key: cord-0560071-pb9lh3sp authors: Liu, Shuo; Han, Jing; Puyal, Estela Laporta; Kontaxis, Spyridon; Sun, Shaoxiong; Locatelli, Patrick; Dineley, Judith; Pokorny, Florian B.; Costa, Gloria Dalla; Leocan, Letizia; Guerrero, Ana Isabel; Nos, Carlos; Zabalza, Ana; Sorensen, Per Soelberg; Buron, Mathias; Magyari, Melinda; Ranjan, Yatharth; Rashid, Zulqarnain; Conde, Pauline; Stewart, Callum; Folarin, Amos A; Dobson, Richard JB; Bail'on, Raquel; Vairavan, Srinivasan; Cummins, Nicholas; Narayan, Vaibhav A; Hotopf, Matthew; Comi, Giancarlo; Schuller, Bjorn title: Fitbeat: COVID-19 Estimation based on Wristband Heart Rate date: 2021-04-19 journal: nan DOI: nan sha: ae313ad702132adecda02577bbb7f4782d2fcf34 doc_id: 560071 cord_uid: pb9lh3sp This study investigates the potential of deep learning methods to identify individuals with suspected COVID-19 infection using remotely collected heart-rate data. The study utilises data from the ongoing EU IMI RADAR-CNS research project that is investigating the feasibility of wearable devices and smart phones to monitor individuals with multiple sclerosis (MS), depression or epilepsy. Aspart of the project protocol, heart-rate data was collected from participants using a Fitbit wristband. The presence of COVID-19 in the cohort in this work was either confirmed through a positive swab test, or inferred through the self-reporting of a combination of symptoms including fever, respiratory symptoms, loss of smell or taste, tiredness and gastrointestinal symptoms. Experimental results indicate that our proposed contrastive convolutional auto-encoder (contrastive CAE), i. e., a combined architecture of an auto-encoder and contrastive loss, outperforms a conventional convolutional neural network (CNN), as well as a convolutional auto-encoder (CAE) without using contrastive loss. Our final contrastive CAE achieves 95.3% unweighted average recall, 86.4% precision, anF1 measure of 88.2%, a sensitivity of 100% and a specificity of 90.6% on a testset of 19 participants with MS who reported symptoms of COVID-19. Each of these participants was paired with a participant with MS with no COVID-19 symptoms. Remote, passive monitoring of physiological and behavioural characteristics using smartphones and wearable devices can be used to rapidly collect a variety of data in huge volumes. Such data has the potential to improve our understanding of the interplay between a variety of health conditions at both the level of the individual and the population, if rigorously collected and validated [1] . Passive data collection is typically implemented with a high temporal reso-lution [2] . Wearable fitness trackers, for example, estimate parameters such as heart rate up to every second, and can be worn 24 hours a day. Monitoring individuals with a range of health states, lifestyles, and demographic variables in combination with data artefacts and missing data leads to high variability. From heart rate and general physical activity to GPS-based location information or the proximity to other individuals detected through Bluetooth, multiple types of data can be monitored remotely. Therefore, studies using wearables and smartphones in this way exhibit several 'v's of big data, namely velocity, volume, variability and variety. As such, advanced analysis methodology such as deep learning have the potential to make a significant contribution [3] . The volume and richness of remotely collected data has particular potential in the context of infectious diseases, such as the novel corona virus (SARS-CoV-2). Specific applications include individual screening and population-level monitoring while minimising contact with infected individuals [4, 5, 6] . Among all remotely collectable physiological data, heart rate is a biomarker of particular value in such applications. Patterns in heart rate fluctuations over time in particular have been found to provide clinically relevant information about the integrity of the physiological system generating these dynamics. Previous studies have not only revealed an altered heart rate variability in a number of medical conditions, such as hypovolaemia [7] , heart failure [8] , or angina [9] , but also demonstrated that the degree of short-term heart rate alteration correlates with illness severity. Analysis of the autonomic regulation of heart rate, such as heart rate variability, has also been discussed as a promising approach for detecting infections earlier than conventional clinical methods and prognosticating infections [10] . Wearables such as Fitbit fitness trackers 1 can provide indirect measurements of the heart rate through pulse rate estimates made using photoplethysmography. Radin and colleagues [4] analysed the resting heart rate collected in this way, alongside with sleep duration data collected from over 47 000 individuals