Size Of Dataset Research Articles

Making waves: Knowledge and data fusion in urban water modelling

Mathematical modeling plays a crucial role in understanding and managing urban water systems (UWS), with mechanistic models often serving as the foundation for their design and operations. Despite the wide adoptions, mechanistic models are challenged by the complexity of dynamic processes and high computational demands. Data-driven models bring opportunities to capture system complexities and reduce computational cost, by leveraging the abundant data made available by recent advance in sensor technologies. However, the interpretability and data availability hinder their wider adoption. This paper advocates for a paradigm shift in the application of data-driven models within the context of UWS. Integrating existing mechanistic knowledge into data-driven modeling offers a unique solution that reduces data requirements and enhances model interpretability. The knowledge-informed approach balances model complexity with dataset size, enabling more efficient and interpretable modeling in UWS. Furthermore, the integration of mechanistic and data-driven models offers a more accurate representation of UWS dynamics, addressing lingering uncertainties and advancing modelling capabilities. This paper presents perspectives and conceptual framework on developing and implementing knowledge-informed data-driven modeling, highlighting their potential to improve UWS management in the digital era.

Size stable batch mode model updating method

In near-infrared spectroscopy analysis, the predictive performance of models may degrade due to variations in the measurement environment and sample properties. Maintenance or updates to the model are essential to uphold prediction accuracy. The state-of-the-art approach for model updates involves labeling new samples and incorporating them into the calibration set. However, this strategy leads to an escalating size of the calibration dataset, and the low ratio of new samples renders the model update process inefficient. To enhance update efficiency, a size-stable updating strategy is presented which involves the simultaneous addition of new samples and removal of old samples. This is achieved by utilizing the new set as a basis for identifying and deleting significantly different old labeled samples. To improve labeling efficiency, a batch mode update is employed. Initially, calibration samples are clustered to yield multiple clusters with comparable sizes to the new update set. Subsequently, differences between multiple old sample clusters and the new set are calculated. The old sample cluster that most different are removed, accompanied by the addition of the new sample set to maintain calibration set size stability. In calculating set similarity, a multi-indices fusion method is employed to ensure accurate judgments. The efficacy of this method is validated through simulated and real data.

Robust deep learning based shrimp counting in an industrial farm setting

Shrimp production is one of the fastest growing sectors in the aquaculture industry. Despite extensive research in recent years, stocking densities in shrimp systems still depend on manual sampling which is neither time nor cost efficient and additionally challenges shrimp welfare. This paper compares the performance of automatic shrimp counting solutions for commercial Recirculating Aquaculture System (RAS) based farming systems, using eight Deep Learning based methods. The entire dataset includes 1379 images of shrimps in RAS farming tanks, taken at a distance using an iPhone 11 mini. These were manually annotated, with bounding boxes for every clearly visible shrimp. The dataset was partitioned into training (60 %, 828 samples), validation (20 %, 276 samples) and test (20 %, 275 samples) splits for purposes of training and evaluating the models. The present work demonstrates that state-of-the-art object detection models outperform manual counting and achieve high performance across the entire production range and at various circumstances known to be challenging for object detection (dim light, overlapping and small animals, various acquisition devices and image resolutions and camera distance to object). Highest counting performance was obtained with models based on YOLOv5m6 and Faster R–CNN (as opposed to neural network autoencoder architecture to estimate a density map). The best model generalizes well on an independent test set and even shows promising performance when tested with different taxa. The model performs best at densities below 200 shrimps per image with an overall error of 5.97 %. It is assumed that this performance can be improved by increasing the dataset size, especially with images at high shrimp stocking density, and it is strongly believed that a performance below the 5 % error threshold is close to being achieved, which will allow for deployment of the model in an industrial setting.

FSDC: Flow Samples and Dimensions Compression for Efficient Detection of DNS-over-HTTPS Tunnels

