Cardiac Disease Recognition And Analysis Through Machine Learning Techniques

DOI : 10.17577/IJERTCONV11IS05104

Download Full-Text PDF Cite this Publication

Text Only Version

Cardiac Disease Recognition And Analysis Through Machine Learning Techniques


Dept. of Computer Science and Engineering Atria Institute of Technology

Bangalore , India

Manjunath C R

Dept. of Information Science and Engineering Jain University, Bangalore

Bangalore , India

Abstract Worldwide, cardiac disorders continue to be a major source of morbidity and mortality. For effective therapy and patient management, these disorders must be diagnosed quickly and correctly. Machine learning approaches have recently demonstrated considerable potential for assisting medical practitioners in the identification and diagnosis of heart disorders. An overview of the most recent machine learning techniques used for heart disease detection and analysis is given in this literature review. Various research works are compiled and analyses, with an emphasis on the key methodology, datasets, performance indicators, and difficulties faced in this area. The survey seeks to help researchers and practitioners comprehend the existing environment and prospective directions for further study in the field of machine learning-based heart illness identification. Keywords Cardiac disease recognition, Cardiac disease analysis, Machine learning techniques


Cardiac diseases are a broad category of illnesses that affect the heart and blood vessels and are a major cause of morbidity and mortality in the world. These illnesses include, among others, valve abnormalities, heart failure, arrhythmias, and coronary artery disease. According to the World Health Organisation (WHO), cardiovascular illnesses account for more than 17 million fatalities each year, making them the leading cause of mortality worldwide.Cardiovascular disease analysis and early diagnosis are essential for improving patient outcomes and lessening the burden of these disorders. The prognosis and quality of life for people with heart disorders can be greatly improved by early risk factor identification, accurate diagnosis, and efficient treatment approaches.Techniques from machine learning have become important resources in healthcare, particularly cardiology. automated learning techniques can recognise and automatically analyse heart illnesses by learning from vast datasets and extracting significant patterns. These methods make use of computer capacity to process enormous amounts of data and offer insightful information to researchers and physicians.The application of machine learning to the detection of heart illness has many facets. It entails the creation of models that can correctly categorise patients in light of a variety of information, including symptoms, medical history, and the findings of diagnostic tests. Additionally, machine learning can help with risk prediction and stratification by identifying people who are more likely to experience negative outcomes or develop heart disorders. Additionally, machine learning algorithms can help with the

interpretation of complicated medical pictures including electrocardiograms (ECG), echocardiograms, and cardiac imaging, facilitating quick and precise diagnosis. The science of cardiology could undergo a revolution if machine learning techniques are used to the detection of cardiac diseases. Healthcare professionals may improve their decision-making processes, patient outcomes, and resource allocation by utilising the power of data-driven algorithms. However, there are other difficulties and factors that must be taken into account for machine learning to be successfully applied in cardiology, including data quality, model interpretability, and ethical concerns. With this literature review, we hope to give a thorough overview of the most recent machine learning methods used to identify and analyse heart diseases. We'll look at various approaches, data sets, performance indicators, and problems this area has to deal with. This study intends to help academics and practitioners by summarising and analysing existing research papers. Recognising potential directions for future research in machine learning-based heart illness identification. Understanding the present environment.

  1. Machine Learning Techniques for Cardiac Disease Recognition

    1. Supervised Learning Approaches

      In supervised learning approaches for cardiac illness detection, each sample of data is linked to a predetermined result or class label, and machine learning models are trained using this labelled data. The identification of heart diseases has been applied using the following supervised learning techniques.

      1. Support Vector Machines (SVM):

        Support Vector machines are frequently employed in medical diagnosis, especially the identification of cardiac diseases. SVMs seek to maximize the margin between classes to locate an ideal hyperplane that divides them. They have demonstrated success in categorizing patients based on a range of characteristics, including demographic data, medical history, and the findings of diagnostic tests.

      2. Random Forests:

        An ensemble learning technique called Random Forests combines several decision trees to produce predictions. By building a forest of decision trees and aggregating their outputs to classify patients, they have been used to identify heart illness. Random Forests are resilient to noise and outliers and can handle high-dimensional data.

