AUTOMATIC METHOD TO DELINEATE OR CATEGORIZE AN ELECTROCARDIOGRAM

[0001] The present invention relates to temporal signal analysis, preferably ECG analysis, using at least one neural network.

[0002] The electrocardiogram (ECG) is a graphic representation of the electrical activity of the heart. It is recorded from the body surface using a number of electrodes placed in specific predefined areas. It is considered as a fundamental tool of clinical practice. It is a simple, non-invasive exam that can be performed by any health professional. Placing the electrodes is not considered as a medical procedure, yet in some countries, the prescription of the ECG by a doctor is essential for it to be performed. The ECG constitutes the first step in cardiovascular diseases (CVD) diagnosis, and is used multiple times throughout the life of a CVD patient. CVD constitute the first global cause of death.

[0003] An ECG is composed of multiple temporal signals, called lead signals, such as the standard 12-lead ECG shown in FIG. 1. An ECG displays repeating patterns usually comprising a P-wave, a QRS complex and a T-wave, respectively corresponding to the depolarization of the atria, depolarization of the ventricles and repolarization of the ventricles. These waves and complex are shown in FIG. 2, which focuses on a couple of beats in one lead signal.

[0004] The ECG allows for the detection of many anomalies, which often in turn point to specific CVD. It is estimated that about 150 measurable anomalies can be identified on ECG recordings today. However, without specific expertise and/or regular training, only a small portion of these anomalies can be easily spotted. Unfortunately, today, it is estimated that only one third of ECGs are performed in settings where cardiology expertise is readily available.

[0005] In order to make ECG interpretation simpler and assist non-specialists, two alternatives exist today, but neither fully satisfy the needs of health professionals:

• • [0006] Telecardiology centers, where an interpretation of an ECG sent by a non-specialist is delivered either by cardiologists or by specialized ECG technicians. Their interpretations are of high quality but are slow and expensive to obtain.
  • [0007] Prior art automated ECG interpretation softwares, which are mainly developed by ECG device manufacturers. They provide low quality interpretation (false alarms are very frequent) but deliver them in seconds.

[0008] Prior art automated ECG interpretation softwares can provide two types of information about an ECG signal:

• • [0009] the temporal locations of each wave, called its “delineation”, and/or
  • [0010] a classification of the ECG as normal/abnormal or labeling its anomalies.

[0011] Two main approaches are used for the delineation of ECG signals.

[0012] The first one is based on multiscale wavelet analysis. This approach looks for wavelet coefficients reaching predefined thresholds at well-chosen scales (Martinez et al, IEEE transactions on biomedical engineering, Vol. 51, No. 4, April 2004, 570-581, Almeida et al., IEEE transactions on biomedical engineering, Vol. 56, No. 8, August 2009, pp 1996-2005, Boichat et al., Proceedings of Wearable and Implantable Body Sensor Networks, 2009, pp 256-261, U.S. Pat. No. 8,903,479, 2014 Dec. 2, Zoicas et al.). The usual process is to look for QRS complexes, and then look for P waves on the signal before the complexes, and after them for T waves. This approach can only handle a single lead at a time, sometimes using projection to one artificial lead (US 2014/0148714-2014 May 29, Mamaghanian et al). This computation is made very unstable by the use of thresholds. The approach is also limited as it can neither deal with multiple P waves nor with “hidden” P waves. A hidden P wave is a P wave which occurs during another wave or complex, such as for example during a T wave.

[0013] The second one is based on Hidden Markov Models (HMM). This machine learning approach considers that the current state of the signal (whether a sample is either part of a QRS complex, a P wave, a T wave or no wave) is a hidden variable that one wants to recover (Coast et al., IEEE transactions on biomedical engineering, Vol. 37, No. 9, September 1990, pp 826-836, Hughes et al., Proceedings of Neural Information Processing Systems, 2004, pp 611-618, U.S. Pat. No. 8,332,017, 2012 Dec. 11, Trassenko et al). To this end, a representation of the signal must be designed using handcrafted “features”, and a mathematical model must be fitted for each wave, based on these features. Based on a sufficient number of examples, the algorithms can learn to recognize each wave. This process can however be cumbersome since the feature design is not straightforward, and the model, usually Gaussian, is not well adapted. Also, none of these works has considered the situation of hidden P waves.

