range from 5 to 36 ms. The memory consumption was well within the range of 5 to 10 kB. Approximately 178 kB of RAM was dedicated to the complete real-time ECG processing stack, including peripheral drivers, data buffers, and the embedded TinyML model.

The optimized models were calculated by using a Pareto analysis over all models as depicted in Figure 12. To find the optimized models, three performance comparisons were carried out for all DNN models: accuracy vs. latency, accuracy vs. memory, and memory vs. latency. The cost function was set according to the resources available in the wearable device. For the accuracy cost function, a threshold of 96% accuracy was set. For latency, a threshold of 50 ms was set; beat-by-beat ECG detection could reliably wrap around that time constraint. The memory consumption threshold was set to 40 kB, considering that the consumption fit within the actual RAM size of 256 kB. As shown in Figure 12, the LSTM models were mostly common in all three Pareto sets. The LSTM model offered better accuracy compared to ANN and CNN models and fit well within the memory and time constraints. By capturing the intricate temporal dependencies in ECG signals over extended sequences, the LSTM could detect the subtle waveform variations critical for achieving more accurate beat identification. The LSTM model L3 was one of the best optimized models among all DNN models. The L3 model offered 98.09% accuracy with 20 ms of inferencing latency and 6.7 kB of memory consumption.

Table 5. The performance metrics of the pretrained L3 model on the test dataset implemented on SoC firmware.

	N	SV	V	F	Q
N	17,798	86	109	80	45
SV	21	516	18	1	0
V	26	3	1410	5	3
F	2	0	2	158	0
Q	7	3	4	0	1594

provides the L3 model performance for the test dataset. The table provides the confusion matrix for the AAMI classes. Each row in the table represents the true classes, and the columns represent the predicted classes. For the test set, the overall accuracy of the L3 model was 98.09%. Overall, there were 21,892 ECG beats in the test dataset. The wearable could successfully detect 17,746, 516, 1410, 158, and 1594 beats, respectively, for the N, SV, V, F, and Q beat classes.

5. Discussion

Table 6. QRS detection and timing performance comparison.

Reference	Actual Beats	TP	FN	FP	Sn%	Pr%	F1%	RPErms (ms)	SDRR (ms)
This Work	109,494	109,123	371	550	99.6	99.50	99.5	6.31	10.69
[52]	109,494	109,359	126	104	99.89	99.91	99.796	7.94	18.55
[51]	75,052	73,647	-	-	-	-	-	-	-
[50]	109,996	-	-	-	99.81	99.88	99.69	12.2	44.6
[49]	109,985	-	-	-	99.56	99.05	99.29	-	-
[47]	103,724	100,135	3,380	144	96.80	99.83	96.66	-	-
[46]	109,510	109,295	215	214	99.80	99.80	99.61	-	-
[45]	109,494	109,270	224	314	99.80	99.71	99.51	-	-
[44]	73,515	72,224	46	319	99.94	99.58	99.76	-	-
[43]	109,490	-	-	-	99.66	99.66	-	-	-
[42]	109,494	108,611	883	380	99.12	99.62	98.70	10.11	280.2

provides a comparison between the proposed QRS detection algorithm and existing work. As shown in the table, our proposed method obtains a 99.66% accuracy. The F1, precision, and sensitivity performance are in the range of 99%, which makes it a reliable approach. The proposed method offers an effective detection rate that is in close proximity with that of the existing literature. Furthermore, the proposed model provides superior timing performance in detecting the R peaks. The RPE_rms and SD_RR results outperform the work by [42,50,52].

Table 7. Comparison of arrhythmia classification systems across hardware platforms.