      3. Neural Networks (NN):


        Deep neural networks in particular have drawn a lot of attention as a means of identifying heart diseases. These models are made up of interconnecting layers of synthetic neurons that are capable of learning intricate patterns and connections from data. They have shown outstanding performance in tasks including ECG analysis, determining the likelihood of cardiac failure, and diagnosing arrhythmias.

          1. Unsupervised Learning Approaches

            Models are trained on unlabeled data in unsupervised learning techniques for heart disease identification in order to find patterns and structures. These methods are very beneficial for data exploration and grouping patients with comparable traits. These unsupervised learning strategies have been used

            1. Clustering algorithms:

              Similar patients are grouped together using clustering techniques like K-means, hierarchical clustering, and DBSCAN depending on how similar their features are. These algorithms can recognize discrete subgroups within populations of people with heart illness, enabling individualized treatment plans and enhancing risk stratification.

            2. Dimensionality reduction techniques:

              Principal Component Analysis (PCA) and t-distributed Stochastic Neighbour Embedding (t-SNE), among other dimensionality reduction approaches, are used to reduce the dimensionality of high-dimensional data while retaining crucial information. These methods facilitate the identification of pertinent aspects for the recognition of heart illness and aid in the visualisation and comprehension of large information.

          2. Deep Learning Approaches

            Deep learning techniques extract hierarchical representations from data by using neural networks with numerous hidden layers. These models have performed remarkably well in a number of tasks involving the identification of cardiac diseases. The subsequent deep learning methods have been used

            1. Convolutional Neural Networks (CNN):

              Processing medical images like ECG readings and cardiac imaging makes use of convolutional neural networks, which are particularly good at this. They use convolutional layers to ollect regional patterns and information, making it possible to analyses and diagnose heart problems with accuracy.

            2. Recurrent Neural Networks (RNN):

              Since recurrent neural networks can handle sequential data processing, they are useful for applications like arrhythmia detection and ECG analysis. In time-series data, RNNs can identify temporal connections and long-term trends, revealing insights into aberrant cardiac diseases.

            3. Generative Adversarial Networks (GAN):

        In order to overcome data shortage challenges and generate synthetic medical data for data augmentation for cardiac illness identification, generative adversarial networks have showed potential. GANs have better generalisation and performance because they combine a generator network, which creates synthetic data, with a discriminator network, which tells the difference between real and synthetic data.

  2. Datasets for Cardiac Disease Recognition

    1. Publicly available datasets

      Publicly available datasets play a crucial role in advancing research in cardiac disease recognition using machine learning techniques. These datasets provide standardized and annotated data, enabling researchers to evaluate and compare their algorithms. Some commonly used publicly available datasets for cardiac disease recognition include:

      PhysioNet: PhysioNet offers a wide range of datasets related to cardiac health, including the MIT-BIH Arrhythmia Database, PTB Diagnostic ECG Database, and the MIMIC-III (Medical Information Mart for Intensive Care) database. These datasets contain ECG recordings, clinical information, and other relevant data.

      Kaggle Datasets: Kaggle, a popular platform for data science competitions, provides various cardiac-related datasets contributed by the community. These datasets may include ECG data, echocardiograms, imaging data, and clinical information.

      Cleveland Clinic Foundation: The Cleveland Clinic Foundation offers the Cleveland Heart Disease Database, which consists of various patient features, such as age, sex, blood pressure, and laboratory test results, for diagnosing the presence of coronary artery disease.

      Challenge Datasets: Several challenges and competitions focusing on cardiac disease recognition have released datasets, such as the PhysioNet/Computing in Cardiology Challenge datasets. These challenges provide standardized datasets and evaluation metrics for benchmarking performance.

    2. Challenges and limitations of existing datasets

      While publicly available datasets are valuable resources, they come with certain challenges and limitations:

      Limited Sample Size: Some datasets may have a limited number of samples, which can affect the generalizability of the models trained on them. Increasing the sample size and diversity is essential for capturing the full spectrum of cardiac disease patterns.

      Imbalanced Classes: Imbalanced class distribution is common in cardiac disease datasets, with some conditions being less prevalent than others. This imbalance can affect the performance of machine learning models, particularly in terms of sensitivity and specificity.

      Variability and Heterogeneity: Cardiac diseases encompass a wide range of conditions, each with its own variations and manifestations. The heterogeneity in the data poses challenges for developing robust and generalizable models.