[0014] As for anomalies and/or CVD detection, most algorithms use rules based on temporal and morphological indicators computed using the delineation: PR, RR and QT intervals, QRS width, level of the ST segment, slope of the T wave, etc. . . . . These rules such as the Minnesota Code (Prineas et al., Springer, ISBN 978-1-84882-777-6, 2009) were written by cardiologists. However, they do not reflect the way the cardiologists analyze the ECGs and are crude simplifications. Algorithms such as the Glasgow University Algorithm are based on such principles (Statement of Validation and Accuracy for the Glasgow 12-Lead ECG Analysis Program, Physio Control, 2009).

[0015] More advanced methods use learning algorithms, and are built using a diagnosis and an adequate representation for each ECG they learn from; however, in these methods, once again, it is necessary to seek a representation of the raw data into a space that preserves the invariance and stability properties. Indeed, an ECG signal varies significantly from one patient to another. It is therefore extremely difficult for an algorithm to learn how to discriminate different diseases by simply comparing raw data. A representation which drastically limits this interpatient variability while preserving the invariance within the same disease class must be chosen.

[0016] In order to solve the above-mentioned issues, the Applicant turned to architectures called neural network. Such architectures have been intensively studied in the field of imaging (Russakovsky et al., arXiv:1409.0575v3, 30 Jan. 2015), but limitations arose when, very recently, the first scientific teams attempted to apply them to ECGs (Zheng et al., Web-Age Information Management, 2014, Vol. 8485, pp 298-310, Jin and Dong, Science China Press, Vol. 45, No 3, 2015, pp 398-416). Indeed, these prior arts only limit the classification to an attempt to identify normal ECG versus abnormal ECG, or to perform a beat-to-beat analysis. The beat-to-beat analysis adds a preprocessing step while reducing the ability of the neural network to learn some anomalies: rhythm disorders, for example, cannot be identified from the observation of a single beat. In fact, these algorithms only consider single-label classification whereas multi-label classification is essential in ECG interpretation, since one ECG can present several anomalies.

[0017] Thus, there is a need for computerized algorithms able to analyze ECG that can:

• • [0018] carry out the analysis without constraints from the ECG recording duration;
  • [0019] carry out the analysis without the need for beat-by-beat processing, or feature extraction;
  • [0020] obtain the delineation of the signal, including identification of hidden P waves;
  • [0021] provide a multi-label classification directly from a full ECG, possibly exhibiting multiple labels;
  • [0022] be fast, stable and reliable.

[0023] To address the above issues in ECG analyses, the Applicant developed two techniques based on convolutional neural networks.

• • [0024] A fully convolutional neural network which first gives a dense prediction of the probability of presence of each wave on each time stamp of the ECG, then post-processes the signal to produce its delineation. This novel approach of the delineation, using convolutional networks, allows the processing of ECGs of any duration, analyzing all types of waves in the same way, without being constrained by their positions.
  • [0025] A recurrent convolutional neural network which predicts directly multiple labels on the whole ECG signal. This structure allows the processing of an ECG of any duration, and takes into account the time dynamic in its analysis. It results in a fixed format output (multi-labels).

[0026] Thus, the present invention relates to a method for computerizing the delineation of an ECG signal, comprising: applying a fully convolutional neural network NN1 to said ECG, whereby the fully convolutional neural network NN1 reads each time point of the ECG signal, analyzes temporally each time point of the ECG signal, assigns to each time point of the ECG a score for at least the following: P-wave, QRS complex, T-wave, and then, optionally and whenever necessary, displaying the scores according to time, optionally with the ECG signal.

