Non-invasive cardiac pattern monitoring is essential for assessing how individuals respond to cognitive and physical stressors, as these stressors influence cardiac function by activating the sympathetic (SNS) and parasympathetic (PNS) branches of the autonomic nervous system (ANS) [1]. Over 70 % of people all over the world experience stress-related symptoms, which can compromise the immune system and lead to severe health issues like hypertension, diabetes, post-traumatic stress disorder (PTSD), and depression [2]. Prolonged stress elevates cortisol levels, a steroid hormone, which worsens cardiovascular and mental illnesses and is associated with leading causes of death, such as heart disease, stroke, cancer, and suicide [3]. Consequently, there has been a surge in research utilizing ECG signals from wearable devices, along with machine-learning and deep-learning, to analyze heart rate (HR) and its relationship with stress. HR serves as a non-invasive marker of ANS activity, regulating vital functions like heart rate, respiration, and blood pressure [4]. Time-domain (TD), frequency-domain (FD), and non-linear statistical features have proven effective in assessing stress-induced HR changes, with increased SNS activity and decreased PNS activity being early indicators of cardiac issues [5,6]. This has spurred the development of machine-learning and deep-learning models for predicting stress and antedating cardiac diseases.

Recent studies have explored stress detection using various machine-learning and deep-learning models, focusing on the impact of stress from activities such as public speaking, cognitive tasks, and exercise [7]. In 2024, one study developed an RF approach to classify two classes (non-stress and stress) using ECG and skin temperature, achieving 98.29 % accuracy and a 97.89 % F1-score [8]. Additionally, XGBoost (XGB) was used to classify three stress levels (low, medium, high) with the same accuracy and F1-score. Roopa et al. [9] also analyzed ECG signals using the DeepChill deep learning model for non-stress and stress recognition, reporting an accuracy of 91.90 % and an F1-score of 81.70 %. Feng et al. [10] proposed a LSTM-1DCNN model to classify three states (neutral, stress, amusement) using ECG and electromyogram (EMG), with 87.82 % accuracy and an F1-score of 86.68 %. Further research applied a CNN model to detect mental stress from ultra-short heart rate variability, achieving 97.75 % accuracy [11]. In 2022, a study used a fuzzy logic approach (Xception) to classify driver stress into three levels using ECG signals, leveraging CWT and Morse wavelet-based scalograms, achieving 98.11 % accuracy and a 97.87 % F1-score [12]. Another study reported that deep-learning algorithms for two-class stress prediction using ECG data yielded accuracies of 98.99 % with CNN and 98.92 % with VGG16 [13]. Kang et al. [14] introduced a CNN-LSTM model using spectrograms to classify mental stress with 98.3 % accuracy. Despite these innovations, challenges remain, particularly in capturing the non-linear nature of cardiac functions during cognitive and physical activities, such as irregular R-peaks generation and RR interval formation, which are difficult for machine-learning and CWT-based deep-learning models to detect. These studies highlight ongoing advancements and emphasize the need for selecting the right models and integrating advanced computational techniques to improve performance and real-world applicability in stress detection.

The Glasgow University ECG database (ECG-GUDB) [15] serves as a valuable resource for analyzing stressed cardiac patterns through ECG signals. While prior models achieved up to 98 % accuracy and a 97.90 % F1-score in classifying two classes (healthy and atrial fibrillation) [[16], [17], [18]], their practical application remains limited due to challenges in detecting non-linear cardiac behaviors. The ECG-GUDB dataset [15] was used to develop a CNN model that classifies five cardiac patterns using bispectrum-based contours, capturing non-linear ECG dynamics with high accuracy. This study aims to provide a robust method for cardiac pattern assessment, supporting improved stress management.

In this study, an image-based CNN deep-learning model (MobileNetV2 architecture) was proposed to efficiently detect cardiac patterns, enhancing accuracy and reliability through bispectrum contour-based image classification. ECG data from ECG-GUDB dataset, involving subjects performing five activities—sitting (relaxed), math-reasoning (cognitive stressor), and walking, jogging, and hand-biking (physical stressor)—were analyzed to assess the heart's physiological responses. TD, FD, and statistical features such as average heart rate (AHR), standard deviation of normal RR intervals (SDNN), root mean square successive difference between RR intervals (RMSSD), normal-to-normal intervals differing by more than 50 milliseconds (NN50), stress index, vagal-sympathetic effect (VSE), parasympathetic nervous system (PNS) index, sympathetic nervous system (SNS) index, SNS-to-PNS ratio, kurtosis, and skewness, were extracted from filtered ECG signals. After that, two feature selection techniques such as Pearson correlation and LASSO regularization were employed to identify the most important features for improved cardiac pattern prediction. The selected features were then used as input for a feature-based RF machine-learning model. Concurrently, CWT-based scalograms and bispectrum-based contours were utilized in an image-based CNN model to classify five distinct cardiac patterns effectively. Bispectrum-based contours highlighted the nonlinearity in cardiac activity by mapping R-peak frequency distributions, enhancing the CNN model's ability to distinguish stress-induced cardiac patterns.