Deep Neural Networks for the Recognition and Classification of Heart Murmurs Using Neuromorphic Auditory Sensors

A NOVEL APPROACH OSA DETECTION USING SINGLE-LEAD ECG SCALOGRAM BASED ON DEEP NEURAL NETWORK

Obstructive sleep apnea (OSA) is the most common and severe breathing dysfunction which frequently freezes the breathing for longer than 10[Formula: see text]s while sleeping. Polysomnography (PSG) is the conventional approach concerning the treatment of OSA detection. But, this approach is a costly and cumbersome process. To overcome the above complication, a satisfactory and novel technique for interpretation of sleep apnea using ECG were recording is under development. The methods for OSA analysis based on ECG were analyzed for numerous years. Early work concentrated on extracting features, which depend entirely on the experience of human specialists. A novel approach for the prediction of sleep apnea disorder based on the convolutional neural network (CNN) using a pre-trained (AlexNet) model is analyzed in this study. After filtering per-minute segment of the single-lead ECG recording accompanied by continuous wavelet transform (CWT), the 2D scalogram images are generated. Finally, CNN based on deep learning algorithm is adopted to enhance the classification performance. The efficiency of the proposed model is compared with the previous methods that used the same datasets. Proposed method based on CNN is able to achieve the accuracy of 86.22% with 90% sensitivity in per-minute segment OSA classification. Based on per-recording OSA diagnosis, our works correctly classify all the abnormal apneic recording with 100% accuracy. Our OSA analysis model using time-frequency scalogram generates excellent independent validation performance with different state-of-the-art OSA classification systems. Experimental results proved that the proposed method produces excellent performance outcomes with low cost and less complexity.

A Wearable, Patient-Adaptive Freezing of Gait Detection System for Biofeedback Cueing in Parkinson's Disease

Freezing of Gait (FoG) is a common motor-related impairment among Parkinson's disease patients, which substantially reduces their quality of life and puts them at risk of falls. These patients benefit from wearable FoG detection systems that provide timely biofeedback cues and hence help them regain control over their gait. Unfortunately, the systems proposed thus far are bulky and obtrusive when worn. The objective of this paper is to demonstrate the first integration of an FoG detection system into a single sensor node. To achieve such an integration, features with low computational load are selected and dedicated hardware is designed that limits area and memory utilization. Classification is achieved with a neural network that is capable of learning in real time and thus allows the system to adapt to a patient during run-time. A small form factor FPGA implements the feature extraction and classification, whereas a custom PCB integrates the system into a single node. The system fits into a 4.5 × 3.5 × 1.5 cm ³ housing case, weighs 32 g, and achieves 95.6% sensitivity and 90.2% specificity when adapted to a patient. Biofeedback cues are provided either through auditory or somatosensory means and the system can remain operational for longer than 9 h while providing cues. The proposed system is highly competitive in terms of classification performance and excels with respect to wearability and real-time patient adaptivity.

Neuromorphic LIF Row-by-Row Multiconvolution Processor for FPGA

Deep Learning algorithms have become state-of-the-art methods for multiple fields, including computer vision, speech recognition, natural language processing, and audio recognition, among others. In image vision, convolutional neural networks (CNN) stand out. This kind of network is expensive in terms of computational resources due to the large number of operations required to process a frame. In recent years, several frame-based chip solutions to deploy CNN for real time have been developed. Despite the good results in power and accuracy given by these solutions, the number of operations is still high, due the complexity of the current network models. However, it is possible to reduce the number of operations using different computer vision techniques other than frame-based, e.g., neuromorphic event-based techniques. There exist several neuromorphic vision sensors whose pixels detect changes in luminosity. Inspired in the leaky integrate-and-fire (LIF) neuron, we propose in this manuscript an event-based field-programmable gate array (FPGA) multiconvolution system. Its main novelty is the combination of a memory arbiter for efficient memory access to allow row-by-row kernel processing. This system is able to convolve 64 filters across multiple kernel sizes, from 1 × 1 to 7 × 7, with latencies of 1.3  μs and 9.01  μs, respectively, generating a continuous flow of output events. The proposed architecture will easily fit spike-based CNNs.

An Accelerator Architecture of Changeable-Dimension Matrix Computing Method for SVM

Matrix multiplication is a critical time-consuming processing step in many machine learning applications. Due to the diversity of practical applications, the matrix dimensions are generally not fixed. However, most matrix calculation methods, based on field programmable gate array (FPGA) currently use fixed matrix dimensions, which limit the flexibility of machine learning algorithms in a FPGA. The bottleneck lies in the limited FPGA resources. Therefore, this paper proposes an accelerator architecture for matrix computing method with changeable dimensions. Multi-matrix synchronous calculation concept allows matrix data to be processed continuously, which improves the parallel computing characteristics of FPGA and optimizes the computational efficiency. This paper tests matrix multiplication using support vector machine (SVM) algorithm to verify the performance of proposed architecture on the ZYNQ platform. The experimental results show that, compared to the software processing method, the proposed architecture increases the performance by 21.18 times with 9947 dimensions. The dimension is changeable with a maximum value of 2,097,151, without changing hardware design. This method is also applicable to matrix multiplication processing with other machine learning algorithms.

