Ensemble of Convolution Neural Networks on Heterogeneous Signals for Sleep Stage Scoring

• •

  Processing heterogeneous signals simultaneously in a Separable CNN and explore the advantages of this approach

• •

  Importance of combining various patients in the train batches

• •

  Explore the ensemble of multi-signal models and its differences with ensemble of single-signal models

• •

  Establish difference between the EEGs signals present on the dataset.

According to the World Health Organisation (WHO), sleep disorders can be behind other major diseases [1], such as depression, stress, heart problems, diabetes or early Alzheimer [2]. With the current pandemic spreading all over the world, the problems to obtain a good quality sleep have been significantly increased [3], resulting in 40% of the population who claim problems to sleep a straight night. As a  consequence, over time, the activity in the sleeping units in most of the hospitals has registered a significantly increase [4]. That increment in the activity goes along with an increment in the number of polysomnographies (PSG) to be processed by the physicians. PSGs are records of simultaneous signals that can be registered during sleep, such as Electrocardiograms, Electroencephalograms (EEGs), Electromyograms (EMG), breath or movement records, etc. These records taken during periods between 6 to 24 hours are the primary tool of the specialist to study, identify and treat the sleep disorders.

To process those records, the physicians have to label the signals identifying the different stages based mainly on their expertise. That labelling process usually follows one of the two guidelines established by the community known as Rechtschaffen & Kales (R&K) [5] and the American Academy of Sleep Medicine (AASM) [6]. The main problem with those guidelines is the absence of a clear set of features to identify the different sleep stages. Therefore, the result is a difficult situation prone to make mistakes due to fatigue or monotony[7], where each specialist has to process a significant quantity of data in a highly time-consuming process.

In order to solve this situation, different approaches have been made to automate this process. However, this process is particularly tough for any machine learning approach due to the unreliability of the ground truth. For example, in [8] the physicians involved in the study only agree in the 81% of labels for the same PSGs, while in [9] it was set around 84%. Therefore, there is much room for personal interpretation bound to this process, which drives to inconsistent ground truth.

Several techniques have been used to deal with this issue especially focusing on using different automatic feature extraction techniques, such as Genetic Programming [10]. However, Convolutional Neural Networks (CNN) has become a hot topic when, due to their capability to extract information, the specific features to solve a problem are unknown. In fact, this kind of technique has been used previously to tackle the same problem [11, 12, 13]. However, independently of the approach, all the previous works share in common that the techniques only deals with EEGs, while the remaining of the PSG is not processed or ignored in order to simplify the processing. Only in [13], other signals are used but they are treated separately before combining the information in the last steps. The tendency of the later approach is to generate silos of information instead of combining the information extracted from the different sources.

Taking as frame the use of Deep Learning (DL) Models to process the information contained in the signals of a PSG, this paper explores which is the impact of the inclusion of the EMG signal in the process of scoring one of the biggest public datasets of PSGs known as SHHS-1 [14]. The idea is to measure if there is any improvement in the convolutional network models and if the CNNs are able to extract information that combines heterogeneous signals in order to improve the performance, as other works have made with only the EEGs [15, 12]. Additionally, with the best models obtained for each combination of signals, different ensembles have been constructed and compare to the ensemble of the single-signal-input models in order to identify if there is additional information extracted by the multi-signal-input models. The result is a model with comparable results to the current state-of-the-art but higher precision and recall for the less common class which also requires a fraction of the parameters of [11].This last point also benefits not only a significative reduction in complexity but in training time.

The outline of the paper is organised as follows: section 2 present a review of related works with special attention to the most modern ones which have used part of the same dataset. Section 3 presents not only the architecture used as base, but a summary of the key elements of convolutional neural networks and a description of the dataset and the signals used on the tests. Those tests and the analysis of the results could be found in 4. Finally, section 5 and section 6 present the conclusions and the future lines to work, respectively.

Over the years, the problem of scoring the signals from PSG has received uncountable efforts to improve the extraction of features. Particularly, the one which has especially hogged the light among any other has been the processing of EEGs, while other uses have been less prominent.

Focusing on EEGs processing, many researchers have attempted different approaches to improve the automation of the sleep stage scoring. Although, the number of work is significantly lower if the spotlight is over those that performed an automatic extraction of features. Probably, [16] is the first worth to be mention in this sense, where the authors combine Fuzzy Logic with an iterative method to identify the sleeping stages by extracting the features from the signal timeline. Another work worth to be mention that also kept the integrity of the signals’ timeline is [17], which applies a decision tree on the raw signal. This point is important due to it could help physicians to understand the decisions of the system. On the other hand, others works have preferred to apply different transformations to the signal, such as the calculation of the entropy together with a Linear Discriminant Analysis (LDA) [18] or the use of wavelets[19]. Going along with the line, [20] proposes to calculate a combination of energies of the signal and used an Artificial Neural Network (ANN) to perform the classification. Also, closely related, there are those works that do not use the signal but they focus on high-level features such as statistical features [21], power spectral density [22], graph theory features [23] or moment features [24].