to improve model predictions of influenza rates in five US states. In the ongoing DETECT 2 study [5] , the same group is now applying this approach to monitor outbreaks of viral infections including COVID-19. In separate work, Fitbit devices have also been used to identify individuals with COVID-19 using heart rate measurements collected before symptom onset [11] . In addition to pre-print studies [12, 13] , other similar ongoing endeavours include the German project Corona-Datenspende 3 , which has a cohort of over 500 000 volunteers, and the TemPredict study in the US 4 . Applied to such data sets, deep learning methodology has the potential to automatically identify individuals with COVID-19 purely on the basis of passive data from wearable devices [5, 13, 11] . Natarajan and colleagues [13] used a convolutional neural network (CNN) to predict illness on a given day using Fitbit data from 1 181 individuals, reporting an area under the receiver operating characteristics curve (AUC) of 0.77 ± 0.03. Similarly, this paper describes a CNN-based deep learning approach for predicting the presence of COVID-like symptoms using Fitbit heart rate data. Treating this task as analogous to anomaly detection, we explore the suitability of using a convolutional autoencoder (CAE) with contrastive loss [14, 15] . The technique's performance is compared against a conventional auto-encoder and a convolutional neural network. The data used in this work was collected as part of the IMI2 RADAR-CNS major programme 5 . RADAR-CNS is an observational, longitudinal, prospective study aimed at developing new ways of measuring major depressive disorder, epilepsy, and multiple sclerosis (MS) using remote monitoring technologies Heart rate data of the participants were assigned into temporal segments, defining a 14-day interval extending from 7 days preceding symptom onset to 7 days following symptom onset in which we sought to identify infection-related variations in heart rate. The interval mainly cover the duration of the COVID-19 incubation period [19, 20, 21] , and minimises the anomalous effects of day-to-day variations in activity, such as those observed between weekdays and weekends. Since the number of available symptomatic and asymptomatic segments was highly imbalanced, we replicated the symptomatic segments to the number of asymptomatic segments to guide the detection model weighted in favour of the minority class. For the LOSO CV procedure, 19 symptomatic and 570 asymptomatic segments were acquired from participants with reported symptoms, and 1 140 asymptomatic segments from the control participants also referred to as control segments. An overview of available symptomatic and asymptomatic segments is given in Table 2 . Ideally, the instantaneous heart rate measurement for an individual was captured every second, and was uploaded every five seconds along with the recording time (blue curve in middle of Fig. 1 ). The mean of 5 minutes is taken to smooth heart rate measurements while still tracking slow short-term changes in the heart rate. Moreover, this approach alleviates the effect of different sampling rates observed in actual heart rate recordings. Missing data over full 5-minutes intervals is filled with the median value of the overall 14-days segment. Finally, we have a single heart rate value per every 5 minutes. The resulting smoothed heart rate trajectory is considered appropriate to detect global heart rate patterns associated with COVID-like symptoms (red curve in middle of Fig. 1 ). The red rectangle covers 7 days heart rate data before and after symptom onset and represents a symptomatic segment, while the green rectangle gives an example of an asymptomatic segment, i. e., a segment in which no symptoms were present and which is located at least 7 days distant from the symptomatic segment. Middle: Symptomatic segment. The blue curve represents the unprocessed heart rate trajectory of the red rectangle above, while the red curve indicates the heart rate trajectory averaged over 5-minutes intervals. Bottom: Representation of the above illustrated symptomatic segment as 24 × 168 sized image of 5-minutes heart rate data related pixels. Each column represents an interval of 2 hours, the 168 columns sum up to 14 days. We then transform the averaged heart rate data of each segment into a feature map, i. e., an image of size 24 × 168 pixels (bottom of Fig. 1 ), in which each pixel represents a 5-minutes heart rate sampling point. Thereby, each column encodes a heart rate trajectory of 2 hours (24 × 5 minutes), resulting in a covered interval of overall 14 days by 168 columns (168 × 2 hours). In our experiments, we verified that this set-up of the feature map is effective as the input of our deep learning models, leading to promising detection results. For detecting the