Deep learning based cardiac disorder classification and user authentication for smart healthcare system using ECG signals

Convolution neural networks

Deep learning, a recent innovation in artificial intelligence, demonstrates near-human proficiency in many applications such as autonomous medical diagnosis, speech-to-text translation, feature detection, and classification. The CNN is a type of DL algorithm that can take in information, add biases and weights to different feedback features, and then discriminate among them (Albawi, Mohammed & Al-Zawi, 2017). A traditional CNN utilized two-dimensional input, including pictures and videos. As a result, it became commonly known as two-dimensional CNN. A more recent development is the 1-dimensional CNN, a two-dimensional CNN revision. The ECG records supplied the CNN model utilized during extracting features. Six layers make up this structure and alternate among 1-dimensional convolution and 1D max-pooling stages. A sub-CNN algorithm was developed for several sessions based on the latest findings. Enhancing the system depth and the specifications of every level in this procedure may enhance the detection accuracy, whereas the total effect is enhanced.

A feature extraction method may incorporate down sampling and convolutional layers. A kind of fuzzy filter known as convolution can enhance information quality while reducing distortion. The top element’s extracted features are convolutional with the current gradient convolution kernel (Mendieta et al., 2019). An equation then returns the outcomes of the convolutional calculations. Every generated vector provided as a target is compared against the inaccuracy using the given equations. If it is fulfilled, the learning is finished, and the numbers and limits are stored in the interim. The classifiers are created under the assumption that each weight is stable. Figure 3 shows the schematic representation of the proposed CNN network.

This CNN is precisely developed and methodically tuned through extensive testing to attain optimal performance in data classification. The model architecture comprises three 1D convolutional layers, each succeeded by batch normalization and max-pooling layers, to extract hierarchical features from the input data. Subsequent to these layers, there exists a flattening layer, followed by two completely connected dense layers using dropout regularization, culminating in a softmax layer for multi-class classification. The architecture comprises an input layer, succeeded by a Conv1D layer featuring 64 filters and a kernel size of 6, employing ReLU activation, batch normalization, and max pooling with a stride of (3, 2). This block is reiterated two additional times. The output features are further flattened and processed through two thick layers, each including 64 units with ReLU activation, accompanied by dropout layers at a rate of 0.2 to mitigate overfitting. The final dense layer contains five units and utilizes a softmax activation function for classification. Each of the five output units corresponds to a specific class in the dataset, reflecting separate categories that the model is designed to differentiate. The softmax function transforms raw output scores into probabilities, with each unit’s output representing the probability that the input sample is associated with the corresponding class. The class with the greatest likelihood is chosen as the definitive forecast. The model employs the Adam optimizer in conjunction with the categorical cross-entropy loss function, facilitating steady and efficient learning.

A grid search approach is used for hyperparameter optimization to enhance the model’s performance. This systematic method examines several values for essential hyperparameters, including the quantity of convolutional filters ([32, 64, 128]), kernel dimensions ([3, 5, 6]), and dropout rates ([0.2, 0.3, 0.5]). Every combination is assessed by 5-fold cross-validation. The efficacy of each model configuration is evaluated using validation accuracy and categorical cross-entropy loss, aiming to identify the combination that attains the highest validation accuracy and the lowest validation loss. Of the evaluated configurations, the model using 64 filters, a kernel size of 6, and a dropout rate of 0.2 demonstrates optimal performance. This configuration has been chosen for the final implementation. The tuning technique greatly enhances the model’s generalization and robustness. Additionally, early halting is employed during training, contingent upon validation loss, to prevent overfitting. Although learning rate scheduling is examined, the default learning rate of the Adam optimizer is adequate for consistent convergence. This experimental process of parameter tuning and layer tweaking is essential for enhancing the CNN model’s classification accuracy.