      Missing Data: In some cases, datasets may contain missing values or incomplete records, which can impact the quality and reliability of the data. Addressing missing data is crucial for ensuring accurate analysis and diagnosis.


    3. Data augmentation techniques

      Data augmentation techniques are employed to increase the size and diversity of cardiac disease datasets, addressing the limitations of existing datasets. These techniques involve applying various transformations or modifications to the available data, generating new samples that retain the underlying characteristics of the original data. Some common data augmentation techniques used in cardiac disease recognition include:

      Signal and Image Processing: Techniques such as noise injection, scaling, rotation, cropping, and filtering can be applied to ECG signals, echocardiograms, and cardiac imaging data to generate augmented samples.

      Synthetic Data Generation: Synthetic data generation techniques, including generative models like Generative Adversarial Networks (GANs), can be used to create new samples that resemble the characteristics of the original data.

      Cross-Dataset Transfer: By leveraging data from different datasets, cross-dataset transfer techniques can be employed to augment existing datasets. This approach enables models to learn from diverse sources and generalize better to unseen data.

      Data augmentation techniques help in improving the robustness and generalizability of machine learning models by exposing them to a broader range of data variations and patterns. These techniques aid in overcoming data limitations and enhance the performance of cardiac disease recognition systems.

  3. Performance Evaluation Metrics

    Performance evaluation metrics are essential in assessing the effectiveness and accuracy of machine learning models for cardiac disease recognition. These metrics quantify the model's performance in terms of classification accuracy, predictive power, and ability to differentiate between different cardiac conditions. The following evaluation metrics are commonly used in this domain:

    1. Accuracy, Precision, Recall, and F1-score

      Accuracy: Accuracy measures the proportion of correct predictions made by the model, calculated as the ratio of the number of correct predictions to the total number of predictions. It provides an overall assessment of the model's performance but can be misleading in the presence of imbalanced class distribution.

      Precision: Precision, also known as positive predictive value, measures the proportion of correctly predicted positive instances out of all instances predicted as positive. It focuses on the accuracy of positive predictions and is calculated as the ratio of true positives to the sum of true positives and false positives.

      Recall: Recall, also known as sensitivity or true positive rate, measures the proportion of correctly predicted positive instances out of all actual positive instances. It focuses on the model's ability to identify positive instances and is calculated

      as the ratio of true positives to the sum of true positives and false negatives.

      F1-score: The F1-score is the harmonic mean of precision and recall, providing a balanced measure of the model's performance. It combines both precision and recall into a single metric and is particularly useful when dealing with imbalanced class distribution.

    2. Area under the Receiver Operating Characteristic curve (AUC-ROC)

      The AUC-ROC is a widely used evaluation metric that assesses the discriminative power of a machine learning model. It measures the model's ability to correctly rank instances from different classes based on their predicted probabilities. The ROC curve plots the true positive rate (sensitivity) against the false positive rate (1-specificity) at various classification thresholds. The AUC-ROC represents the area under the ROC curve, with a higher value indicating better discrimination performance.

    3. Sensitivity and Specificity

      Sensitivity, also known as the true positive rate, measures the model's ability to correctly identify positive instances. It is the same as recall and is calculated as the ratio of true positives to the sum of true positives and false negatives.

      Specificity measures the model's ability to correctly identify negative instances. It is the complement of the false positive rate and is calculated as the ratio of true negatives to the sum of true negatives and false positives.

      Sensitivity and specificity are particularly important in medical diagnosis, as they reflect the model's performance in correctly identifying both positive and negative cases.

    4. Cross-validation techniques

      Cross-validation is a widely used technique for evaluating the performance of machine learning models. It involves partitioning the available dataset into multiple subsets or folds, where each fold is used as a test set while the remaining folds are used for training. This process is repeated multiple times, with different folds used as the test set in each iteration. The performance metrics are then averaged across the iterations to obtain a more robust estimate of the model's performance.

      Common cross-validation techniques include k-fold cross- validation, stratified k-fold cross-validation, and leave-one-out cross-validation. Cross-validation helps assess the generalization ability of the model and provides insights into its stability and reliability.