presence of COVID-like symptoms in a given 14 days heart rate segment, we investigate several different deep learning methods to represent the problem by feature maps, including CNNs [22, 23] and a conventional CAE [24, 25, 26, 27] . Further, we formulise the task as anomaly detection, and propose a strategy for training a CAE, by means of fitting the reconstruction error into the format of contrastive loss instead of conventional loss like root mean square error (RMSE). A CNN typically consists of a sequence of convolutional layers, with every convolutional layer containing its own learnable parameters, i. e., the filter weights and biases, to extract features from its input feature map. The stacking of convolutional layers allows a hierarchical decomposition of the input, a deeper CNN is expected to develop more complex representations of the the feature maps [28, 29, 30] . Subsequently, fully-connected layers map the learnt representations to the desired classes. Such CNN architectures have been successfully applied in many tasks, such as image and video recognition [31, 32] , sequential data processing [33, 34] , and medical applications [35, 36, 37, 38] including those recent works for Covid-19 diagnosis [39, 40, 41, 42] . The alternative approach to learn representations from an input image is to use a CAE [43, 26, 27 ]. An auto-encoder contains two parts, i. e., an encoder and a decoder. The encoder learns latent attributes of the original input, whereas the decoder aims at reconstructing the original input from the learnt latent attribute. The dimensionality of the latent attributes is designed as a bottleneck imposed in the architecture; it hence can be seen as a compressed knowledge representation of the input. In order to guarantee the reproduction of the original input at the output of the decoder, the reconstruction error is minimised when optimising an auto-encoder network. In this work, we keep the architecture of our CAE encoder as consistent with our CNNs, and the decoder is arranged in a mirror-like fashion. Since the conventional CAE is an unsupervised learning approach, an additional multi-layer perception (MLP) [44] follows to classify the learnt attributes. This strategy has been widely used for medical and health related tasks [45, 46, 47] , including tasks similar to ours, such as anomaly detection [48, 49, 50, 51] . As the strategy for optimising a conventional CAE does not take class information into account, the learnt latent attributes may contain redundant representations that are of no sufficient importance for the final classification. To incorporate the class information during training of the CAE, we apply contrastive loss [52] to the CAE reconstruction error in order to guide it to learn sufficiently discriminative latent attributes for different classes. The three considered neural network architectures, i. e., CNN, conventional CAE, contrastive CAE, used in this work are described in details in the following separate sections. Our CNN architecture is comprised of a sequence of convolutional layers, an example of 4 layers is illustrated in Fig. 2 . Each convolutional layer contains its own kernel size, stride, and number of channels as given in the specifications that appear in Table 3 . For one layer convolution, kernel size and stride together determine the receptive field, i. e., the perception scope on the original input. Zero-padding is typically utilised to maintain a proper size of the feature map. More kernel filters result in better representation diversity. In between convoutional layers, batch normalisation can be used reduce the effect of internal covariate shift (ICS), an effect caused by different batches of training data having slightly different distributions [53, 54] . Hence, the vanishing gradient issue can be alleviated. A parametric rectified linear unit (PReLU) [55] performs as the activation function for each layer. It carries on the merits of ReLU activation [56] in stable training while avoiding dead neurons using a learnable slope parameter. A max-pooling layer processes the activations to reduce the spatial size of the feature size, leading to less parameters in the following network layers and, therefore, alleviating the potential over-fitting issues. The CNN is therefore a sequential cascade of convolutional layer -batch normalisation -PReLU. Its output is averaged across all locations, resulting in a representation vector. The representation vector is then projected onto the two classes, symptomatic or asymptomatic, through two fully-connected layers. In our experiments, we tested CNNs with different number of layers, ranging from 1 to 6, in order to optimise the performance. The encoder part of our CAE keeps the identical structure as the CNN introduced in Section 3.1, but contains only one FC