The CNN was configured with the following layer: Inputs $\to$ Convolution-1D (64 filters, map size 6, Relu) $\to$ Batch Normalization $\to$ Max Pool Layer (stride 3, 2) $\to$ Convolution-1D (64 filters, map size 6, Relu) $\to$ Batch Normalization $\to$ Max Pool Layer (stride 3, 2) $\to$ Convolution-1D (64 filters, map size 6, Relu) $\to$ Batch Normalization $\to$ Max Pool Layer (stride 3, 2) Flatten Layer $\to$ Dense (64 units, Relu) $\to$ Dropout (0.2) $\to$ Dense (64 units, Relu) $\to$ Dropout (0.2) $\to$ Dense (5 units, Relu) $\to$ SoftMax. The Adam optimization technique with categorical cross entropy produces better outcomes. The output of the convolution layer:

(1) $$X_J^I=f\left(\sum _{i\in M_j}{X}_J^{I-1}\ast W_{ij}^I+b_J^I\right).$$

The information shared between hidden layers is represented by:

(2) $$X_J^I=f\left(B_j^I\mathrm{down}\left(X_j^{I-1}\right)+b_j^I\right)$$the output are calculated by decreasing the error by the following equation:

(3) $$E=\frac{1}{2}\sum _{k=0}^{n-1}{\left(d_k-y_k\right)}^2\le \epsilon .$$

Long short-term memory

Long short-term memory is a recurrent neural network that can comprehend and predict consecutive information. RNN does have a limited ability to sustain long-term memory. The LSTM was created; as a result to address this problem by incorporating gating operations and a storage cell that can maintain its configuration over the duration (Sainath et al., 2015). Figure 4 shows the proposed LSTM model. A memory location, a recall gateway, an initialization vector, and an update gate are the essential parts of the LSTM. The activated sigmoid function manages the input gate’s output vector. We provide an enhanced LSTM that is specially made for identifying ECG signals to help solve the issue. An exponent is incorporated into the LSTM in this method to facilitate efficient data distribution and enhance interpretation. The formulas for the current block, potential cell state, and LSTM output are given:

(4) $${\tilde{c}}_t=\tanh \left(w_c\left[h_{t-1},x_t\right]+b_c\right)$$

(5) $$c_t=f_t\ast c_{t-1}+i_t\ast {\tilde{c}}_t$$

(6) $$h_t=O_t\ast \tanh \left(c^t\right).$$

Equations (4) to (5) delineate the internal mechanisms of a LSTM unit, which is extensively employed for simulating sequential data. In Eq. (4), the prospective cell state.

The calculation involves applying a tanh activation function to a linear transformation of the concatenated prior hidden state $h_{t-1}$ and the current input $x_t$. This candidate cell state serves as a prospective enhancement to the memory. Equation (5) modifies the current cell state $c_t$ by integrating the prior cell state $c_{t-1}$, adjusted by the forget gate $f_t$, and the candidate cell state ${\tilde{c}}_t$. t, modulated by the input gate. This mechanism enables the LSTM to selectively preserve pertinent information from the past and assimilate fresh input. Ultimately, Eq. (6) establishes the current hidden state, which functions as the output of the LSTM unit at time step t. The concealed state is derived by applying the output gate $O_t$ to the tanh activation of the present cell state. The gated mechanisms of the LSTM facilitate the management of long-range dependencies and temporal dynamics, rendering it particularly adept for tasks like ECG signal classification and authentication.

In order to see if the outcomes were notably superior to those of other networks, we introduced a long short-term memory layer to the most current optimization technique. The network’s layer was configured as follows: LSTM Layer (16 filters) $\to$, Dropout (0.2 rates) $\to$ LSTM Layer (32 filters) $\to$ Dropout Layer (0.2 rates) $\to$ LSTM Layer (64 filters) $\to$ Dropout Layer (0.1 rates) $\to$ Flatten Layer $\to$ Dense Layers. The selection of 64 filters and a kernel size of 6 was determined through empirical testing and literature demonstrating their efficacy in detecting pertinent temporal patterns in ECG signals. Batch normalization was employed to stabilize and expedite training, whilst max pooling facilitated dimensionality reduction and computational efficiency. Dropout layers with a rate of 0.2 were utilized to mitigate overfitting. The model was optimized with Adam and categorical cross-entropy, resulting in improved convergence and classification accuracy. These decisions together enhanced performance in multi-class ECG classification tasks. The Adam optimization technique with categorical cross-entropy yields superior outcomes.