[0027] According to an embodiment, the method further comprises a pre-treatment step, wherein the pre-treatment comprises denoising and removing the baseline of the ECG signal as well as expressing it at a chosen frequency prior to the application of NN1.

[0028] According to an embodiment, the method further comprises a post-treatment step, so as to produce the time points of the beginning and the end of each wave in the ECG signals, called the onsets and offsets.

[0029] The invention also comprises a software comprising a trained neural network for delineation of an ECG signal. The invention also comprises a computer device comprising a software implementing a method for delineation of an ECG signal, comprising applying a fully convolutional neural network NN1 to said ECG, as described above.

[0030] This invention also includes a method for computerizing multi-label classification of an ECG signal, comprising applying a convolutional neural network NN2 to said ECG, whereby the recurrent neural network NN2 reads each time point of the ECG signal, analyzes each time point of the signal, computes scores for each anomaly and allots to an ECG at least one disease-related label, and then, optionally, displaying the labels, preferably in the form of a list of detected anomalies.

[0031] According to an embodiment, the method further comprises a pre-treatment step, wherein the pre-treatment comprises denoising and removing the baseline of the ECG signal as well as expressing it at a chosen frequency prior to the application of NN2.

[0032] According to an embodiment, the method further comprises a post-treatment step, so as to produce the onset and offset of each wave in the ECG signal.

[0033] The invention also comprises a software comprising a trained neural network for multi-label classification of an ECG signal. The invention also comprises a computer device comprising a software implementing a method for multi-label classification of an ECG signal, comprising applying a recurrent neural network NN2 to said ECG, as described above.

[0034] Furthermore, the invention also includes a method for computerizing delineation and multi-label classification of an ECG signal, comprising applying a trained recurrent neural network NN3 to said ECG, whereby the recurrent neural network NN3 reads each time point of the ECG signal, analyzes temporally each time point of the signal, assigns a score for at least the following: P-wave, QRS complex, T-wave, no wave; computes scores for each anomaly, and then, optionally, displaying the interval labels according to time and the anomaly scores, preferably in the form of a list of detected, optionally with the ECG signal.

[0035] According to an embodiment, the method further comprises a pre-treatment step, wherein the pre-treatment comprises denoising and removing the baseline of the ECG signal as well as expressing it at a chosen frequency prior to the application of NN3.

[0036] According to an embodiment, the method further comprises a post-treatment step, so as to produce the onset and offset of each wave in the ECG signal.

[0037] The invention also comprises a software comprising a trained neural network for delineation and multi-label classification of an ECG signal. The invention also comprises a computer device comprising a software implementing a method for delineation and multi-label classification of an ECG signal, comprising applying a neural network NN3 to said ECG, as described above.

[0038] The present invention relates to temporal signal analysis, preferably ECG analysis, using at least one convolutional neural network.

[0039] The framework used here is the one of supervised learning. The aim of supervised learning is to predict an output vector Y from an input vector X. In the Applicant embodiment, X is an ECG (a multivariate signal) as a matrix of size m×n. As for Y, in the Applicant embodiment, it can be:

• • [0040] the delineation, providing a score for each sample of X to be part of one of the different waves as a matrix of size p×n;
  • [0041] the scores for each anomaly as a vector of size q;
  • [0042] the set composed of both the delineation and the vector of scores.

[0043] The problem of supervised learning can also be stated as follows: designing a function f such that for any input X, f(X)≈Y. To this end, the function f is parametrized, and these parameters are “learned” (parameters are optimized with regards to an objective loss function, for example, by means of a gradient descent (Bishop, Pattern Recognition and Machine Learning, Springer, 2006, ISBN-10: 0-387-31073-8).