layer before mapping into latent attributes. Given the feature map of a heart rate segment x i , in which i indicates segment index, the encoder f enc (·) processes the feature map and produces the latent attributes The decoder presents an inverse processing of the encoder as seen in 3. For each decoder layer, the input feature map mainly passes through transposed convolution and transposed max-pooling, also known as de-convolution and depooling. Batch normalisation is employed in between, followed by PReLU as the activation function. The decoding procedure can be represented aŝ where f dec (·) represent the operations in the decoder. Its outputx i , seen as the reconstructed feature map, is then compared to the original input to compute the reconstruction error. An auto-encoder is optimised by minimising the reconstruction error, one typical such error is root mean squared error (RMSE) [57] : where N stands for the number of samples. The specifications of our CAE are given in Table 4 . In experiments, we considered different number of convolutional layers in the CAE; the dimensionality of the flatten layer is determined by the last encoder layer. Hence, for different numbers of layers, we adjust the length of the fully-connected layer to optimise the CAE performance. The difficulty in finding good latent attributes in the framework of an autoencoder lies in setting it to a proper dimensionality. Too long latent attributes may contain redundancies for easier reconstructing the original input, but fall short of concentrating on learning the saliently discriminative features for different classes. Meanwhile, shorter latent attributes can have less or limited representation capability. The conventional optimisation strategy for an autoencoder considers no class information [26, 27] , and hence the learnt latent attributes is not well oriented to be discriminative for different classes. Specif-ically, for our detection task, the auto-encoder may tend to learn the latent attributes that can better reconstruct the original feature map, while ignoring some salient attributes that indicate the difference between symptomatic and asymptomatic segments. To incorporate class information, symptomatic and asymptomatic, into the optimisation of CAE, we fit the reconstruction error of the two classes into contrastive loss [52] . As analogues to anomaly detection, we expect the CAE to output a low reconstruction error for asymptomatic segments, and a high reconstruction error for symptomatic segments. Therefore, the loss function for our contrastive CAE can be seen as where the superscripts p and n are used to distinguish positive (symptomatic) and negative (asymptomatic) samples. Ideally, the reconstruction error for a positive pair, original and reconstructed feature map for asymptomatic segments, is expected to be 0, indicating a successful reconstruction of the original input at the decoder. In contrast, the reconstruction error for the negative input pair, original and reconstructed feature map for symptomatic segments, is expected to be close to the margin value m. Therefore, the difference in classes leads to different reconstruction errors from our CAE. Contrastive loss has been successfully exploited in several image processing tasks, especially those recently proposed for CT and X-ray diagnosis of COVID-19 [58, 59, 60] . We conducted a series of experiments in order to test the models presented in Section III for detecting COVID-like symptoms in the given 14-days heart rate segments. The neural networks of different architectures are pre-trained with the heart rate segments of 49 participants that reported COVID-like symptoms, but had not met the CD1 or CD2 criterion. We apply leave-one-subject-out cross-validation to the heart rate segments of the 19 individuals who meet CD1 or CD2, and further test the models on the corresponding symptom-free control group. The performance of our most effective method, i. e., the Contrastive CAE to assure stable and fast convergence of our models. In the following, we first compare our proposed contrastive CAE with a CNN model, both of which use a different number of layers. As an additional comparison method, we applied an MLP directly to the one-dimensional 5-min average heart rate segment without formatting it into feature maps. and 75.4% specificity as given in Table 6 . This indicates that the CAE trained with RMSE has a limited capability in learning discriminative lattent attributes between symptomatic and asymptomatic segments, since the attributes learning procedure considers no class information during optimisation, and, hence leaves the classification difficulty to the final MLP classifiers. Besides, the optimised performance of the conventional CAE is worse than that of the CNN models, further stressing the need of involving