This paper proposes an innovative approach capitalized on the distinctive characteristics of command and control (C&C) beacons, namely, time intervals and frequency between consecutive unique connections, to compress the network flow dataset. While previous studies on the same matter used single technique, we propose a multi-technique approach for efficient detection of DoH tunnels. We use a baseline public dataset, CIRA-CIC-DoHBrw-2020, containing over a million network flow properties and statistical features of DoH, tunnels, benign DoH and normal browsing (HTTPS) traffic. Each sample is represented by 33 features with a timestamp. Our methodology combines star graph and bar plot visualizations with supervised and unsupervised learning techniques. The approach underscores the importance of C&C beacon characteristic features in compressing a dataset and reducing a flow dimension while enabling efficient detection of DoH tunnels. Through compression, the original dataset size and dimensions are reduced by approximately 95% and 94% respectively. For supervised learning, RF emerges as the top-performing algorithm, attaining precision and recall scores of 100% each, with speed increase of ≈6796 times faster in training and ≈55 in testing. For anomaly detection models, OCSVM emerges as the most suitable choice for this purpose, with precision (88.89) and recall (100). Star graph and bar graph models also show a clear difference between normal traffic and DoH tunnels. The reduction in flow sample size and dimension, while maintaining accuracy, holds promise for edge networks with constrained resources and aids security analysts in interpreting complex ML models to identify Indicators of Compromise (IoC).

Self-attention CNN based indoor human events detection with UWB radar

In the era of smart homes and healthcare automation, the ability to accurately monitor and detect indoor human activities is paramount. Ultra-wideband (UWB) radar has emerged as a promising means for event detection, given its non-invasive nature and easy deployment in diverse environments. However, despite the advances in radar-based event detection, challenges remain, such as distinguishing between similar events like falls and rapid sitting. To address these challenges, for the first time, we propose an impulse radio-ultrawideband (IR-UWB) radar system to collect over ten thousand radar echo signals of eight similar actions from different angles and design a self-attention-based low-complexity convolutional neural network (CNN) model for event classification. The model leverages global correlations in radar signal spectrograms to efficiently extract features. A comparative simulation study is conducted to evaluate the detection accuracy of the proposed model and some of the existing methods based on different dataset sizes and CNN configurations. Moreover, the influence of different self-attention structures on precision and model parameter count is analyzed. Our findings reveal that the proposed self-attention-based CNN model significantly outperforms other traditional machine learning techniques while maintaining a low level computational complexity.

Advancing medical imaging with GAN-based anomaly detection

Anomaly detection in medical imaging is a complex challenge, exacerbated by limited annotated data. Recent advancements in generative adversarial networks (GANs) offer potential solutions, yet their effectiveness in medical imaging remains largely uncharted. We conducted a targeted exploration of the benefits and constraints associated with GAN-based anomaly detection techniques. Our investigations encompassed experiments employing eight anomaly detection methods on three medical imaging datasets representing diverse modalities and organ/tissue types. These experiments yielded notably diverse results. The results exhibited significant variability, with metrics spanning a wide range (area under the curve (AUC): 0.475-0.991; sensitivity: 0.17-0.98; specificity: 0.14-0.97). Furthermore, we offer guidance for implementing anomaly detection models in medical imaging and anticipate pivotal avenues for future research. Results unveil varying performances, influenced by factors like dataset size, anomaly subtlety, and dispersion. Our findings provide insights into the complex landscape of anomaly detection in medical imaging, offering recommendations for future research and deployment.

Generalizable clinical note section identification with large language models.

Clinical note section identification helps locate relevant information and could be beneficial for downstream tasks such as named entity recognition. However, the traditional supervised methods suffer from transferability issues. This study proposes a new framework for using large language models (LLMs) for section identification to overcome the limitations. We framed section identification as question-answering and provided the section definitions in free-text. We evaluated multiple LLMs off-the-shelf without any training. We also fine-tune our LLMs to investigate how the size and the specificity of the fine-tuning dataset impacts model performance. GPT4 achieved the highest F1 score of 0.77. The best open-source model (Tulu2-70b) achieved 0.64 and is on par with GPT3.5 (ChatGPT). GPT4 is also found to obtain F1 scores greater than 0.9 for 9 out of the 27 (33%) section types and greater than 0.8 for 15 out of 27 (56%) section types. For our fine-tuned models, we found they plateaued with an increasing size of the general domain dataset. We also found that adding a reasonable amount of section identification examples is beneficial. These results indicate that GPT4 is nearly production-ready for section identification, and seemingly contains both knowledge of note structure and the ability to follow complex instructions, and the best current open-source LLM is catching up. Our study shows that LLMs are promising for generalizable clinical note section identification. They have the potential to be further improved by adding section identification examples to the fine-tuning dataset.