The framework incorporates LSTM layers because of their proficiency in capturing long-range relationships in sequential data. LSTMs are adept at handling tasks with temporal dynamics due to their ability to save knowledge throughout prolonged time intervals through internal memory cells and gating processes. The architecture employs layered LSTM layers to augment the model’s ability to learn intricate temporal characteristics. This study selects LSTM networks above gated recurrent units (GRUs), despite the latter’s computational efficiency and reduced parameters, due to LSTMs’ superior performance on datasets characterized by extended and complex temporal dependencies. LSTMs provide enhanced control over information flow via distinct input, output, and forget gates, which is beneficial for modeling intricate temporal patterns. The preference for LSTM over GRU is substantiated by preliminary experiments, which demonstrate that LSTM models regularly exceed GRU models in validation accuracy and loss, rendering them a more appropriate option for the specified classification task.

The initial procedures used to analyze and evaluate the ECG data are preprocessing and normalization. This step’s objective is to acquire information from various Electrocardiogram systems, which comprises amplitude lowering, frequencies homogeneity, and static component reduction, as well as from various patients, which includes signaling scope standardization. Two preprocessing techniques, the MinMax scaler and the standard scaler are used in this research to increase the classifier accuracy of the data. The application of these strategies greatly enhances the capability of the models.

MinMax scaling and Standard scaling are preferred over alternative methods, such as Robust scaler or MaxAbs scaler, due to their capacity to enhance model performance in particular scenarios (Sinsomboonthong, 2022). The MinMax scaler is especially efficient when the model necessitates features to be within a specified range, often [0, 1], which is essential for algorithms like as neural networks, where activation functions (e.g., ReLU or sigmoid) operate optimally when inputs are confined to a limited range. MinMax scaling standardizes the data within a specified range, preventing features with bigger numerical values from overshadowing the model’s learning process and ensuring that all features contribute equitably to the model’s performance. Moreover, it facilitates expedited convergence by ensuring that the gradient steps in optimization are more uniform across all characteristics, hence diminishing the probability of sluggish or inadequate convergence.

Conversely, the Standard scaler is employed when the model presumes that the data adheres to a Gaussian (normal) distribution, a prevalent assumption in algorithms like logistic regression and support vector machines (Thara, Prema Sudha & Xiong, 2019). This scaler normalizes the data to achieve a mean of zero and a variance of one, hence standardizing the features. This transformation is especially advantageous when features exhibit varying scales or when certain features contain outliers. Standard scaling alleviates the influence of outliers by centering the mean and normalizing the features, resulting in enhanced optimization stability. Moreover, numerous machine learning algorithms exhibit enhanced performance when the data is centered around zero and normalized, as this mitigates the risk of characteristics with greater volatility overshadowing the model’s training process.

Other scalers, such as the Robust scaler, effectively manage outliers by scaling according to the interquartile range; nevertheless, they are primarily appropriate for datasets with substantial outliers that deviate from normal distribution assumptions (Napier et al., 2024). Conversely, MinMax scaling and Standard scaling are typically more effective for datasets where characteristics are either already confined to a specific range or approximately adhere to a normal distribution. This analysis indicates that both scalers consistently enhance model accuracy and training stability across various classifiers, illustrating their optimal balance between data normalization and model performance.

Database description

The PTB Diagnostic (Bousseljot, Kreiseler & Schnabel, 1995) and MIT-BIH Arrhythmia (Moody & Mark, 2001) ECG Collection, two well-known datasets in heartbeat classifications, are used to create two sets of beating signals that make up this database. These databases include enough samples to allow for the development of a neural network presented in Table 1. With the use of artificial neural network models, this information has been utilized to investigate the categorization of heartbeats and examine some of the training possibilities. For both the usual case and cases afflicted by various rhythms and infarction, the signals correlate to the ECG forms of heartbeats. Each segment of these signals, which has undergone preprocessing and segmentation, corresponds to a pulse.