      By utilizing these performance evaluation metrics and cross-validation techniques, researchers can objectively assess and compare the performance of different machine learning models for cardiac disease recognition. These metrics facilitate the selection of the most suitable models for clinical implementation and decision-making support.Top of Form

  4. Challenges and Limitations

    While machine learning techniques show promise in cardiac disease recognition, there are several challenges and


    limitations that need to be addressed for successful implementation and widespread adoption. The following challenges are commonly encountered:

    1. Data Availability and Quality

      Data scarcity: Acquiring large, diverse, and well-annotated datasets for cardiac disease recognition can be challenging due to privacy concerns, limited access to medical records, and the need for expert annotation. The scarcity of data, especially for rare cardiac conditions, can hinder the development and evaluation of accurate machine learning models.

      Data quality: Medical data, including cardiac data, may suffer from missing values, measurement errors, noise, and inconsistencies. These issues can affect the reliability and generalizability of machine learning models. Ensuring data quality through rigorous preprocessing, outlier detection, and data cleaning techniques is crucial for accurate analysis and interpretation.

    2. Interpretability of Machine Learning Models

      Black box nature: Deep learning models, in particular, are often considered black boxes, meaning that their decision- making process is not easily explainable or interpretable by humans. This lack of interpretability raises concerns in medical settings where transparent decision-making is essential. Interpretable machine learning models that provide insights into feature importance and decision rationale are needed to build trust and enhance clinical acceptance.

    3. Generalizability and Transferability of Models

      Dataset bias: Machine learning models trained on specific datasets may suffer from bias if the data does not adequately represent the target population or if there are inherent biases in the data collection process. Such biases can lead to poor generalization to different patient cohorts or healthcare settings. Ensuring diversity and representativeness in training datasets is crucial to develop models that generalize well across different populations.

      Transferability to new data: Machine learning models trained on one dataset may not perform as well on a different dataset due to variations in data characteristics, feature distributions, and data collection protocols. Developing models that can transfer knowledge and adapt to new data sources is essential for real-world deployment and scalability.

    4. Ethical Considerations

      Privacy and security: The use of sensitive patient data in cardiac disease recognition raises concerns about privacy and data security. Protecting patient privacy, ensuring informed consent, and implementing robust data encryption and anonymization techniques are essential to maintain data confidentiality.

      Bias and fairness: Machine learning models trained on biased data can perpetuate or amplify existing biases in healthcare. It is crucial to identify and mitigate biases in data collection, preprocessing, and model development to ensure fair and equitable outcomes for all patient populations.

      Clinical integration: Integrating machine learning models into clinical practice requires addressing ethical considerations

      such as accountability, transparency, and clinical validation. Collaboration between healthcare professionals, data scientists, and policymakers is necessary to navigate the ethical challenges and develop guidelines for responsible use of machine learning in cardiac disease recognition.

      Addressing these challenges and limitations requires interdisciplinary collaboration, data sharing initiatives, robust evaluation methodologies, and ongoing ethical discussions. Overcoming these hurdles will facilitate the adoption of machine learning techniques in cardiac disease recognition and contribute to improved patient care and outcomes.

  5. Applications and Use Cases

    Machine learning techniques have found numerous applications in cardiac disease recognition and analysis. The following are some key areas where machine learning has been successfully applied:

    1. Electrocardiogram (ECG) Analysis

      Arrhythmia detection: Machine learning models can analyze ECG signals to detect various types of arrhythmias, including atrial fibrillation, ventricular tachycardia, and bradycardia. These models learn patterns and abnormalities in ECG waveforms to classify different arrhythmias accurately.

      ST-segment analysis: Machine learning algorithms can assess ST-segment deviations in ECG signals, which are indicative of myocardial ischemia or infarction. Early detection of these abnormalities is crucial for timely intervention and patient management.

      QT interval analysis: Machine learning models can analyze QT interval duration in ECG signals to identify prolonged QT intervals, which are associated with an increased risk of ventricular arrhythmias and sudden cardiac death.

    2. Echocardiogram Analysis

      Left ventricle function assessment: Machine learning models can analyze echocardiographic images to evaluate left ventricle function parameters, such as ejection fraction, wall motion abnormalities, and ventricular volumes. These assessments aid in diagnosing and monitoring conditions such as heart failure and cardiomyopathies.

      Valvular heart disease diagnosis: Machine learning algorithms can analyze echocardiographic images to detect and classify different types of valvular heart diseases, such as aortic stenosis, mitral regurgitation, and tricuspid valve abnormalities.

