Corrigendum: SemNet: Learning semantic attributes for human activity recognition with deep belief networks.

[This corrects the article DOI: 10.3389/fdata.2022.879389.].

A Study on Different Deep Learning Algorithms Used in Deep Neural Nets: MLP SOM and DBN.

Deep learning is a wildly popular topic in machine learning and is structured as a series of nonlinear layers that learns various levels of data representations. Deep learning employs numerous layers to represent data abstractions to implement various computer models. Deep learning approaches like generative, discriminative models and model transfer have transformed information processing. This article proposes a comprehensive review of various deep learning algorithms Multi layer perception, Self-organizing map and deep belief networks algorithms. It first briefly introduces historical and recent state-of-the-art reviews with suitable architectures and implementation steps. Moreover, the various applications of those algorithms in various fields such as wireless networks, Adhoc networks, Mobile ad-hoc and vehicular ad-hoc networks, speech recognition engineering, medical applications, natural language processing, material science and remote sensing applications, etc. are classified.

A deep belief network-based clinical decision system for patients with osteosarcoma.

Osteosarcoma was the most frequent type of malignant primary bone tumor with a poor survival rate mainly occurring in children and adolescents. For precision treatment, an accurate individualized prognosis for Osteosarcoma patients is highly desired. In recent years, many machine learning-based approaches have been used to predict distant metastasis and overall survival based on available individual information. In this study, we compared the performance of the deep belief networks (DBN) algorithm with six other machine learning algorithms, including Random Forest, XGBoost, Decision Tree, Gradient Boosting Machine, Logistic Regression, and Naive Bayes Classifier, to predict lung metastasis for Osteosarcoma patients. Therefore the DBN-based lung metastasis prediction model was integrated as a parameter into the Cox proportional hazards model to predict the overall survival of Osteosarcoma patients. The accuracy, precision, recall, and F1 score of the DBN algorithm were 0.917/0.888, 0.896/0.643, 0.956/0.900, and 0.925/0.750 in the training/validation sets, respectively, which were better than the other six machine-learning algorithms. For the performance of the DBN survival Cox model, the areas under the curve (AUCs) for the 1-, 3- and 5-year survival in the training set were 0.851, 0.806 and 0.793, respectively, indicating good discrimination, and the calibration curves showed good agreement between the prediction and actual observations. The DBN survival Cox model also demonstrated promising performance in the validation set. In addition, a nomogram integrating the DBN output was designed as a tool to aid clinical decision-making.

SemNet: Learning semantic attributes for human activity recognition with deep belief networks.

Human Activity Recognition (HAR) is a prominent application in mobile computing and Internet of Things (IoT) that aims to detect human activities based on multimodal sensor signals generated as a result of diverse body movements. Human physical activities are typically composed of simple actions (such as "arm up", "arm down", "arm curl", etc.), referred to as semantic features. Such abstract semantic features, in contrast to high-level activities ("walking", "sitting", etc.) and low-level signals (raw sensor readings), can be developed manually to assist activity recognition. Although effective, this manual approach relies heavily on human domain expertise and is not scalable. In this paper, we address this limitation by proposing a machine learning method, SemNet, based on deep belief networks. SemNet automatically constructs semantic features representative of the axial bodily movements. Experimental results show that SemNet outperforms baseline approaches and is capable of learning features that highly correlate with manually defined semantic attributes. Furthermore, our experiments using a different model, namely deep convolutional LSTM, on household activities illustrate the broader applicability of semantic attribute interpretation to diverse deep neural network approaches. These empirical results not only demonstrate that such a deep learning technique is semantically meaningful and superior to its handcrafted counterpart, but also provides a better understanding of the deep learning methods that are used for Human Activity Recognition.

Approximation properties of Gaussian-binary restricted Boltzmann machines and Gaussian-binary deep belief networks.

Despite the successful use of Gaussian-binary restricted Boltzmann machines (GB-RBMs) and Gaussian-binary deep belief networks (GB-DBNs), little is known about their theoretical approximation capabilities to represent distributions of continuous random variables. In this paper, we address the expressive properties of GB-RBMs and GB-DBNs, contributing theoretical insights to the optimal number of hidden variables. We first treat the GB-RBM's unnormalized log-likelihood as a sum of a special two-layer feedforward neural network and a negative quadratic term. Then, a series of simulation results are established, which can be used to relate GB-RBMs to general two-layer feedforward neural networks whose expressive properties are much better understood. On this basis, we show that a two-layer ReLU network with all weights in the second layer being 1, along with a negative quadratic term, can approximate all continuous functions. In addition, we provide qualified lower bounds for the number of hidden variables of GB-RBMs required to approximate distributions whose log-likelihood are given by some classes of smooth functions. Moreover, we further study the universal approximation of GB-DBNs with two hidden layers by providing a sufficient number of hidden variables O(ɛ-2) that are guaranteed to approximate any given strictly positive continuous distribution within a given error ɛ. Finally, numerical experiments are carried out to verify some of the proposed theoretical results.