Reference	Classes	Records and Beats	Method	HW Platform	System Performance
[54]	5	Records: 48, beats: 100,062	Train/test (%): 50/50, 1D-CNN	–	Acc: 98.56, Sn: 97.68, Sp: 99.5
[55]	5	Records: 30, beats: 2520	Train/test (%): 80/20, Fast compression DNN	–	Acc: 95.16
[53]	6	Records: 30, beats: 10,502	Train/test (%): 75/25, DWT-based analysis	–	Acc: 99.99, Sn: 99.99, Sp: 99.97
[61]	4	Records: 22, beats: 50,948	SE-RESNET		Acc: 88.4, F1: 87.4
[57]	5	Records: 40	Train/test (%): 50/50, Wavelet-based QCNN	ARM M4	Model size: 25 kB, Tinf: 172.3 ms, Acc: 95.1
[56]	5	Records: 48, beats: 100,062	Train/test (%): 50/50, QCNN	ARM A55	Acc: 96.2, Sn: 98.5, Sp: 99.8, Tinf: 7.65 ms, Pavg: 765 mW
[52]	5	Records: 48, beats: 109,494	Train/test (%): 50/50, Adaptive QRS, CNN	ARM M4 168 MHz	Tproc: 46.61 s/30 min, Acc: 96.97, Sn: 95.1, Sp: 99.1
[58]	4	Records: 48	Train/test (%): 50/50, spike timing-based NN	ARM A53	Acc: 97.9, Pavg: 1.78 $\mu$J/beat
[59]	5	Records: 48 beats: 100,703	Train/val/test (%): 70/10/20, Transformer model	GAP9, 49 kB RAM	Acc: 98.97, Pavg: 0.09 mJ/beat, Tinf: 4.28 ms
[60]	5	Records: 48, beats: 100,062	Train/val/test (%): 75/5/20, RNN	ARM M4 80 MHz	Acc: 90, Tinf: 150 ms
[62]	5	Records: 48	Train/val/test (%): 70/15/15, energy-aware NAS	ARM M4 80 MHz	Acc: 89.7, Pavg: 4.85 mJ
This work	5	Records: 48, beats: 109,494	Train/val/test (%): 60/20/20, LSTM	ARM M4 64 MHz	Acc: 98.09, Sn: 98.2, Sp: 99.5, Pavg: 17.7 mW, Tinf: 20 ms

Acc = accuracy, T_inf = inference time, T_proc = processing time, P_avg = average power.

presents the performance of the hardware–firmware co-design for different wearable systems. The proposed method classifies into five AAMI classes, whereas four- and six-class classifications are also seen in [53,58,61]. Most of the existing work utilizes all 48 records from the MIT-BIH dataset for training and testing purposes. The quantized CNN (QCNN) models are commonly implemented in low-resource wearable devices. As shown in the table, a wide range of hardware (HW) platforms support edge computing. Different series of ARM^® M4 chips are popular for lightweight model classification. An energy-efficient AI model using the NAS framework is exploited in the ARM^® M4 80 MHz chip for low-power applications. The proposed wearable utilizes an ARM^® M4 64 MHz chip. The achieved inferencing latency is 20 ms, which is superior to [57,60]. The low-resource chip offers an average of power consumption during active periods as low as 5.9 mA. With this power consumption, the device would last for approximately 59 h. Moreover, the overall accuracy of the system outperforms the inferencing performance by [55,56,57,58,59,60,61,62]. Because of the imbalanced dataset, certain arrhythmic beats like SV and F are less understood by the model, as fewer beats are available. In the future, we will focus on data augmentation so that the number of underrepresented arrhythmia beats increases during model training and testing.

6. Conclusions

A low-power QRS detection and arrhythmia classification algorithm using an LSTM model was proposed in this paper. The adaptive thresholding approach in the QRS detection algorithm offered reliable performance in detecting the R peaks. The proposed method could accurately detect 99.7% of R peaks when tested on the MIT-BIH dataset. The algorithm offered an efficient R-peak location estimation, providing an error of only 6.31 ms RPE. The work proposed a combination of time series, frequency domain, and ECG morphological features for classifying the cardiac beats into five different arrhythmia classes. A lightweight LSTM model was proposed that provided optimized performance over 90 different DNN models. The optimized model was determined by comparing the model’s accuracy, inference latency, and memory consumption over the test dataset. The proposed model offered an overall accuracy of 98.09% for beat-by-beat classification. Furthermore, the model enhanced the detection accuracy for SV and V beats that have a limited dataset. The lightweight model was integrated into a low-resource ARM^® M4 chip. Because the model fit into a resource-constrained environment, it only used 17.7 mW of power and required 20 ms of inference time. These make the wearable suitable for real-time and long-term arrhythmia monitoring and classification. The current model inference takes 20 ms, which can be improved further. The footprint of the current model can be further reduced to facilitate the integration of additional external sensors into the system. The current work did not include testing on real patients. Future work will continue to integrate other cardiac datasets and more complex models with reduced latency in resource-constrained chips. We will also focus on model testing on clinical cardiac patients. Furthermore, we will work on adding multimodal sensor data to improve cardiac arrhythmia and other cardio-pulmonary disease detection.