[0044] A neural network is a particular type of function f, aiming at mimicking the way biological neurons work. One of the most basic and earliest neural network is the perceptron (Rosenblatt, Psychological Review, Vol. 65, No. 6, 1958, pp 386-408). From the input X, it computes linear combinations (i.e. weighted sums) of the elements of X through a multiplication with a matrix W, adds an offset b, and then applies a non-linear function σ, such as for instance a sigmoid, on every element of the output:

f(X)=σ(WX+B)

[0045] The parameters which are learned in a perceptron are both W and B. In practice, more general neural networks are just compositions of perceptrons:

f(X)=σ_n(W_n . . . σ_n(W₁X+B₁)+B_n)

[0046] The output of a perceptron can be sent as input to another one. The input, the final output, and the intermediate states are called layers. The intermediate ones are more specifically called hidden layers, since only the input and the final output are observed. For instance, a neural network with one hidden layer can be written as:

f(X)=σ₂(W₂σ₁(W₁X+B₁)+B₂)

[0047] Such a network is shown in a graphic form as an example in FIG. 3. The vector X enters the network as the input layer, each element of the hidden layer is then computed from linear combinations of all elements of X (hence all the links), and the element of the output layer are then computed from linear combinations of all elements of the hidden layer.

[0048] It has been shown that neural networks in their general form are able to approximate all kinds of functions (Cybenko, Math. Control Signals Systems, Vol. 2, 1989, pp 303-314). The term “deep learning” is used when a neural network is composed of many layers (though the threshold is not perfectly defined, it can be set to about ten). This field arose mostly in the last decade, thanks to recent advances in algorithms and in computation power.

[0049] Convolutional neural networks are a particular type of neural networks, where one or more of the matrices W_i which are learned do not encode a full linear combination of the input elements, but the same local linear combination at all the elements of a structured signal such as for example an image or, in this specific context, an ECG, through a convolution (Fukushima, Biol. Cybernetics, Vol. 36, 1980, pp 193-202, LeCun et al., Neural Computation, Vol. 1, 1989, pp 541-551). An illustration of a convolutional neural network is shown in FIG. 6. Most convolutional neural networks implement a few convolutional layers and then standard layers so as to provide a classification. A network which only contains convolutional networks is called a fully convolutional neural network. Finally, a recurrent convolutional neural network is a network composed of two sub-networks: a convolutional neural network which extracts features and is computed at all time points of the ECG, and a neural network on top of it which aggregates in time the outputs of the convolutional neural network. An illustration of a recurrent convolutional neural network is provided in FIG. 7.

[0050] As mentioned above, an ECG is represented as a matrix of real numbers, of size m×n. The constant m is the number of leads, typically 12, though networks could be taught to process ECG with any number of leads. The number of samples n provides the duration of the ECG n/f, with f being the sampling frequency of the ECG. A network is trained for a given frequency, such as for example 250 Hz or 500 Hz or 1000 Hz, though any frequency could be used. A same network can however process ECG of any length n, thanks to the fact that it is fully convolutional in the embodiment of the delineation, or thanks to the use of a recurrent neural network in the embodiment of the anomaly network.

[0051] In both the delineation and the multi-label classification embodiment s, networks are expressed using open softwares such as for example Theano, Caffe or Torch. These tools provide functions for computing the output(s) of the networks and for updating their parameters through gradient descent. The exact structure of the network is not extremely important as long as they are deep structures: fully convolutional in the situation of the delineation network (Long et al., Proceedings of Computer Vision and Pattern Recognition, 2015, pp 3431-3440), and convolutional (Krizhevsk et al., Proceedings of Neural Information Processing Systems, 2012, pp 1097-1105), potentially recurrent in the situation of the multi-label classification network (Donahue et al., arXiv:1411.4389v3, 17 Feb. 2015 and Mnih et al., arXiv:1406.6247v1, 24 Jun. 2014). The 2D convolutional layers which were used on images are then easily converted into 1D convolutional layers in order to process ECG signals.

[0052] This invention also pertains to a method for manufacturing a neural network for delineation of an ECG, by training it.