However, the ones that have focused the attention on recent times have been those related to any Deep Learning approach, such as, Random Belief Networks [25] or Convolutional Neural Networks [26] due to their capabilities to identify the features to perform the classification. For example, [27] uses Autoencoders to score a well-known dataset of 20-patient [28]. Supratak et. al. [29] proposed an alternative architecture named DeepSleep which was used with the same dataset but also tested with another very common one [30], although at the cost of a very big architecture with a dual pipeline. Trying to address this issue, Fernandez-Blanco et. al. proposes the use of Depth-wise separable convolutional neural networks (DS-CNN). This latter work goes along the line of shown in [12] which points out to an advantage of processing several signals at the same time with a Convolutional Neural network. The results were equivalent to architectures 1000 times bigger in the number of parameters [31, 11] which also use the same dataset, known as SHHS1 [14]. This dataset is up to date the biggest collection of PSG recorded by National Sleep Research Resource with records of 5804 patients.

The main problem of any of the previously mention works is that, without exception, all of them use only one or two at most of the data signals present in the PSG, and they are always EEGs. However, in any PSG, there are more signals than the EEGs and its addition can show benefits in terms of performance. Therefore, recently, some works have tried to explore the incorporation of alternative signal into the classification. Most of those approaches, e.g. [32, 33, 34], are very limited attending to its extension with a very small number of patients. The first work which is worth to be mention is [13], which focused on using CNN to process each one of the types of signal in the study and combine the result with a  bidirectional long short-term memory in order to focus on the temporal component of the signals, but the separation of the signals in 3 different CNNs make to the network impossible to extract features combining the information of both sources. On the other hand, it is also worth to be mention [35]. This work explores the same database among others with an ensemble approach by developing different CNN models trained with several sources. However, it also presents the same problem as [13] because the models used on the ensemble are always developed on a single type of signal and then combined in an ensemble. Additionally, this work does not use the complete dataset, but it takes a subset of patients which makes their results very difficult to compare. This work is also going to incorporate a different kind of signal in the classification, Electromyogram (EMG). However, as the main difference with the previously mentioned ones, in this work, the models are going to explore the active combination of the signals in a single model. This allows the network to extract features that can combine both signals in order to obtain more information. Additionally, and using the whole SHHS-1 dataset, it is going to measure the impact of using those new heterogeneous-signal models in different ensembles in order to complement the information extracted by another model.

In this paper, two sets of tests are presented to analyse the influence of using additional signals to score the sleep stages of a PSG. The first set of tests will explore the influence of use as input more signals and its impact on the classification. Additionally, these first tests were also used to measure the influence of combining more patients during training in the batches. After that set, this paper goes a step further and it also explores the impact of those compose models when they are used in an ensemble schema instead of simple ones. The aim of that last set is to identify if the models with more than one signal can identify additional information which cannot be capture with the ensemble of simple models.

For the test, the SHHS-1 dataset was split into 3 subsets for training, validation, and test. Two points should be highlighted about this division, first, the split process was done according to the number of patients instead of the patterns or sections. The reason is to perform what is called a patient-record division in order to keep the 3 different datasets as much separated as possible and closer to a real scenario. Second, the division was performed with a hold-out schema according to 70:10:20 proportion for each one of the three subsets: training, validation, and test, respectively. These subsets were saved and reused in each of the tests of this paper in order to ensure the comparability of the results.

Once the datasets were defined, a series of experiments were carried out following always the same pipeline with the parameters described in Table 2. All tests shown in this paper have been executed in the same machine, an Intel i7 2.6GHz with 32Gbytes of RAM and the support of an NVidia Titan V graphical card. To run each experiment, i.e. run the training and test, the described machine has taken between 2 and 8 days.

Starting from the same architecture, a training process was performed in mini-batches of 32 patterns. These training steps used as input patterns 5 consecutive sections of 30 seconds sampled at 125Hz, as it was described in section 3.3, and the label of the section in the middle as the desired output of the system. The training is performed by calculating the error of the mini-batches with the cross-entropy as loss function, this measures the difference between the output of the network which uses a Softmax function and the desired output codified as a one-hot encoded. In order to reduce the error, the adjustment is performed through a gradient optimisation technique called ADAM [50] with an initial learning rate of $10^{-4}$. It is also worth to be mention that previous to the last layer, which is the only fully-connected layer, it was set a Dropout layer with a probability of 0.5 in order to improve the generalisation. A threshold of 100 iterations in training was set, after each of which the validation dataset was evaluated. The validation dataset was used to set an early stop threshold if after 10 training cycles the loss of the validation did not improve. The system would return the model with the lowest validation error and this is the one which is going to undertake the evaluation with the test dataset. As previously mentioned, the test dataset was composed by 1161 patients previously seen by the network neither on training nor the validation. In Table 3, a summary of the sections contained in the test is presented along with the percentage of each particular label with respect to the total.