Variational autoencoder-based dimension reduction of Ichimoku features for improved financial market analysis

Financial markets are complex and dynamic, and accurately predicting market trends is crucial for traders and financial analysts. Ichimoku-based features have gained significant attention in financial market analysis due to their ability to capture essential market signals and patterns. This significant compression retains essential patterns related to trends, support/resistance levels, and trading signals. The reduced dimensionality improves computational efficiency and could allow for more accurate predictive modeling by traders. However, real-world testing is needed because compressing data risks losing useful nuances. In this study, we utilize an autoencoder for the dimensionality reduction of Ichimoku-based features in financial market analysis. The autoencoder, a neural network architecture, compresses high-dimensional data into a lower-dimensional representation by learning important features and patterns. The experiments conducted on a Euro/Dollar market dataset spanning 1990, comprising 16 columns with Ichimoku features, reveal the remarkable reduction of dataset size from 2,269,500 to 756,375, equivalent to a decrease of 66.67 %. These results highlight the efficiency of the proposed approach in reducing the dimensionality of financial market data, suggesting its potential as a valuable tool for traders and financial analysts to predict market trends and make informed decisions in the financial markets.

A Federated Learning-Based Industrial Health Prognostics for Heterogeneous Edge Devices Using Matched Feature Extraction

Data-driven industrial health prognostics require rich training data to develop accurate and reliable predictive models. However, stringent data privacy laws and the abundance of edge industrial data necessitate decentralized data utilization. Thus, the industrial health prognostics field is well suited to significantly benefit from federated learning (FL), a decentralized and privacy-preserving learning technique. However, FL-based health prognostics tasks have hardly been investigated due to the complexities of meaningfully aggregating model parameters trained from heterogeneous data to form a high performing federated model. Specifically, data heterogeneity among edge devices, stemming from dissimilar degradation mechanisms and unequal dataset sizes, poses a critical statistical challenge for developing accurate federated models. We propose a pioneering FL-based health prognostic model with a feature similarity-matched parameter aggregation algorithm to discriminatingly learn from heterogeneous edge data. The algorithm searches across the heterogeneous locally trained models and matches neurons with probabilistically similar feature extraction functions first, before selectively averaging them to form the federated model parameters. As the algorithm only averages similar neurons, as opposed to conventional naive averaging of coordinate-wise neurons, the distinct feature extractors of local models are carried over with less dilution to the resultant federated model. Using both cyclic degradation data of Li-ion batteries and non-cyclic data of turbofan engines, we demonstrate that the proposed method yields accuracy improvements as high as 44.5\% and 39.3\% for state-of-health estimation and remaining useful life estimation, respectively.

Transferable machine learning model for the aerodynamic prediction of swept wings

With their development, machine learning models can be used instead of computational fluid dynamics simulations to predict flow fields in aerodynamic optimization. However, it is difficult to construct a prediction model for swept wings with various planform geometries because too many samples are required to cover the parameter space. In the present paper, a new model framework is proposed to predict wing surface pressure and friction distributions with fewer samples. The distributed geometry parameters along spanwise are used as model inputs instead of the global planform parameters, and processors are designed to help the model better learn the local effect of geometric variation. The model is trained and tested on simple swept wings with single segment and linear twist distribution, where it outperforms the global input model by 57.6% in terms of lift coefficient prediction errors on small dataset sizes. The distributed input also enables the model to be transferred from single wings to more engineering-practical yet complex kink wings. After fine-tuning with a few samples, model accuracy for kink wings can be similar to that of simple wings, which proves the model for wings with complex planform geometries can be efficiently built with the proposed method.

Early Classification of Emphysema in CT Scans Using Deep Learning

Abstract: Learning about Lung Diseases and their characterization is one of the most interesting research topics in recent years. With the various uses of medical images in hospitals, pathologies, and diagnostic centers, the size of the medical image datasets is also expanding expeditiously to capture the diseases in hospitals. Though a lot of research has been done on this particular topic still this field is confusing and challenging. There are lots of techniques in literature to classify medical images. The main drawback of traditional methods is the semantic gap that exists between the low-level visual information captured by imaging devices and high-level semantic information perceived by a human being. The difficulty of querying and managing the large datasets leads to a new mechanism called deep convolutional neural network. Chronic Obstructive Pulmonary Disease (COPD) is a prevalent respiratory disorder characterized by airflow limitation, often caused by emphysema. Early detection and classification of emphysema severity play a crucial role in effective treatment and management. Deep learning techniques have recently achieved an impressive result in the field of computer vision along with Medical Engineering. Our aim was to determine whether participant-level emphysema pattern could predict impairment and mortality when classified by using a deep learning method.