the class information in training a more efficient CAE for the detection task. For the bianry classification task, the classes' difference can be implicitly modelled in the contrastive loss as in Eq. (4) for training the CAE, since the positive and negative reconstruction error are guided to produce a margin between each other in a discriminative manner. Hence, the contrastive CAE is capable of learning latent attributes that represent salient features to distinguish between symptomatic and asymptomatic segments. Again, a 2-layer MLP is used to classify the learnt latent attributes into classes, symptomatic or asymptomaic. In our experiments, the contrastive CAE with an attribute dimensionality of 100 achieves its best result in terms of UAR, and when the dimensionality increases to 300, the proposed contrastive CAE achieves a balanced results between the UAR and precision measure, leading to its best F1 measure, i. e., 84.6%. In addition, the model obtained the optimum sensitivity and a high specificity in Since our contrastive CAE was trained to create a sufficient margin between the reconstruction errors of symptomatic and asymptomatic segments, applying classification directly on the reconstruction errors, rather than the learnt latent attributes, is a more efficient way to use the contrastive CAE for our binary decision task. The decision threshold between the reconstruction errors of symptomatic and asymptomatic classes is determined by the means of applying logistic regression [61] on the training part for each cross-validation round. A 14-days heart rate segment is decided for as COVID-19 symptomatic (CD1/CD2 criterion) if the reconstruction error is above the decision boundary. The best performance, shown in Table 7 and decoder by removing the latent attributes layer. The performance, however, maintains stable as given in the last row of Table 7 . We further explore the possibility to make binary COVID-19 yes/no (based on the symptom CD1/CD2 definitions above) decisions earlier with our proposed contrastive CAE. To this end, we shift the window for sliding over the symptomatic segments to earlier and later days, but still contain the onset date. On the other hand, in order to test the model's validality on later days, we shift the sliding window for asymptomatic segments to later days that contain the onset date. In our experiment, we keep the asymptomatic segments as in the previous experiments, but replace the previous symptomatic segments with the shifted symptomatic segments. The experimental results, as seen in Table 8 , reveal that the model works well for the heart rate segments that are sliced one day earlier or three days later, leading to no performance degradation. This is potentially due to the fact that the model was trained with segments centred by onset date, and the shifted symptomatic segments share similar internal structure as the original symptomatic segments. However, segments that further deviate from the original symptomatic segments, i. e., shifting the sliding window to two more previous days or four days later, results in decreased classification accuracy. Throughout all experiments, we kept assuming that the participant-reported onset date is identical to the real symptom onset. As seen in Section 4.3, the contrastive CAE performs effective on the symptomatic segments that are centred relating to the reported symptom onset. In addition, we observe the CAE to work robustly on the segments one day earlier and three days later, and further shifting of the symptomatic segments leads to decrease in performance. However, considering that the participants may have hesitated before reporting the appearance of their symptoms, the true symptom onset may happen earlier than the reported onset date. As the experimental results given in Table 8 , we see that the proposed model starts giving good results one day earlier, although the model was optimised only with symptomatic segments and their control segments, indicating that potentially some participants were conscious of the symptoms and reported the symptoms on the second day. Furthermore, shifting the segments up to a few days later (maximally three days in our experiments) leads to higher certainty that symptoms are indeed contained, and hence, our model achieves stable performance in this case. Potentially, there exist some participants whose true symptom started on earlier days, but relieved soon. In this case, the segments, that are shifted to a later point in time by many days, may exclude the true symptomatic part, leading to a low detection accuracy when assuming that the considered segments contain symptoms. In this work, we base our study on 5-minutes mean of the heart rate measurements. In practice, we also considered the heart rate variance every 5 minutes as input features