[0053] The training phase of the neural networks in the embodiment of delineation consists in the following steps:

• • [0054] taking one ECG from a dataset containing ECGs and their known delineation; the ECG being expressed as a matrix of size m×n with m fixed and at a predefined frequency;
  • [0055] expressing the delineation of this ECG under the form of a matrix y of size p×n where p is the number of annotated types of wave; typically p=3, so as to identify P waves, QRS complexes, and T waves; annotations are usually expressed as lists of wave with their start and end points such as for example: (P, 1.2 s, 1.3 s), (QRS 1.4 s 1.7 s), (T, 1.7, 2.1), (P, 2.2, 2.3); in this example, the first row of y, corresponding to P waves, will be 1 for samples corresponding to times between 1.2 and 1.3 s, and between 2.2 and 2.4 s, and 0 otherwise; row 2 will correspond to QRS complexes and row 3 to T waves;
  • [0056] computing the output of the network for this ECG;
  • [0057] modifying the parameters of the network so as to decrease a cost function comparing the known delineation and the output of the network; a cross-entropy error function is used so as to allow for multi-labeling (allowing for multiple waves at a given instant); this minimization can be done though a gradient step
  • [0058] repeating steps 1 to 4 at least once for each ECG of the dataset;
  • [0059] recovering the neural network.

[0060] This invention also provides a method for manufacturing a neural network for the categorization of an ECG signal, by training it.

[0061] In a multi-label classification, the manufacturing/training process includes the following steps:

• • [0062] taking one ECG from a dataset containing ECGs and their known anomaly labels; the ECG must be expressed as a matrix of size m×n with m fixed and at a predefined frequency;
  • [0063] expressing the anomalies as a vector of size q, with q the number of anomalies to identify; this vector could be [0; 1; 0; 0; 1; 0; 0; 0] for q=8; a 1 is set in the vector at the index corresponding to the anomalies which are present: in the above example, the ECG exhibits two anomalies;
  • [0064] computing the output of the network for this ECG;
  • [0065] modifying the parameters of the network so as to decrease a cost function comparing the known label vector and the output of the network; a cross-entropy error function is used so as to allow for multi-labeling (allowing for multiple anomalies for an ECG); this minimization can be done though a gradient step;
  • [0066] repeating steps 1 to 4 at least once for each ECG of the dataset;
  • [0067] recovering the neural network.

[0068] This invention also provides a method for manufacturing a neural network for both the delineation and the categorization of an ECG signal, by training it.

[0069] In the embodiment of the combination of delineation with multi-label classification, the manufacturing process includes the following steps:

• • [0070] taking one ECG from a dataset containing ECGs and their known anomaly labels; the ECG must be expressed as a matrix of size m×n with m fixed and at a predefined frequency;
  • [0071] expressing the anomalies as a vector of size q, with q the number of anomalies to identify; this vector could be [0; 1; 0; 0; 1; 0; 0; 0] for q=8; a 1 is set in the vector at the index corresponding to the anomalies which are present: in the above example, the ECG exhibits two anomalies;
  • [0072] expressing the delineation of this ECG under the form of a matrix Y of size p×n where p is the number of waves to identify; typically p=3, so as to identify P waves, QRS waves, and T waves; annotations are usually expressed as lists wave type with their start and end points such as for example: (P, 1.2 s, 1.3 s), (QRS 1.4 s 1.7 s), (T, 1.7, 2.1), (P, 2.2, 2.3); in this example, the first row of Y, corresponding to P waves, will be 1 for samples corresponding to times between 1.2 and 1.3 s, and between 2.2 and 2.4 s, and 0 otherwise; row 2 will correspond to QRS complexes and row 3 to T waves;
  • [0073] computing both outputs of the network for this ECG;
  • [0074] modifying the parameters of the network so as to decrease the sum of a cost function comparing the known label vector and one of the output of the network, and a cost function comparing the delineation and the other output; cross-entropy error functions are used to allow for multi-labeling (allowing for multiple anomalies for an ECG as well as multiple waves at any time point); this minimization can be done though a gradient step;
  • [0075] repeating steps 1 to 4 at least once for each ECG of the dataset;
  • [0076] recovering the neural network.