RESEARCH OF REPRESENTATIVENESS OF KAZAKH LANGUAGE CORPORA BY WORD STEMS FOR THE SUMMARIZATION

In this work, we investigated the dependence of the work of the summarization model on the number of word stems in it. The work was performed on a synthetic summarization dataset for the Kazakh language. Taking the number of word stems as a metric of representativeness, an analysis of the quality of work of three summation models was performed depending on the number of word stems in the training dataset. To obtain three datasets, we divided the training dataset into three parts. BLEU estimates were obtained for each model on the test files. The experimental part of the work showed that the model with the largest number of stems shows the highest BLEU score. But the score does not directly depend on the number of word stems. Two models trained on datasets of different sizes show approximately the same scores.

Enhancing Medical Imaging Segmentation with GB-SAM: A Novel Approach to Tissue Segmentation Using Granular Box Prompts.

Recent advances in foundation models have revolutionized model development in digital pathology, reducing dependence on extensive manual annotations required by traditional methods. The ability of foundation models to generalize well with few-shot learning addresses critical barriers in adapting models to diverse medical imaging tasks. This work presents the Granular Box Prompt Segment Anything Model (GB-SAM), an improved version of the Segment Anything Model (SAM) fine-tuned using granular box prompts with limited training data. The GB-SAM aims to reduce the dependency on expert pathologist annotators by enhancing the efficiency of the automated annotation process. Granular box prompts are small box regions derived from ground truth masks, conceived to replace the conventional approach of using a single large box covering the entire H&E-stained image patch. This method allows a localized and detailed analysis of gland morphology, enhancing the segmentation accuracy of individual glands and reducing the ambiguity that larger boxes might introduce in morphologically complex regions. We compared the performance of our GB-SAM model against U-Net trained on different sizes of the CRAG dataset. We evaluated the models across histopathological datasets, including CRAG, GlaS, and Camelyon16. GB-SAM consistently outperformed U-Net, with reduced training data, showing less segmentation performance degradation. Specifically, on the CRAG dataset, GB-SAM achieved a Dice coefficient of 0.885 compared to U-Net's 0.857 when trained on 25% of the data. Additionally, GB-SAM demonstrated segmentation stability on the CRAG testing dataset and superior generalization across unseen datasets, including challenging lymph node segmentation in Camelyon16, which achieved a Dice coefficient of 0.740 versus U-Net's 0.491. Furthermore, compared to SAM-Path and Med-SAM, GB-SAM showed competitive performance. GB-SAM achieved a Dice score of 0.900 on the CRAG dataset, while SAM-Path achieved 0.884. On the GlaS dataset, Med-SAM reported a Dice score of 0.956, whereas GB-SAM achieved 0.885 with significantly less training data. These results highlight GB-SAM's advanced segmentation capabilities and reduced dependency on large datasets, indicating its potential for practical deployment in digital pathology, particularly in settings with limited annotated datasets.

A New Alternating Suboptimal Dynamic Programming Algorithm with Applications for Feature Selection

Feature selection is predominantly used in machine learning tasks, such as classification, regression, and clustering. It selects a subset of features (relevant attributes of data points) from a larger set that contributes as optimally as possible to the informativeness of the model. There are exponentially many subsets of a given set, and thus, the exhaustive search approach is only practical for problems with at most a few dozen features. In the past, there have been attempts to reduce the search space using dynamic programming. However, models that consider similarity in pairs of features alongside the quality of individual features do not provide the required optimal substructure. As a result, algorithms, which we will call suboptimal dynamic programming algorithms, find a solution that may deviate significantly from the optimal one. In this paper, we propose an iterative dynamic programming algorithm, which invertsthe order of feature processing in each iteration. Such an alternating approach allows for improving the optimization function by using the score from the previous iteration to estimate the contribution of unprocessed features. The iterative process is proven to converge and terminates when the solution does not change in three successive iterations or when the number of iterations reaches the threshold. Results in more than 95% of tests align with those of the exhaustive search approach, being competitive and often superior to the reference greedy approach. Validation was carried out by comparing the scores of output feature subsets and examining the accuracy of different classifiers learned on these features across nine real-world applications, considering different scenarios with various numbers of features and samples. In the context of feature selection, the proposed algorithm can be characterized as a robust filter method that can improve machine learning models regardless of dataset size. However, we expect that the idea of alternating suboptimal optimization will soon be generalized to tasks beyond feature selection.