when starting with standard CNNs, and did not see additional improvement in terms of accuracy. Therefore, in further studies of conventional CAEs and contrastive CAEs, we concentrated on the mean value rather than variance. The mean heart rate values as features of our deep learning models may be sufficient, since the global variance is retained in segmental mean values. In some previous works, the class information can be applied to latent attribute layers, leading to the supervised auto-encoder introduced in [62] . Crossentropy losses are used to minimise the difference between predicted labels from latent attributes and true labels. This approach provides a certain preservation of the reconstructed feature map, taking the cross-entropy loss as a regularisation method. The reconstruction error and cross entropy loss are jointly optimised. However, the optimisation of the joint loss requires a proper combination factor in order to balance the optimisation on reconstruction error and prediction error. Since the two types of errors originate from different stages of the auto-encoder model, leading to their different scale level, the difficulty lies in seeking a good combination scale. In our proposed method, the contrastive loss can be employed directly on the reconstruction error for positive and negative input pairs. The training procedure is simpler, since the reconstruction errors of the positive pairs and negative pairs present in the same scale level at the beginning of the training stage. The method also provides the chance to validate whether the model has learnt discriminative latent attributes for different classes in the auto-encoder framework. In this work, we explored several deep learning models, including conventional CNNs, conventional CAEs, and contrastive CAEs, to make a machinelearning-based COVID-19 yes/no decision based on symptoms' presence as definded by two criteria (CD1/CD2) given 14-days heart rate measurements from a Fitbit wristband. The models were pre-trained based on the heart rate data of 49 participants with MS who reported to have COVID-like symptoms but did not satisfy CD1 or CD2. However, the reported symptoms were highly related to COVID-19 symptoms. The models were then tested on data of 19 MS participants whose reported symptoms met the criteria of CD1 or CD2, by means of LOSO CV. In this process, each of the 19 symptomatic MS participants was paired with a site-, gender-, and age-matched symptom-free MS participants. Experimental results indicate that our proposed contrastive CAE approach, incorporating class information into optimising the CAE with contrastive loss, achieved considerable improvements over the conventional CNN and CAE models in terms of performance, including UAR, precision, F1 measure, as well as sensitivity, and specificity. The effectiveness of contrastive learning as demonstrated in this work shall motivate further research, including the investigation of different bio-signal measurements, features, and model architectures. Several recent works [5, 11] have shown the potential of using more than one bio-signal measurement for Covid-19 recognition. Such additional measurements can be audio and speech signal [63, 64] which can also be achieved from wearable devices. Since the set-up of our experiments was chosen to detect whether or not the COVID-19-like symptoms appeared during a period of heart rate data, the models show limitations in a causal set-up, i. e., when trying to predict potential symptoms before they are present. To this end, future work shall try to answer the question of how many days in advance we will already be able to reliably predict the potential imminent onset of COVID-like symptoms. As the acquisition of data in the RADAR-CNS programme is still ongoing, the improvement of our proposed binary COVID-19 yes/no (based on the symptom CD1/CD2 definitions above) classification model based on a broader data foundation is expected. Further to that, other windows of time should be analysed. Overall, we are optimistic that an applicable decision can be made as to COVID-19 presence based on the symptoms defined herein based on machine learning analysis of consumer-type heart rate measurement. This project has received funding from the Innovative Medicines Initiative The rise of consumer health wearables: Promises and barriers RADAR-base: Open source mobile health platform for collecting, monitoring, and analyzing data using sensors, wearables, and mobile devices Deep Learning for IoT big data and streaming analytics: A survey Harnessing wearable device data to improve state-level real-time surveillance of influenza-like illness in the USA: A population-based study Wearable sensor data and selfreported symptoms for COVID-19 detection Learning from large-scale wearable device data for