[0077] This invention also pertains to a method and a device for delineation of an ECG signal, implementing a fully convolutional neural network trained for delineation of an ECG signal as described above.

[0078] As a basis, it shall be understood that the ECG is expressed as a matrix X of size m×n at the frequency used for training the networks. The ECG is used as input of the trained neural network.

[0079] The neural network then reads each time point of the ECG signal, analyzes spatio-temporally each time point of the ECG signal, assigns a temporal interval score to anyone of at least the following: P-wave, QRS complex, T-wave. It then recovers the output of the neural network, as a matrix Y of size p×n. An example is shown in FIG. 4: the first signal shows one of the leads of the ECG (to help the visualization), the following 3 signals are the outputs of the network, providing scores for P waves, QRS waves and T waves. As it can be seen, these scores are synchronized with the appearance of the aforementioned waves in the signal.

[0080] In a preferred embodiment, the neural network provides scores at each time point as a matrix Y, and a post-processing allows the allocation of each time point to none, single, or several waves, and provides the onset and offset of each of the identified waves. For instance, a sample can be affected to the waves for which the score on the corresponding row of Y is larger than 0.5. This provides a delineation sequence of type (P, 1.2 s, 1.3 s), (QRS 1.4 s 1.7 s), (T, 1.7 s, 2.1 s), (P, 2.2 s, 2.3 s), as recorded in the annotations.

[0081] In an embodiment, finally, the labels may be optionally displayed according to time, optionally with the ECG signal.

[0082] This invention also pertains to a method and a device for multi-label classification of an ECG signal, implementing Long-term Recurrent Convolutional Networks (LRCN, (Donahue et al., arXiv:1411.4389v3, 17 Feb. 2015). These neural networks are trained for multi-label classification of an ECG signal as described above.

[0083] As a basis, it shall be understood that the ECG is expressed as a matrix of size m×n at the frequency used for training the networks. Then, the ECG is used as input of the trained neural network.

[0084] The neural network then reads each time point of the ECG signal, analyzes temporally each time point of the ECG signal, computes a score for each anomaly, recovers the output of the neural network, and optionally displays the labels.

[0085] In a preferred embodiment, the neural network recovers the output as a vector of size q. This vector contains scores for the presence of each anomaly.

[0086] According to an embodiment, the neural network displays the list of found anomalies as the elements for which the score in the vector are higher than a predefined threshold, typically 0.5.

[0087] This invention also pertains to a method and a device for delineation and multi-label classification of an ECG signal, implementing a neural network trained for delineation and multi-label classification of an ECG signal as described above.

[0088] As a basis, it shall be understood that the ECG is expressed as a matrix of size m×n at the frequency used for training the networks. Then, the ECG is used as input of the trained neural network.

[0089] The neural network then reads each time point of the ECG signal, analyzes temporally each time point, assigns a temporal score to all of the following at least: P-wave, QRS complex, T-wave. It then computes a score for each anomaly, recovers both the outputs of the neural network: the first as a matrix y of size p×n, providing scores for at least P waves, QRS waves and T waves; and the second as a vector of size q, said vector containing scores for the presence of each anomaly.

[0090] In a preferred embodiment, a post-processing of the delineation output allows to affect each time point to none, single, or several waves, and provides the onset and offset of each of the identified waves. For instance, a sample can be affected to the waves for which the score on the corresponding row of Y is larger than 0.5. This provides a delineation sequence of type (P, 1.2 s, 1.3 s), (QRS 1.4 s 1.7 s), (T, 1.7 s, 2.1 s), (P, 2.2 s, 2.3 s), as recorded in the annotations.