The selection of these datasets over others relies on many critical criteria: clinical diversity, data quality, annotation accuracy, and extensive utilization in benchmarking heartbeat classification models. In contrast to smaller or less diverse ECG datasets, PTB and MIT-BIH provide more therapeutically pertinent scenarios and have been extensively utilized in prior research, facilitating comparative analysis and repeatability. Moreover, both datasets are publically accessible and expertly annotated, facilitating consistent and significant model training. The ECG signals in these databases are preprocessed and split into discrete pulses, each denoting a heartbeat classed as either normal or abnormal. This segmentation enables the model to concentrate on specific waveform characteristics, improving its capacity to discern nuanced variations linked to various heart diseases. Consequently, these datasets provide a solid basis for examining heartbeat categorization efficacy across various cardiac health conditions.

While the PTB Diagnostic and MIT-BIH Arrhythmia databases offer numerous benefits, it is crucial to recognize certain intrinsic limitations that may affect the generalizability of the study’s results. Both datasets derive from distinct clinical and demographic populations, with the PTB dataset predominantly consisting of myocardial infarction patients and the MIT-BIH dataset featuring a restricted representation of certain arrhythmia types and underrepresented demographics. This may introduce bias related to age, gender, and ethnic diversity, thereby constraining the model’s efficacy in more heterogeneous real-world environments. Moreover, despite the meticulous annotation of the signals, there exists a potential for subjective labeling discrepancies or annotation inaccuracies, especially in marginal or ambiguous instances. These parameters may influence the classifier’s capacity to generalize across diverse patient populations or unexpected signal anomalies. Future research may resolve these issues by integrating more diverse, multicenter datasets or employing domain adaptation methods to reduce dataset-specific biases and improve model resilience.

Feature extraction

Electrocardiogram signals that indicate the electrical and muscular activity of the heart are used to define cardiac muscle activity. ECG signals were chosen for this project due to technological limitations and their well-known waveform, which allows for improved classification and recognition. By attaching electrodes to the skin to show the primary pulses P, QRS, and T in a normal situation, the electrocardiogram symbolizes the assessment of the heart’s electrical activity as a response to time, evaluating the cardiovascular health of the atrial repolarization, the ventricles depolarization and ventricular repolarization of the myocardium. Figure 5 shows the ECG signal of normal and abnormal persons.

This study prioritizes the PQRST wave components due to their essential function in delineating the heart’s normal and pathological states. Each element of the ECG waveform signifies a specific electrical occurrence in the cardiac cycle: the P wave denotes atrial depolarization, the QRS complex signifies ventricular depolarization, and the T wave represents ventricular repolarization. These regions contain essential diagnostic information, and deviations in their shape, length, or amplitude frequently associate with particular cardiac conditions such as arrhythmias, ischemia, and myocardial infarction. Concentrating on these critical parts enhances the clinical significance of the feature extraction process, enabling the model to acquire features that are both discriminative and physiologically pertinent. Moreover, emphasizing the PQRST waves improves the model’s interpretability, as decisions are grounded in comprehensible cardiac events. This focused methodology diminishes extraneous noise and irrelevant data from the signal, hence enhancing the overall precision and resilience of the classification model. The decision to prioritize PQRST characteristics is based on clinical relevance and conforms to optimal methodologies in ECG signal processing and machine learning applications for cardiovascular diagnosis.

The six significant waveforms identified and separated are p, Q, R, S, T, and U. They all represent the electrical activity of distinct cardiac areas. Consequently, their magnitude and patterns can be utilized to diagnose cardiovascular problems. The P wave often has a circular, homogeneous, regular shape that denotes depolarization of the ventricles. The length, intensity, variation, and median are determined by segmenting the ECG data and detecting a P Pulse pattern. Three waves make up the QRS component, which shows complete the ventricles are depolarized. The initial positive wave is known as R Pulse, while the primary and secondary inverse waves following P are known as Q and S. An R signal is recognized. Its length and intensity are calculated, similarly to the P-waves.