Following the described process, a set of experiments were executed. These experiments used the same architecture described in the section 3.4, while they only change the number of inputs between the 3 possibilities as it can be seen in Table 4. For each of those combination of signals, different training processes using alternative number of individuals to compose the mini-batches were also explored and F1 results are presented in Table 5. In that last Table, the best model for each combination of signals is bold. The reason behind also testing the number of patients in the batches is to identify if there is any influence in the training. Due to the fact of the big size of the data, the whole dataset cannot fit into memory and it has been studied if this has any effect in the training by changing the number of patients between 1 to 8, which is the biggest size that the computer used on the test can handle.

Even though the single-signal models have also good results, the results point out to the improvement of the multi-signal models. When we compare the multi-signal model with the models that use one of the signals used as the input, the result is better in the majority of the cases. Another point that needs also to be highlighted is the fact that, when additional inputs are incorporated into the model, mixing sections from various patients results essential for the generalisation of the model. This last point is especially evident in Figure 5 where each model is represented according to its $F_{1}$ and the number of patients combined in the training mini-batches. That figure exhibits an increasing Although there is not a clear winner between single-signal and multi-signal models, the data clearly show that without the exception of the model that uses the signal C4A1 the multi-signal ones are usually better especially as more patients are combined in the mini-batches. In that same figure, the plot on the right shows the same information that the one on the left but removing the models that perform less than 0.70 in $F_{1}$ to improve the readability. It can be seen an increasing tendency on the multi-signal models while the C4A1 model stabilises its progression since 5 patient.

On the Figure 6, the models which score at least 0.73 in the $F_{1}$ are represented. This threshold was set considering the average of the models which perform over 0.7 and they could be considered comparative to the actual state of the art. A closer look to the information on that figure shows two key points. First, most of the models on the table are multi-signal models. In fact, if a we put the spotlight on the top 10 models represented in Table 6, the multi-signal models achieve 8 out of 10 in that table which ranks the models according to their performance in $F_{1}$ along with the Accuracy and $\kappa$. The second point to highlight is that C4A1 seems to have more information in this dataset while, in most of the single-signal models of the literature, both EEGs have been indistinctly used. The models using this signal systematically obtain better results and it is in most of the best models as can be seen in Figure 6 and it is also on all top models in Table 6.

In Table 7,  the best models for each input combination can also be found. It shows how the inclusion of new signals in the input always improve the best result of the single-signal model in Accuracy, $\kappa$ and $F_{1}$ with a minimum additional cost in terms of resources needed by the model. The multi-signal models have the capability to extract features which relate to both signals instead of only one signal which is not possible is to improve the results from a better quality data, i.e., the model with C3A2 and EMG signals do not improve the model which uses C4A1 alone, however, it significantly improves the results from C3A1 model and EMG model.

Moreover, following the ensemble approach described in Section 3.5, all possible combinations of ensembles between the best models in Table 7 were tested. The aim of these second set of tests is to check if the multi-signal models are able to capture additional information that single-signal models cannot. In figure 7, with the exception of the model only using EMG signal due to its very poor performance, the best single-signal models are represented with the ensembles of all the models presented in Table 7. It is also clear that when the multi-signal models are included the performance in Accuracy and F1 is improved. The top 10 ensemble models can be seen on Table 8 together with the results for the ensemble of the single-signal models and the best single-signal model as reference.

Therefore, the best models include at least one model with multi-signal and, in particular, with a heterogeneous mixture of signals. In fact, when the results are compared to the ensemble of single-signal models, it is clear that at least the mixture of several signals improve the classification as it is shown on the two best models which have all three members of the ensemble as multi-signal models. Additionally, analysing the confusion matrix, Table 9 of the best ensemble it is clear that is able to be a very robust and precise model compared to the single models in most of the classes except with the class N1 which is quite underrepresented in the dataset and even the physicians have problems to identify it, too.

Overall, the general performance of the network can be seen in Table 10 together with the other works which has used the same dataset. It should be highlighted that the presented work, oppositely to the one presented by [11], do not perform any additional filtering or clipping of the signals and different to  [13] do not treat the imbalanced by weighting the training.