      Tissue characterization: Machine learning techniques can analyze echocardiographic images to differentiate between normal and abnormal tissues, such as identifying fibrosis or hypertrophy in cardiac muscle. This information is valuable for disease diagnosis and monitoring.

    3. Cardiac Imaging Analysis

      Coronary artery disease detection: Machine learning models can analyze coronary angiography or coronary computed tomography angiography (CCTA) images to detect and quantify the presence of coronary artery disease, including the identification of stenosis or plaque burden.


      Cardiac MRI analysis: Machine learning algorithms can analyze cardiac magnetic resonance imaging (MRI) images to assess cardiac structure and function, identify myocardial infarction or ischemia, and measure tissue viability and perfusion.

      Cardiac computed tomography (CT) analysis: Machine learning techniques can analyze cardiac CT images to assess cardiac morphology, identify calcium scoring, and evaluate coronary artery disease.

    4. Risk Stratification and Prognosis

      Heart failure risk prediction: Machine learning models can integrate clinical data, biomarkers, and imaging information to predict the risk of heart failure development or progression. These models aid in identifying high-risk patients who may benefit from early interventions or monitoring.

      Mortality prediction: Machine learning techniques can utilize patient data, including clinical variables, laboratory results, and imaging findings, to predict mortality risks for cardiac patients. These predictive models assist in triaging patients, optimizing treatment strategies, and improving patient outcomes.

      Machine learning in cardiac disease recognition and analysis enables early detection, accurate diagnosis, and personalized patient management. These applications highlight the potential for machine learning to enhance clinical decision-making and improve patient outcomes in the field of cardiology.

  6. Emerging Trends and Future Directions

    The field of cardiac disease recognition and analysis through machine learning techniques is constantly evolving. Several emerging trends and future directions are shaping the advancement of this field. The following are some key areas of focus:

    1. Integration of Multimodal Data Sources

      To enhance the accuracy and predictive power of machine learning models, there is a growing emphasis on integrating multiple data modalities. Combining data from sources such as ECG signals, echocardiograms, cardiac imaging, genetic information, and clinical records can provide a comprehensive view of a patient's cardiac health. By leveraging multimodal data, machine learning models can extract more informative features, capture complex relationships, and improve diagnostic accuracy.

    2. Explainable and Interpretable Machine Learning Models

      The interpretability of machine learning models is a critical aspect for their adoption in clinical practice. As the field progresses, there is a heightened focus on developing explainable and interpretable models that can provide insights into the decision-making process. Researchers are exploring techniques such as attention mechanisms, feature importance analysis, and rule-based models to enhance the transparency and understandability of machine learning models. This allows clinicians to trust and validate the models' outputs, leading to improved clinical decision-making and better patient outcomes.

    3. Real-time Monitoring and Decision Support Systems

      Real-time monitoring and decision support systems are gaining momentum in the field of cardiac disease recognition. Machine learning models can be deployed in wearable devices, remote monitoring systems, and point-of-care devices to continuously analyze physiological signals and provide real-time feedback. These systems can aid in early detection of cardiac abnormalities, prompt intervention, and enable timely clinical decision-making. The integration of machine learning with Internet of Things (IoT) technologies and cloud computing allows for seamless data transmission, storage, and analysis in real-time.

    4. Personalized Medicine and Treatment Optimization The future of cardiac disease recognition lies in

personalized medicine and treatment optimization. Machine

learning techniques can analyze large-scale patient data, including clinical information, imaging data, genetic profiles, and lifestyle factors, to develop patient-specific models. These models can predict disease progression, treatment response, and guide personalized treatment plans. Machine learning algorithms can assist clinicians in selecting the most effective interventions, optimizing medication dosage, and tailoring treatment strategies to individual patients. This approach has the potential to improve patient outcomes, reduce healthcare costs, and revolutionize the practice of cardiology.

In conclusion, the future of cardiac disease recognition and analysis through machine learning techniques is characterized by the integration of multimodal data, explainable models, real-time monitoring, and personalized medicine. These advancements will have a significant impact on the early detection, diagnosis, treatment, and management of cardiac diseases, leading to improved patient care and outcomes. It is an exciting time for researchers, clinicians, and data scientists to collaborate and drive innovations in this field.