Power of a test for assessing interlaboratory consensus of nominal and ordinal characteristics of a substance, material, or object

A concept of the consensus among different laboratories participating in an interlaboratory comparison, classifying a substance, material, or object according to its nominal and ordinal (i.e. categorical) characteristics, is devised using decomposition of the total variation of the laboratory responses. One of the components of the total variation is caused by the between-laboratory differences, and the second—by conditions associated with the applied experimental design (for example, temperature of test items, technician experience, etc). This decomposition is based on the recently developed two-way CATANOVA for nominal variables and two-way ORDANOVA for ordinal variables. The consensus is tested as hypotheses about homogeneity, i.e. insignificance of the corresponding components of the total variation. The consensus power is taken to be the power of the homogeneity test. A methodology for evaluation of the consensus power and corresponding risks of false decisions versus the dataset size of categorical characteristics obtained in an interlaboratory comparison is detailed. Examples of evaluation of the power and risks are discussed using previously-published datasets of an interlaboratory comparison of identification of weld imperfections, and an examination of the intensity of the odor of drinking water. An example of computer code in the R programming environment is presented for the power calculations in the case of nominal variables, using a chi-square distribution. A newly developed tool for ordinal variables, an Excel spreadsheet with macros, which is based on Monte Carlo draws from a multinomial distribution, is also available.

Optimizing Performance of Transformer-based Models for Fetal Brain MR Image Segmentation.

Purpose To test the performance of a transformer-based model when manipulating pretraining weights, dataset size, and input size and comparing the best model with the reference standard and state-of-the-art models for a resting-state functional (rs-fMRI) fetal brain extraction task. Materials and Methods An internal retrospective dataset (172 fetuses, 519 images; collected 2018-2022) was used to investigate influence of dataset size, pretraining approaches, and image input size on Swin-U-Net transformer (UNETR) and UNETR models. The internal and external (131 fetuses, 561 images) datasets were used to cross-validate and to assess generalization capability of the best model versus state-of-the-art models on different scanner types and number of gestational weeks (GWs). The Dice similarity coefficient (DSC) and the balanced average Hausdorff distance (BAHD) were used as segmentation performance metrics. Generalized equation estimation multifactorial models were used to assess significant model and interaction effects of interest. Results The Swin-UNETR model was not affected by the pretraining approach and dataset size and performed best with the mean dataset image size, with a mean DSC of 0.92 and BAHD of 0.097. Swin-UNETR was not affected by scanner type. Generalization results on the internal dataset showed that Swin-UNETR had lower performance compared with the reference standard models and comparable performance on the external dataset. Cross-validation on internal and external test sets demonstrated better and comparable performance of Swin-UNETR versus convolutional neural network architectures during the late-fetal period (GWs > 25) but lower performance during the midfetal period (GWs ≤ 25). Conclusion Swin-UNTER showed flexibility in dealing with smaller datasets, regardless of pretraining approaches. For fetal brain extraction from rs-fMR images, Swin-UNTER showed comparable performance with that of reference standard models during the late-fetal period and lower performance during the early GW period. Keywords: Transformers, CNN, Medical Imaging Segmentation, MRI, Dataset Size, Input Size, Transfer Learning Supplemental material is available for this article. © RSNA, 2024.

Temporal heterogeneity in the performance of machine learning models for PM2.5 concentration estimation

Machine learning (ML) methods have been applied extensively to simulate air pollutant concentrations and assess individual exposure in epidemiological studies. However, there is still a paucity of research on the temporal heterogeneity of ML model performance and the impact of dataset size. To explore the temporal heterogeneity in model performance when estimating daily concentrations of fine particulate matter (PM2.5) across China in 2021, we compared five decision tree-based ML models (Random Forest (RF), Categorical Boosting (CatBoost), Gradient Boost Regression Tree (GBRT), eXtreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)) across daily scales within three distinct timeframes. The performance of all models was evaluated using cross-validation. We observed that the performance of ML models varied with time, which showed a significant correlation with PM2.5 concentration. Among the 365 days in 2021, RF model performed best, the annual mean R2 was 0.86, with a minimum of 0.84 and a maximum of up to 0.95. For RF, we chose a cubic polynomial curve to fit the relationship between model performance and PM2.5 concentrations, and based on this, we devised a model selection strategy for different time scales, achieving an accuracy rate of up to 79.45 %, with the selected models having an average R2 of 0.85, and a maximum of up to 0.95. Additionally, we found that increasing the dataset size did not significantly improve model performance. Instead, it resulted in considerably longer runtime and increased memory usage. The methodology and findings of this study hold significant value for advancing the development of more efficient and precise modeling approaches for air pollutant concentrations. Furthermore, this research provides a foundation for regional air pollutant governance and future health-related research.