Deep Learning in Cardiology

The medical field is creating large amount of data that physicians are unable to decipher and use efficiently. Moreover, rule-based expert systems are inefficient in solving complicated medical tasks or for creating insights using big data. Deep learning has emerged as a more accurate and effective technology in a wide range of medical problems such as diagnosis, prediction, and intervention. Deep learning is a representation learning method that consists of layers that transform data nonlinearly, thus, revealing hierarchical relationships and structures. In this review, we survey deep learning application papers that use structured data, and signal and imaging modalities from cardiology. We discuss the advantages and limitations of applying deep learning in cardiology that also apply in medicine in general, while proposing certain directions as the most viable for clinical use.

Supervised threshold-based heart sound classification algorithm

OBJECTIVE

Deep classification networks have been one of the predominant methods for classifying heart sound recordings. To satisfy their demand for sample size, the most commonly used method for data augmentation is that which divides each heart sound instance into a number of segments, with each segment labelled as the same category as its origin and used as a new sample for training or forecasting. However, performing this poses a crucial issue as to how to determine the category of a predicted heart sound instance from its segments' prediction results.

APPROACH

To solve this issue, this paper establishes a mathematical formula to connect the classification performance of these heart sound instances with the prediction results of their segments via a threshold which is supervised by the training set. The optimal value of the proposed threshold is calculated by maximizing the prediction accuracy of the training instances. Seeking the optimal threshold by a gradient-based method, we prove that a continuous function can closely approximate a part of the function of accuracy which transforms the discrete function of accuracy into a continuous function. The optimal threshold is used to recognize the undetermined heart sound recording.

MAIN RESULTS

Experimental results show the classification performance from a 10-fold cross-validation, measured by the commonly used scales of sensitivity, specificity and mean accuracy (MAcc). The proposed algorithm improves the MAcc by about 4% by modifying the baseline. In addition, the MAcc surpasses the champion of the PhysioNet/Computing in Cardiology Challenge 2016.

SIGNIFICANCE

Our study develops a methodology to determine the category of a predicted heart sound instance from its segments' prediction results, thus assisting in the data augmentation exercise which is necessary to provide sufficient data for deep classification networks. Our method significantly improves the classification performance.

Deep Learning on 1-D Biosignals: a Taxonomy-based Survey

OBJECTIVES

 Deep learning models such as convolutional neural networks (CNNs) have been applied successfully to medical imaging, but biomedical signal analysis has yet to fully benefit from this novel approach. Our survey aims at (i) reviewing deep learning techniques for biosignal analysis in computer- aided diagnosis; and (ii) deriving a taxonomy for organizing the growing number of applications in the field.

METHODS

 A comprehensive literature research was performed using PubMed, Scopus, and ACM. Deep learning models were classified with respect to the (i) origin, (ii) dimension, and (iii) type of the biosignal as input to the deep learning model; (iv) the goal of the application; (v) the size and (vi) type of ground truth data; (vii) the type and (viii) schedule of learning the network; and (ix) the topology of the model.

RESULTS

 Between January 2010 and December 2017, a total 71 papers were published on the topic. The majority (n = 36) of papers are on electrocariography (ECG) signals. Most applications (n = 25) aim at detection of patterns, while only a few (n = 6) at predection of events. Out of 36 ECG-based works, many (n = 17) relate to multi-lead ECG. Other biosignals that have been identified in the survey are electromyography, phonocardiography, photoplethysmography, electrooculography, continuous glucose monitoring, acoustic respiratory signal, blood pressure, and electrodermal activity signal, while ballistocardiography or seismocardiography have yet to be analyzed using deep learning techniques. In supervised and unsupervised applications, CNNs and restricted Boltzmann machines are the most and least frequently used, (n = 34) and (n = 15), respectively.

CONCLUSION

 Our key-code classification of relevant papers was used to cluster the approaches that have been published to date and demonstrated a large variability of research with respect to data, application, and network topology. Future research is expected to focus on the standardization of deep learning architectures and on the optimization of the network parameters to increase performance and robustness. Furthermore, application-driven approaches and updated training data from mobile recordings are needed.

Real-Time Event-Driven Classification Technique for Early Detection and Prevention of Myocardial Infarction on Wearable Systems

A considerable portion of government health-care spending is allocated to the continuous monitoring of patients suffering from cardiovascular diseases, particularly myocardial infarction (MI). Wearable devices present a cost-effective means of monitoring patients' vital signs in ambulatory settings. A major challenge is to design such ultra-low energy devices for long-term patient monitoring. In this paper, we present a real-time event-driven classification technique based on the random forest classification scheme, which uses a confidence-related decision-making process. The main goal of this technique is to maintain a high classification accuracy while reducing the complexity of the classification algorithm. We validate our approach on a well-established and complete MI database (Physiobank, PTB Diagnostic ECG database). Our experimental evaluation demonstrates that our real-time classification scheme outperforms the existing approaches in terms of energy consumption and battery lifetime by a factor of 2.60, with no classification quality loss.