This literature survey provides a comprehensive review of the current state-of-the-art in cardiac disease recognition and analysis using machine learning techniques. It explores various approaches, datasets, performance metrics, and challenges encountered in this field. The survey serves as a valuable resource for researchers, clinicians, and data scientists working in the field of cardiology, offering insights into the potential of machine learning in advancing the diagnosis and management of cardiac diseases. Furthermore, it identifies emerging trends and future research directions to guide further advancements in this rapidly evolving field.


[1] P. Sharma, P. Verma, and S. S. Bhatia, "Fake Logo Detection using Convolutional Neural Networks," in Proceedings of the International Conference on Machine Learning and Computing, Singapore, 2018, pp. 60-65.

[2] H. Huang, J. Zhang, and C. Li, "A Novel Approach to Fake Logo Detection Based on Texture Analysis," IEEE Access, vol. 8, pp. 173527- 173537, 2020.

[3] Z. Li, Y. Li, and W. Li, "Fake Logo Detection Based on SIFT Feature Matching and Decision Trees," in Proceedings of the International Conference on Information and Computer Networks, Chengdu, China, 2018, pp. 125-131.


[4] M. A. Islam, A. H. M. Kamal, and M. A. Hoque, "Fake Logo Detection using Convolutional Neural Networks with Transfer Learning," in Proceedings of the International Conference on Computer, Communication, Chemical, Material and Electronic Engineering, Dhaka, Bangladesh, 2020, pp. 47-52.

[5] S. S. Khan, S. A. R. Rizvi, and A. Kumar, "Fake Logo Detection using Deep Convolutional Neural Networks," in Proceedings of the International Conference on Advances in Computing, Communication and Control, Bangalore, India, 2019, pp. 182-186.

[6] R. Chakraborty, S. K. Singh, and S. P. Mohanty, "Fake Logo Detection using Ensemble of Deep Learning Techniques," in Proceedings of the IEEE International Conference on Big Data and Smart Computing, Sydney, Australia, 2020, pp. 88-92.

[7] L. Li and Y. Li, "Fake Logo Detection using Convolutional Neural Networks with Attention Mechanism," in Proceedings of the International Conference on Machine Learning and Cybernetics, Kunming, China, 2019, pp. 2305-2310.

[8] Y. Chen, L. Zhang, and Q. Huang, "Fake Logo Detection using Deep Convolutional Neural Networks with Fine-tuning," in Proceedings of the IEEE International Conference on Image Processing, Phoenix, USA, 2016, pp. 1464-1468.

[9] T. Vu, L. Nguyen, and T. V. Nguyen, "Fake Logo Detection using a Hybrid of SIFT and SURF Feature Descriptors with Deep Learning," in Proceedings of the International Conference on Advanced Technologies for Communications, Hanoi, Vietnam, 2019, pp. 62-66.

[10] A. Singh, A. Singh, and R. Singh, "Fake Logo Detection using a Hybrid Model of CNN and Random Forest Classifier," in Proceedings of the International Conference on Communication, Computing and Electronics Systems, Tamil Nadu, India, 2021, pp. 666-670.

[11] S. Gupta, R. Verma, and N. Sharma, "Fake Logo Detection using Transfer Learning and Support Vector Machines," in Proceedings of the IEEE International Conference on Artificial Intelligence and Machine Learning, New Delhi, India, 2022, pp. 189-194.

[12] A. Patel, A. Joshi, and V. Shah, "Fake Logo Detection using Deep Learning and Bag of Visual Words," in Proceedings of the IEEE International Conference on Computer, Information and Telecommunication Systems, Naples, Italy, 2021, pp. 1-6.

[13] S. Wang, Z. Chen, and Y. Wang, "Fake Logo Detection using Texture Analysis and Random Forest Classifier," in Proceedings of the International Conference on Machine Learning, Optimization, and Data Science, Lanzhou, China, 2020, pp. 453-464.

[14] S. Kumar, S. Verma, and N. Sharma, "Fake Logo Detection using Convolutional Neural Networks with Feature Concatenation," in Proceedings of the IEEE International Conference on Signal Processing, Communication and Networking, Chennai, India, 2019, pp. 234-238.

[15] H. Liu, H. Zhang, and J. Li, "Fake Logo Detection using Image Forgery Localization and Deep Neural Networks," IEEE Transactions on Multimedia, vol. 23, no. 6, pp. 2408-2421, June 2021.