Even though the model has significantly more parameters, its performance is equal o better to the state of the art. Only [11] presents better results, however, this work removes some patients from the dataset and clips the signals in order to remove the periods of awake previous to the start of the sleep and it has treated the imbalance during training. Therefore the results could not be exactly reproduced, although it should be mentioned that the model presented in this paper C4A1 uses a similar architecture while using the same datasets for train, validate and test than this one. The results of that model are on Table 8 and, as it was mentioned the proposed ensemble model improves those values. Moreover, comparing the precision in Table 11, it shows this work is the best to deal with the class N1 while showing high values on the remaining classes.it should be mentioned that the precision per class in [31] was not reported, so it was taken out of the comparison.

Several conclusions can be extracted from this work, first and foremost, a new approach based on an ensemble of PS-CNN improve how to tackle the problem of scoring the sleep stages. The proposed model has shown to be able to identify the less frequent class better than any previous approach while it keeps similar percentages in the remaining classes. Different from previous approaches such as [11, 13], this point has been achieved without any additional imbalance treatment, as seen on table 11. Moreover, the proposed approach shows, on the authors’ best known, the best results on this dataset up to date, without trimming the signal or removing patients.

Additionally, it is also worth mentioning the deep study of the different combinations of signals which has to draw two main conclusions. First, the inequality of the EEGs recorded in the dataset. As it can be seen in Table 5, the channel C4A1 seems to have more information alone than the other two signals while using the same datasets. However, the second conclusion is that this information can be complemented with additional signals, which allow to mine information that relates the two signals in a single convolutional model. Even with the EMG which has shown the poorest performance of all signals alone, when it is combined with one of the EEGs channels the results are significantly better than the results of models using each EEG alone.

Those models were later used to compose different ensembles, and those ensembles have shown how the use of multiple signals, especially when they are heterogeneous, has better results than the ensemble of single-signals models of any models alone.

Finally, it should also be mentioned that this type of ensemble is possible due to the use of models which use Separable Convolutional layers. The reduction in the number of parameters allows the composition, train and use of this kind of model not only in high-end infrastructures but in modest computers. For example, as it can be seen on table 10 a similar ensemble with the model proposed in [11] would require no less than 600 million parameters while the proposed here are roughly 51,000.

From this work, there are several lines which can be explored. Going along with the same dataset, the SHHS-1 has several signals more, such as electrooculograms or electrocardiograms. However, these other signals have a different sample ratio than the EEG ad EMG, therefore, exploring the best way to incorporate that information in the classification models could be the next step. Although some recent works such as [51] have shown their advantages on small datasets. Several are the possibilities in this sense, from the inclusion of specific models for signals with a different ratio in an ensemble model to the incorporation of that signals in a Deep Learning model.

Second, another point that should be explored, which is related to the previous one, is how to treat the missing values. Some recent works have started to dig into this question, such as [13], performing an evaluation in the time domain based on Bidirectional Long Short Term Memory (Bi-LSTM) architecture. However, other possibilities should be explored from the use of Generative Adversarial Networks to a simply interpolation.

Third, another point to be improved is the problem that this dataset has with the class N1 which is significantly underrepresented in the whole dataset. Although the presented model represents an improvement over the state-of-the-art, dealing with this problem could be another object of study. For example, Weighting during training the mistakes classified of the less represented could be one of the approaches to compare with other.

Fourth, although the presented model shows a significant performance on the SHHS-1, other works have used different channels or signals from other datasets. How the transfer learning [52, 53] would work in this kind of signals, using a different and smaller dataset such as[30], is a question to check.

Finally, the issue of the inconsistent and unreliable ground truth should also be tackle. A schema such as the ones recently presented like the man-in-the-loop [54] architecture could be a possible approach. In fact, this should not be focused exclusively on the improvement of the performance but the identification of the right set features to label the signals instead of the expertise.

The authors want to acknowledge the support from NVdivia, who has donated the GPU used in the experiments of this publication, and Centro de Supercomputación de Galicia (CESGA), who allows to conduct the first exploratory tests on their installations.

This work is partially supported by Instituto de Salud Carlos III, grant number PI17/01826 (Collaborative Project in Genomic Data Integration (CICLOGEN) funded by the Instituto de Salud Carlos III from the Spanish National Plan for Scientific and Technical Research and Innovation 2013–2016 and the European Regional Development Funds (FEDER)—“A way to build Europe.”. It was also partially supported by different grants and projects from the Xunta de Galicia [ED431D 2017/23; ED431D 2017/16; ED431G/01; ED431C 2018/49; IN845D-2020/03]. Finally, another source of support was the CYTED network (PCI2018_093284) funded by the Spanish Ministry of Innovation and Science.