Image Classification and Recognition of Rice Diseases: A Hybrid DBN and Particle Swarm Optimization Algorithm.

Rice blast, rice sheath blight, and rice brown spot have become the most popular diseases in the cold areas of northern China. In order to further improve the accuracy and efficiency of rice disease diagnosis, a framework for automatic classification and recognition of rice diseases is proposed in this study. First, we constructed a training and testing data set including 1,500 images of rice blast, 1,500 images of rice sheath blight, and 1,500 images of rice brown spot, and 1,100 healthy images were collected from the rice experimental field. Second, the deep belief network (DBN) model is designed to include 15 hidden restricted Boltzmann machine layers and a support vector machine (SVM) optimized with switching particle swarm (SPSO). It is noted that the developed DBN and SPSO-SVM can simultaneously learn three proposed features including color, texture, and shape to recognize the disease type from the region of interest obtained by preprocessing the disease images. The proposed model leads to a hit rate of 91.37%, accuracy of 94.03%, and a false measurement rate of 8.63%, with the 10-fold cross-validation strategy. The value of the area under the receiver operating characteristic curve (AUC) is 0.97, whose accuracy is much higher than that of the conventional machine learning model. The simulation results show that the DBN and SPSO-SVM models can effectively extract the image features of rice diseases during recognition, and have good anti-interference and robustness.

An Integrated Deep Network for Cancer Survival Prediction Using Omics Data.

As a highly sophisticated disease that humanity faces, cancer is known to be associated with dysregulation of cellular mechanisms in different levels, which demands novel paradigms to capture informative features from different omics modalities in an integrated way. Successful stratification of patients with respect to their molecular profiles is a key step in precision medicine and in tailoring personalized treatment for critically ill patients. In this article, we use an integrated deep belief network to differentiate high-risk cancer patients from the low-risk ones in terms of the overall survival. Our study analyzes RNA, miRNA, and methylation molecular data modalities from both labeled and unlabeled samples to predict cancer survival and subsequently to provide risk stratification. To assess the robustness of our novel integrative analytics, we utilize datasets of three cancer types with 836 patients and show that our approach outperforms the most successful supervised and semi-supervised classification techniques applied to the same cancer prediction problems. In addition, despite the preconception that deep learning techniques require large size datasets for proper training, we have illustrated that our model can achieve better results for moderately sized cancer datasets.

Deep Neural Networks for Depression Recognition Based on 2D and 3D Facial Expressions Under Emotional Stimulus Tasks.

The proportion of individuals with depression has rapidly increased along with the growth of the global population. Depression has been the currently most prevalent mental health disorder. An effective depression recognition system is especially crucial for the early detection of potential depression risk. A depression-related dataset is also critical while evaluating the system for depression or potential depression risk detection. Due to the sensitive nature of clinical data, availability and scale of such datasets are scarce. To our knowledge, there are few extensively practical depression datasets for the Chinese population. In this study, we first create a large-scale dataset by asking subjects to perform five mood-elicitation tasks. After each task, subjects' audio and video are collected, including 3D information (depth information) of facial expressions via a Kinect. The constructed dataset is from a real environment, i.e., several psychiatric hospitals, and has a specific scale. Then we propose a novel approach for potential depression risk recognition based on two kinds of different deep belief network (DBN) models. One model extracts 2D appearance features from facial images collected by an optical camera, while the other model extracts 3D dynamic features from 3D facial points collected by a Kinect. The final decision result comes from the combination of the two models. Finally, we evaluate all proposed deep models on our built dataset. The experimental results demonstrate that (1) our proposed method is able to identify patients with potential depression risk; (2) the recognition performance of combined 2D and 3D features model outperforms using either 2D or 3D features model only; (3) the performance of depression recognition is higher in the positive and negative emotional stimulus, and females' recognition rate is generally higher than that for males. Meanwhile, we compare the performance with other methods on the same dataset. The experimental results show that our integrated 2D and 3D features DBN is more reasonable and universal than other methods, and the experimental paradigm designed for depression is reasonable and practical.