Feature engineering

This essential task is carried out during the registration and authentication processes. This might comprise selecting and extracting features, then training a classification system using those features. Machine learning algorithms have several shortcomings, according to prior studies. To build the authentication system, they need a very intricate basis. Utilizing the 200 Hertz frequency specimens that represent the typical sampling rate obtained per second, these characteristics were retrieved from the Hea file. The psychographic and empirical data included gender, ECG mean, patient, RR, age, variance, and standard deviation and patient tests. Furthermore, we utilized exclusively the recently obtained electrocardiograph data, subsequent to additional processing to detect QRS spikes. The patient was the subject, with ECG assessment utilized as the predictor. Figure 6 shows the difference between normal, supraventricular, and unknown heartbeats extracted from the ECG signal.

This study highlights manual feature engineering, informed by established clinical knowledge and prior research, despite the prevalent use of automated feature selection techniques like recursive feature elimination (RFE), principal component analysis (PCA), and mutual information-based selection in ECG classification tasks. Current research indicates that although automated techniques can efficiently diminish dimensionality and emphasize statistically relevant features, they may neglect physiologically crucial signals, particularly in delicate biomedical contexts such as ECG analysis. Previous studies (Acharya et al., 2017) indicate that manually extracted features—such as RR intervals, QRS complex characteristics, and waveform morphology—consistently enhance the interpretability and diagnostic relevance of models. Consequently, in accordance with these findings, this study emphasizes domain-specific feature extraction from raw ECG signals, using temporal and statistical descriptors to preserve classification accuracy and clinical reliability. Future endeavors may integrate automated feature selection as an auxiliary method to enhance and corroborate the feature set.

To evaluate the influence of manual feature engineering on model performance, we perform an ablation study comparing two models: one trained exclusively on the raw, as reported in Fig. 7, preprocessed ECG waveforms and the other enhanced with engineered features (gender, age, RR intervals, QRS complex characteristics, variance, and standard deviation). Integrating these domain-specific characteristics results in a significant enhancement across all critical metrics: validation accuracy escalates from 88.1% to 92.4%, the area under the ROC curve (AUC) advances from 0.90 to 0.95, and the false acceptance rate decreases by 12%. Moreover, the improved model demonstrates greater stability in convergence, decreasing the number of epochs required to achieve ideal performance by 20%. The results statistically demonstrate that incorporating clinically relevant variables into the classifier markedly enhances its discriminative capability and resilience.

Threshold-based authentication system

To trust the people and equipment involved in a telehealth environment, authentication is a crucial requirement. One-time verification has its drawbacks, which have been addressed by secure authentication. The individual is verified once in conventional authentication schemes. ML provides the key to using interactive verification. A key measure to prevent unwanted administrator privileges is to use ML to authenticate all organizations, such as IoT Connected systems, medical professionals, and healthcare workers, before allowing them to communicate with system resources. The resemblance of a specimen toward every input point distribution is calculated using deep learning classifiers. They establish a probability measure based on how similar the classes are, and then pick the class with the highest likelihood. The authentication mechanism, however, is more complicated than a simple classifier. Following the system designed in the figure, if the predicted identity and the claimed identity are the same, confidentiality criteria must be checked by comparing the likelihood of this class with a threshold. This test label will be approved if the probability exceeds the specified level. In order to maximize the system’s sensitivity and specificity, an adaptive threshold that is determined by cross-validation may be used to govern the system’s strictness against incursion. Tighter approval standards can be managed for more delicate conditions by raising the threshold level. We specified the threshold level as seventy percent in our experiments, which means only those ECG signals are considered whose matching probability is more significant than 70%. Figure 8 presents the block diagram of the proposed threshold-based policy for user authentication using ECG signals.