Adaptive Randomization Method to Prevent Extreme Instances of Group Size and Covariate Imbalance in Stroke Trials.

A recent review of randomization methods used in large multicenter clinical trials within the National Institutes of Health Stroke Trials Network identified preservation of treatment allocation randomness, achievement of the desired group size balance between treatment groups, achievement of baseline covariate balance, and ease of implementation in practice as critical properties required for optimal randomization designs. Common-scale minimal sufficient balance (CS-MSB) adaptive randomization effectively controls for covariate imbalance between treatment groups while preserving allocation randomness but does not balance group sizes. This study extends the CS-MSB adaptive randomization method to achieve both group size and covariate balance while preserving allocation randomness in hyperacute stroke trials. A full factorial in silico simulation study evaluated the performance of the proposed new CSSize-MSB adaptive randomization method in achieving group size balance, covariate balance, and allocation randomness compared with the original CS-MSB method. Data from 4 existing hyperacute stroke trials were used to investigate the performance of CSSize-MSB for a range of sample sizes and covariate numbers and types. A discrete-event simulation model created with AnyLogic was used to dynamically visualize the decision logic of the CSSize-MSB randomization process for communication with clinicians. The proposed new CSSize-MSB algorithm uniformly outperformed the CS-MSB algorithm in controlling for group size imbalance while maintaining comparable levels of covariate balance and allocation randomness in hyperacute stroke trials. This improvement was consistent across a distribution of simulated trials with varying levels of imbalance but was increasingly pronounced for trials with extreme cases of imbalance. The results were consistent across a range of trial data sets of different sizes and covariate numbers and types. The proposed adaptive CSSize-MSB algorithm successfully controls for group size imbalance in hyperacute stroke trials under various settings, and its logic can be readily explained to clinicians using dynamic visualization.

Multi-class Waste Classification Using Convolutional Neural Network

This study explores a method of waste classification using deep learning, specifically employing the Convolutional Neural Network (CNN). This research involves the creation of a unique dataset, a hybrid of publicly accessible data and a newly compiled collection of images across 13 waste classes: paper, glass, wood, metal, clothes, PCB e-waste, non-PCB e-waste, PET, HDPE, LDPE, PP, PVC, and PS. The development of the CNN model was approached in two ways: transfer learning and full learning. In the transfer learning approach, two pre-trained models, MobileNetV2 and DenseNet121, were utilized. While in the full learning approach, the architecture is constructed using the sequential method. The experimental results indicated that the DenseNet121 transfer learning model outperformed others, achieving an impressive accuracy of 95.2% and an average F-1 score of 0.95 on test data. This was closely followed by the MobileNetV2 transfer learning model, which attained an accuracy of 92% and an average F-1 score of 0.92. In comparison, the full learning model reached an accuracy of 65% and an average F-1 score of 0.65. Generally, transfer learning models yielded more optimal results than those full learning model. This efficiency can be attributed to the pre-existing knowledge in the transfer learning models, which eliminates the need to learn input patterns from the ground up. However, it's important to note that the dataset size of 4586 images across 13 classes may not be sufficient for developing a robust machine learning model from scratch.

Classification of Cardiovascular Arrhythmia Using Deep Learning Techniques: A Review

Deep Learning (DL), an offshoot of Machine Learning (ML) has emerged as a powerful and feasible solution for medical image analysis due to advancements in robust computer software and hardware technologies. It plays a key role in Cardiovascular disease (CVD) diagnosis by detecting anomalies in Electrocardiogram (ECG) signals. Cardiac arrhythmia, which refers to irregular heartbeat, may signal an early symptom of CVD and can lead to fatal outcomes if ignored. Accurate detection of arrhythmia is very challenging even for experts to distinguish between acute and chronic conditions in ECG readings. This triggered the focus of researchers to explore the application of Artificial Intelligence for ECG classification. Traditional machine learning methods use handcrafted features that require domain knowledge. The new era in DL makes the automatic detection of Cardiovascular Disease (CVD) possible. In this paper, an exhaustive review of DL-based techniques for ECG classification has been presented. Research findings in this survey indicate the challenges and issues with arrhythmia detection, such as single lead and multiple lead ECG signals, choice of the size of the training data set, and the number of arrhythmia classes, etc. The study also signifies that there is great scope for improving the performance of arrhythmia prediction models by employing hybrid ensemble learning, time series analysis using Recurrent Neural Network architectures and identification of unexplored classes of arrhythmia.