Multi-step rainfall forecasting using deep learning approach.

Rainfall prediction is immensely crucial in daily life routine as well as for water resource management, stochastic hydrology, rain run-off modeling and flood risk mitigation. Quantitative prediction of rainfall time series is extremely challenging as compared to other meteorological parameters due to its variability in local features that involves temporal and spatial scales. Consequently, this requires a highly complex system having an advance model to accurately capture the highly non linear processes occurring in the climate. The focus of this work is direct prediction of multistep forecasting, where a separate time series model for each forecasting horizon is considered and forecasts are computed using observed data samples. Forecasting in this method is performed by proposing a deep learning approach, i.e, Temporal Deep Belief Network (DBN). The best model is selected from several baseline models on the basis of performance analysis metrics. The results suggest that the temporal DBN model outperforms the conventional Convolutional Neural Network (CNN) specifically on rainfall time series forecasting. According to our experimentation, a modified DBN with hidden layes (300-200-100-10) performs best with 4.59E-05, 0.0068 and 0.94 values of MSE, RMSE and R value respectively on testing samples. However, we found that training DBN is more exhaustive and computationally intensive than other deep learning architectures. Findings of this research can be further utilized as basis for the advance forecasting of other weather parameters with same climate conditions.

A Multi-Level Output-Based DBN Model for Fine Classification of Complex Geo-Environments Area Using Ziyuan-3 TMS Imagery.

Fine-scale land use and land cover (LULC) data in a mining area are helpful for the smart supervision of mining activities. However, the complex landscape of open-pit mining areas severely restricts the classification accuracy. Although deep learning (DL) algorithms have the ability to extract informative features, they require large amounts of sample data. As a result, the design of more interpretable DL models with lower sample demand is highly important. In this study, a novel multi-level output-based deep belief network (DBN-ML) model was developed based on Ziyuan-3 imagery, which was applied for fine classification in an open-pit mine area of Wuhan City. First, the last DBN layer was used to output fine-scale land cover types. Then, one of the front DBN layers outputted the first-level land cover types. The coarse classification was easier and fewer DBN layers were sufficient. Finally, these two losses were weighted to optimize the DBN-ML model. As the first-level class provided a larger amount of additional sample data with no extra cost, the multi-level output strategy enhanced the robustness of the DBN-ML model. The proposed model produces an overall accuracy of 95.10% and an F1-score of 95.07%, outperforming some other models.

[Detection of microcalcification clusters regions in mammograms combining discriminative deep belief networks].

In order to overcome the shortcomings of high false positive rate and poor generalization in the detection of microcalcification clusters regions, this paper proposes a method combining discriminative deep belief networks (DDBNs) to automatically and quickly locate the regions of microcalcification clusters in mammograms. Firstly, the breast region was extracted and enhanced, and the enhanced breast region was segmented to overlapped sub-blocks. Then the sub-block was subjected to wavelet filtering. After that, DDBNs model for breast sub-block feature extraction and classification was constructed, and the pre-trained DDBNs was converted to deep neural networks (DNN) using a softmax classifier, and the network is fine-tuned by back propagation. Finally, the undetected mammogram was inputted to complete the location of suspicious lesions. By experimentally verifying 105 mammograms with microcalcifications from the Digital Database for Screening Mammography (DDSM), the method obtained a true positive rate of 99.45% and a false positive rate of 1.89%, and it only took about 16 s to detect a 2 888 × 4 680 image. The experimental results showed that the algorithm of this paper effectively reduced the false positive rate while ensuring a high positive rate. The detection of calcification clusters was highly consistent with expert marks, which provides a new research idea for the automatic detection of microcalcification clusters area in mammograms.

Detection of microcalcification clusters regions in mammograms combining discriminative deep belief networks / 生物医学工程学杂志

In order to overcome the shortcomings of high false positive rate and poor generalization in the detection of microcalcification clusters regions, this paper proposes a method combining discriminative deep belief networks (DDBNs) to automatically and quickly locate the regions of microcalcification clusters in mammograms. Firstly, the breast region was extracted and enhanced, and the enhanced breast region was segmented to overlapped sub-blocks. Then the sub-block was subjected to wavelet filtering. After that, DDBNs model for breast sub-block feature extraction and classification was constructed, and the pre-trained DDBNs was converted to deep neural networks (DNN) using a softmax classifier, and the network is fine-tuned by back propagation. Finally, the undetected mammogram was inputted to complete the location of suspicious lesions. By experimentally verifying 105 mammograms with microcalcifications from the Digital Database for Screening Mammography (DDSM), the method obtained a true positive rate of 99.45% and a false positive rate of 1.89%, and it only took about 16 s to detect a 2 888 × 4 680 image. The experimental results showed that the algorithm of this paper effectively reduced the false positive rate while ensuring a high positive rate. The detection of calcification clusters was highly consistent with expert marks, which provides a new research idea for the automatic detection of microcalcification clusters area in mammograms.