Table 2 presents the number of scenarios that can be encountered by the authentication system. In scenario 1 if the 1 has more probability then the authentication system will allow it to go through to access the telehealth system, otherwise, it will deny the entry. In scenario 2, the system will block all the input ECG signals because the probability score is less than the threshold.

The threshold level of 70% is selected arbitrarily in this implementation, rather than being optimized by cross-validation or parameter tuning. This conclusion is consistent with existing literature on ECG-based authentication, which normally employs fixed threshold values ranging from 60% to 80% to achieve a balance between false acceptance and rejection rates. The objective is to implement a conservative yet pragmatic threshold that functions dependably without introducing complexity to the authentication system. This study does not address adaptive thresholding, which could improve system sensitivity and specificity through ROC analysis or validation-based tuning. Future research may explore dynamic threshold selection to more effectively address fluctuations in signal quality and individual user characteristics.

To manage edge factors, such as signals that fluctuate around the threshold, the system may establish a buffer zone around the threshold. This strategy entails establishing a limited range or “buffer zone” surrounding the threshold to tolerate signals that are proximate to the decision boundary. If the threshold is established at 70% signals with probability ranging from 65% to 75% may be regarded as within the buffer zone.

Within this buffer zone, the system may implement other measures to enhance decision-making. Signals within the buffer zone may initiate a secondary verification process. This may entail employing an alternative classification model that is more attuned to the distinctive characteristics of the ECG signal or reassessing the judgment by examining supplementary attributes or contextual data. For borderline signals, the system may solicit further information, such as supplementary samples or postponed verification, to enhance decision-making confidence.

This method guarantees that signals close to the threshold do not result in capricious acceptance or rejection. It introduces adaptability, enabling the system to manage confusing situations with greater diligence and accuracy. The supplementary analysis offered by the buffer zone might diminish false positives and negatives, so strengthening the integrity of the authentication process.

For instance, consider a remote patient monitoring system where in a nurse authenticates many times daily via ECG-based verification. Occasionally, minor alterations in sensor positioning or bodily movement may result in the ECG signal yielding a similarity score of 68%, which resides within the buffer zone. Rather than outright refusing access, the system initiates a re-verification prompt, instructing the nurse to sit still and attempt authentication again. If the system identifies recent successful authentications from the same device and user context, it may confer a contextual confidence enhancement and permit access. This adaptive strategy guarantees operational continuity in practical situations while preserving system security.

Evaluation metrices

The confusion matrix is a fundamental tool for evaluating a classifier’s performance. It provides a detailed breakdown of classification outcomes by comparing predicted results with actual labels, yielding four key values: true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). These values form the basis for computing important performance metrics such as precision, sensitivity (recall), specificity, and accuracy. Among these, accuracy—defined as the proportion of correctly classified instances across all classes—is the most commonly used metric, especially in cross-validation settings, where it reflects the number of correctly identified sequences relative to the total number of batches evaluated.

(7) $$ACC=\left(\sum _{i=1}^N\frac{TP+TN}{TP+FP+TN+FN}\right)\cdot 100\mathrm{\%}/N.$$

The classification algorithm’s sensitivity determines how successful it is in diagnosing all ill persons; specificity refers to the fraction of negative samples that the classification model properly classified as negative instances. The ROC curve is another crucial quality indicator that is used to visually assess the effectiveness of the suggested solution.

(8) $$SEN=\left(\sum _{i=1}^N\frac{TP}{TP+FN}\right)\cdot 100\mathrm{\%}/N$$

(9) $$SPE=\left(\sum _{i=1}^N\frac{TN}{FP+TN}\right)\cdot 100\mathrm{\%}/N$$

(10) $$PRE=\left(\sum _{i=1}^N\frac{TP}{FP+TP}\right)\cdot 100\mathrm{\%}/N$$

(11) $$F1\text{-}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\left(\sum _{\mathrm{i}=1}^{\mathrm{N}}\frac{2(\mathrm{T}\mathrm{P})}{2(\mathrm{T}\mathrm{P})+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}\right)\cdot 100\mathrm{\%}/\mathrm{N}.$$