predicting epidemics trend of COVID-19 Mild hypovolemic stress alters autonomic modulation of heart rate Independent and incremental prognostic value of heart rate variability in patients with chronic heart failure Heart rate variability depression in patients with unstable angina Clinical review: A review and analysis of heart rate variability and the diagnosis and prognosis of infection Pre-symptomatic detection of COVID-19 from smartwatch data Measuring COVID-19 and influenza in the real world via person-generated health data Assessment of physiological signs associated with COVID-19 measured using wearable devices Clustering and unsupervised anomaly detection with l2 normalized deep auto-encoder representations Autoencoder-based network anomaly detection The RADAR-CNS consortium, Real-time assessment of COVID-19 prevalence among multiple sclerosis patients: A multicenter European study Clinical features of patients infected with 2019 novel coronavirus in Performance evaluation of classification algorithms by k-fold and leave-one-out cross validation Incubation period of 2019 novel coronavirus (2019-nCoV) infections among travellers from Wuhan The incubation period of coronavirus disease 2019 (COVID-19) from publicly reported confirmed cases: Estimation and application Use of physiological data from a wearable device to identify SARS-CoV-2 infection and symptoms and predict COVID-19 diagnosis: Observational study Imagenet classification with deep convolutional neural networks Fast and scalable distributed deep convolutional autoencoder for fMRI big data analytics Deep convolutional autoencoderbased lossy image compression Image classification based on convolutional denoising sparse autoencoder Stacked convolutional denoising auto-encoders for feature representation Deep inside convolutional networks: Visualising image classification models and saliency maps Visualizing and understanding convolutional networks Conceptual understanding of convolutional neural network -a deep learning approach Object detection with deep learning: A review Large-scale video classification with convolutional neural networks Deep learning for time series classification: A review, Data Mining and Knowl Sound classification using convolutional neural network and tensor deep stacking network Deep learning with convolutional neural network for differentiation of liver masses at dynamic contrastenhanced CT: A preliminary study Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs van der Laak, the CAME-LYON16 Consortium, Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer Proceedings of International Conference on Medical Image Computing & Computer Assisted Intervention Deep learning for image-based cancer detection and diagnosis: A survey Review of artificial intelligence techniques in imaging data acquisition, segmentation and diagnosis for COVID-19 Automatic COVID-19 lung infected region segmentation and measurement using CT-scans images Predicting COVID-19 Malignant Progression with AI Techniques A noisy sparse convolution neural network based on stacked auto-encoders Artificial neural networks (the multilayer perceptron) -A review of applications in the atmospheric sciences Deep features learning for medical image analysis with convolutional autoencoder neural network Stacked deep convolutional auto-encoders for emotion recognition from facial expressions Prediction of drug-likeness using deep autoencoder neural networks Anomaly detection with robust deep autoencoders Deep autoencoding Gaussian mixture model for unsupervised anomaly detection Spatio-temporal autoencoder for video anomaly detection Clustering and unsupervised anomaly detection with l2 normalized deep auto-encoder representations Proceedings of Conference on Neural Information Processing Systems Batch normalization: Accelerating deep network training by reducing internal covariate shift Understanding batch normalization Delving deep into rectifiers: Surpassing human-Level performance on ImageNet classification Understanding and improving convolutional neural networks via concatenated rectified linear units MetaCOVID: A Siamese neural network framework with contrastive loss for n-shot diagnosis of COVID-19 patients Multi-task contrastive learning for automatic CT and X-ray diagnosis of COVID-19 Momentum contrastive learning for few-shot COVID-19 diagnosis from chest CT images An introduction to logistic regression: From basic concepts to interpretation with particular attention to nursing domain Supervised autoencoders: Improving generalization performance with unsupervised regularizers COVID-19 and computer audition: An overview on what speech & sound analysis could contribute in the SARS-CoV-2 Corona crisis An early study on intelligent analysis of speech under COVID-19: Severity, sleep quality, fatigue, and anxiety