Influence of Primary Auditory Cortex in the Characterization of Autism Spectrum in Young Adults using Brain Connectivity Parameters and Deep Belief Networks: An fMRI Study.

BACKGROUND: The spectrum of autism encompasses High Functioning Autism (HFA) and Low Functioning Autism (LFA). Brain mapping studies have revealed that autism individuals have overlaps in brain behavioural characteristics. Generally, high functioning individuals are known to exhibit higher intelligence and better language processing abilities. However, specific mechanisms associated with their functional capabilities are still under research. OBJECTIVE: This work addresses the overlapping phenomenon present in autism spectrum through functional connectivity patterns along with brain connectivity parameters and distinguishes the classes using deep belief networks. METHODS: The task-based functional Magnetic Resonance Images (fMRI) of both high and low functioning autistic groups were acquired from ABIDE database, for 58 low functioning against 43 high functioning individuals while they were involved in a defined language processing task. The language processing regions of the brain, along with Default Mode Network (DMN) have been considered for the analysis. The functional connectivity maps have been plotted through graph theory procedures. Brain connectivity parameters such as Granger Causality (GC) and Phase Slope Index (PSI) have been calculated for the individual groups. These parameters have been fed to Deep Belief Networks (DBN) to classify the subjects under consideration as either LFA or HFA. RESULTS: Results showed increased functional connectivity in high functioning subjects. It was found that the additional interaction of the Primary Auditory Cortex lying in the temporal lobe, with other regions of interest complimented their enhanced connectivity. Results were validated using DBN measuring the classification accuracy of 85.85% for high functioning and 81.71% for the low functioning group. CONCLUSION: Since it is known that autism involves enhanced, but imbalanced components of intelligence, the reason behind the supremacy of high functioning group in language processing and region responsible for enhanced connectivity has been recognized. Therefore, this work that suggests the effect of Primary Auditory Cortex in characterizing the dominance of language processing in high functioning young adults seems to be highly significant in discriminating different groups in autism spectrum.

Identification of vowels in consonant-vowel-consonant words from speech imagery based EEG signals.

Retrieval of unintelligible speech is a basic need for speech impaired and is under research for several decades. But retrieval of random words from thoughts needs a substantial and consistent approach. This work focuses on the preliminary steps of retrieving vowels from Electroencephalography (EEG) signals acquired while speaking and imagining of speaking a consonant-vowel-consonant (CVC) word. The process, referred to as Speech imagery is imagining of speaking to oneself silently in the mind. Speech imagery is a form of mental imagery. Brain connectivity estimators such as EEG coherence, Partial Directed Coherence, Directed Transfer Function and Transfer Entropy have been used to estimate the concurrency and causal dependence (direction and strength) between different brain regions. From brain connectivity results it has been observed that the left frontal and left temporal electrodes were activated for speech and speech imagery processes. These brain connectivity estimators have been used for training Recurrent Neural Networks (RNN) and Deep Belief Networks (DBN) for identifying the vowel from the subject's thought. Though the accuracy level was found to be varying for each vowel while speaking and imagining of speaking the CVC word, the overall classification accuracy was found to be 72% while using RNN whereas a classification accuracy of 80% was observed while using DBN. DBN was found to outperform RNN in both the speech and speech imagery processes. Thus, the combination of brain connectivity estimators and deep learning techniques appear to be effective in identifying the vowel from EEG signals of subjects' thought.

Improved Deep Learning Based Method for Molecular Similarity Searching Using Stack of Deep Belief Networks.

Virtual screening (VS) is a computational practice applied in drug discovery research. VS is popularly applied in a computer-based search for new lead molecules based on molecular similarity searching. In chemical databases similarity searching is used to identify molecules that have similarities to a user-defined reference structure and is evaluated by quantitative measures of intermolecular structural similarity. Among existing approaches, 2D fingerprints are widely used. The similarity of a reference structure and a database structure is measured by the computation of association coefficients. In most classical similarity approaches, it is assumed that the molecular features in both biological and non-biologically-related activity carry the same weight. However, based on the chemical structure, it has been found that some distinguishable features are more important than others. Hence, this difference should be taken consideration by placing more weight on each important fragment. The main aim of this research is to enhance the performance of similarity searching by using multiple descriptors. In this paper, a deep learning method known as deep belief networks (DBN) has been used to reweight the molecule features. Several descriptors have been used for the MDL Drug Data Report (MDDR) dataset each of which represents different important features. The proposed method has been implemented with each descriptor individually to select the important features based on a new weight, with a lower error rate, and merging together all new features from all descriptors to produce a new descriptor for similarity searching. Based on the extensive experiments conducted, the results show that the proposed method outperformed several existing benchmark similarity methods, including Bayesian inference networks (BIN), the Tanimoto similarity method (TAN), adapted similarity measure of text processing (ASMTP) and the quantum-based similarity method (SQB). The results of this proposed multi-descriptor-based on Stack of deep belief networks method (SDBN) demonstrated a higher accuracy compared to existing methods on structurally heterogeneous datasets.

Recent advances of deep learning in psychiatric disorders.

Deep learning (DL) is a recently proposed subset of machine learning methods that has gained extensive attention in the academic world, breaking benchmark records in areas such as visual recognition and natural language processing. Different from conventional machine learning algorithm, DL is able to learn useful representations and features directly from raw data through hierarchical nonlinear transformations. Because of its ability to detect abstract and complex patterns, DL has been used in neuroimaging studies of psychiatric disorders, which are characterized by subtle and diffuse alterations. Here, we provide a brief review of recent advances and associated challenges in neuroimaging studies of DL applied to psychiatric disorders. The results of these studies indicate that DL could be a powerful tool in assisting the diagnosis of psychiatric diseases. We conclude our review by clarifying the main promises and challenges of DL application in psychiatric disorders, and possible directions for future research.

Optimizing neural networks for medical data sets: A case study on neonatal apnea prediction.

OBJECTIVE: The neonatal period of a child is considered the most crucial phase of its physical development and future health. As per the World Health Organization, India has the highest number of pre-term births [1], with over 3.5 million babies born prematurely, and up to 40% of them are babies with low birth weights, highly prone to a multitude of diseases such as Jaundice, Sepsis, Apnea, and other Metabolic disorders. Apnea is the primary concern for caretakers of neonates in intensive care units. The real-time medical data is known to be noisy and nonlinear and to address the resultant complexity in classification and prediction of diseases; there is a need for optimizing learning models to maximize predictive performance. Our study attempts to optimize neural network architectures to predict the occurrence of apneic episodes in neonates, after the first week of admission to Neonatal Intensive Care Unit (NICU). The primary contribution of this study is the formulation and description of a set of generic steps involved in selecting various model-specific, training and hyper-parametric optimization algorithms, as well as model architectures for optimal predictive performance on complex and noisy medical datasets. METHODS: The data used for the study being inherently complex and noisy, Kernel Principal Component Analysis (PCA) is used to reduce dataset dimensionality for the analysis such as interpretations and visualization of the dataset. Hyper-parametric and parametric optimization, in different categories, are considered, including learning rate updater algorithms, regularization methods, activation functions, gradient descent algorithms and depth of the network, based on their performance on the validation set, to obtain a holistically optimized neural network, that best model the given complex medical dataset. Deep Neural Network Architectures such as Deep Multilayer Perceptron's, Stacked Auto-encoders and Deep Belief Networks are employed to model the dataset, and their performance is compared to the optimized neural network obtained from the parametric exploration. Further, the results are compared with Support Vector Machine (SVM), K Nearest Neighbor, Decision Tree (DT) and Random Forest (RF) algorithms. RESULTS: The results indicate that the optimized eight layer Multilayer Perceptron (MLP) model, with Adam Decay and Stochastic Gradient Descent (AUC 0.82) can outperform the conventional machine learning models, and perform comparably to the Deep Auto-encoder model (AUC 0.83) in predicting the presence of apnea in neonates. CONCLUSION: The study shows that an MLP model can undergo significant improvements in predictive performance, by the proposed step-wise optimization. The optimized MLP is proved to be as accurate as deep neural network models such as Deep Belief Networks and Deep Auto-encoders for noisy and nonlinear data sets, and outperform all conventional models like Support Vector Machine (SVM), Decision Tree (DT), K Nearest Neighbor and Random Forest (RF) algorithms. The generic nature of the proposed step-wise optimization provides a framework to optimize neural networks on such complex nonlinear datasets. The investigated models can help neonatologists as a diagnostic tool.

[Automatic classification method of arrhythmia based on discriminative deep belief networks].

Existing arrhythmia classification methods usually use manual selection of electrocardiogram (ECG) signal features, so that the feature selection is subjective, and the feature extraction is complex, leaving the classification accuracy usually affected. Based on this situation, a new method of arrhythmia automatic classification based on discriminative deep belief networks (DDBNs) is proposed. The morphological features of heart beat signals are automatically extracted from the constructed generative restricted Boltzmann machine (GRBM), then the discriminative restricted Boltzmann machine (DRBM) with feature learning and classification ability is introduced, and arrhythmia classification is performed according to the extracted morphological features and RR interval features. In order to further improve the classification performance of DDBNs, DDBNs are converted to deep neural network (DNN) using the Softmax regression layer for supervised classification in this paper, and the network is fine-tuned by backpropagation. Finally, the Massachusetts Institute of Technology and Beth Israel Hospital Arrhythmia Database (MIT-BIH AR) is used for experimental verification. For training sets and test sets with consistent data sources, the overall classification accuracy of the method is up to 99.84% ± 0.04%. For training sets and test sets with inconsistent data sources, a small number of training sets are extended by the active learning (AL) method, and the overall classification accuracy of the method is up to 99.31% ± 0.23%. The experimental results show the effectiveness of the method in arrhythmia automatic feature extraction and classification. It provides a new solution for the automatic extraction of ECG signal features and classification for deep learning.

Diagnosis of Breast Hyperplasia and Evaluation of RuXian-I Based on Metabolomics Deep Belief Networks.

Breast cancer is estimated to be the leading cancer type among new cases in American women. Core biopsy data have shown a close association between breast hyperplasia and breast cancer. The early diagnosis and treatment of breast hyperplasia are extremely important to prevent breast cancer. The Mongolian medicine RuXian-I is a traditional drug that has achieved a high level of efficacy and a low incidence of side effects in its clinical use. However, for detecting the efficacy of RuXian-I, a rapid and accurate evaluation method based on metabolomic data is still lacking. Therefore, we proposed a framework, named the metabolomics deep belief network (MDBN), to analyze breast hyperplasia metabolomic data. We obtained 168 samples of metabolomic data from an animal model experiment of RuXian-I, which were averaged from control groups, treatment groups, and model groups. In the process of training, unlabelled data were used to pretrain the Deep Belief Networks models, and then labelled data were used to complete fine-tuning based on a limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) algorithm. To prevent overfitting, a dropout method was added to the pretraining and fine-tuning procedures. The experimental results showed that the proposed model is superior to other classical classification methods that are based on positive and negative spectra data. Further, the proposed model can be used as an extension of the classification method for metabolomic data. For the high accuracy of classification of the three groups, the model indicates obvious differences and boundaries between the three groups. It can be inferred that the animal model of RuXian-I is well established, which can lay a foundation for subsequent related experiments. This also shows that metabolomic data can be used as a means to verify the effectiveness of RuXian-I in the treatment of breast hyperplasia.

Automatic classification method of arrhythmia based on discriminative deep belief networks / 生物医学工程学杂志

Existing arrhythmia classification methods usually use manual selection of electrocardiogram (ECG) signal features, so that the feature selection is subjective, and the feature extraction is complex, leaving the classification accuracy usually affected. Based on this situation, a new method of arrhythmia automatic classification based on discriminative deep belief networks (DDBNs) is proposed. The morphological features of heart beat signals are automatically extracted from the constructed generative restricted Boltzmann machine (GRBM), then the discriminative restricted Boltzmann machine (DRBM) with feature learning and classification ability is introduced, and arrhythmia classification is performed according to the extracted morphological features and RR interval features. In order to further improve the classification performance of DDBNs, DDBNs are converted to deep neural network (DNN) using the Softmax regression layer for supervised classification in this paper, and the network is fine-tuned by backpropagation. Finally, the Massachusetts Institute of Technology and Beth Israel Hospital Arrhythmia Database (MIT-BIH AR) is used for experimental verification. For training sets and test sets with consistent data sources, the overall classification accuracy of the method is up to 99.84% ± 0.04%. For training sets and test sets with inconsistent data sources, a small number of training sets are extended by the active learning (AL) method, and the overall classification accuracy of the method is up to 99.31% ± 0.23%. The experimental results show the effectiveness of the method in arrhythmia automatic feature extraction and classification. It provides a new solution for the automatic extraction of ECG signal features and classification for deep learning.