[0091] According to an embodiment, the neural network displays the list of found anomalies as the elements for which the score in the vector are higher than a predefined threshold, typically 0.5, as well as the delineation, optionally with the ECG signal.

[0092] The invention also comprises a computer device implemented software comprising a trained neural network for delineation of an ECG signal. The invention also comprises a device, such as for example a cloud server, a commercial ECG device, a mobile phone or a tablet, comprising a software implementing a method for delineation, multi-label classification or both, of an ECG signal, as described above.

[0093] According to an embodiment of the invention, a step to prepare the signal and create input variables for classification is further carried out (“pre-treatment”). The purpose of this pre-treatment is to remove the disturbing elements of the signal such as for example noise and baseline, low frequency signal due to respiration and patient motion, in order to facilitate classification. For noise filtering, a multivariate approach functional analysis proposed by (Pigoli and Sangalli, Computational Statistics and Data Analysis, vol. 56, 2012, pp 1482-1498) can be used. The low frequencies of the signal corresponding to the patient's movements may be removed using median filtering as proposed by (Kaur et al., Proceedings published by International Journal of Computer Applications, 2011, pp 30-36).

[0094] According to an embodiment of the invention, a post-treatment step is added, so as to produce the onset and offset of each wave in the ECG signal.

[0095] This invention brings to the art a number of advantages, some of them being described below:

• • [0096] The input of the networks are one or multilead ECG signals with variable length, possibly preprocessed so as to remove noise and baseline wandering due to patients movements, and express the signal at a chosen frequency.
  • [0097] The output of a classification network is a vector of scores for anomalies. These are not classification scores since one ECG can present several anomalies. For example, the output of such network could be a vector [0.98; 0.89; 0.00; . . . ] with the corresponding labels for each element of the vector (Right Bundle Branch Bloc; Atrial Fibrillation; Normal ECG; . . . ). Scores are given between a scale of [0, 1] and the example output vectors therefore indicates a right bundle branch block and atrial fibrillations. A recurrent neural network architecture can be added on the top of the convolutional network (Donahue et al., arXiv:1411.4389v3, 17 Feb. 2015 and Mnih et al., arXiv:1406.6247v1, 24 Jun. 2014). In this way, the convolution network acts as a pattern detector whose output will be aggregated in time by the recurrent network.
  • [0098] The output of the delineation network is a set of signals spanning the length of the input ECG signal, providing the score for being in P wave, a QRS complex, a T wave and potentially other types of waves such as for example premature ventricular complexes, flutter waves, or U waves. An example of output signals is provided in FIG. 5.
  • [0099] The delineation network is not limited to recovering at most one wave at each time point and therefore can identify several waves at any time point, such as for instance hidden P waves.
  • [0100] No works applying convolutional networks to the delineation have been made so far.

[0101] The underlying structure of the networks is not fundamental as long as it is a recurrent network for the multi-label classification network and a fully convolutional network for the delineation network. One can use a structure such as RLCN (Donahue et al., arXiv:1411.4389v3, 17 Feb. 2015 and Mnih et al., arXiv:1406.6247v1, 24 Jun. 2014) for classification and a network similar as the one in (Long et al., Proceedings of Computer Vision and Pattern Recognition, 2015, pp 3431-3440) for delineation. In both embodiments, convolutional layers must be modified as 1D convolutions instead of 2D convolutions.

[0102] A hybrid network, sharing the first convolutional layers and diverging so as to provide both the delineation as one output, and the multi-label classification as another output is also used. This combination has the advantage of being able to produce a multi-label classification helped by the identification of the ECG waves.

[0103] The present invention is further illustrated by the following examples.

[0104] This training was performed on 630 ECGs and the network evaluated on about 900 beats from 70 different patients which were not used for the training phase. The false positive rate was 0.3% for QRS complexes and T waves detection and their false negative rate was 0.2%. They were respectively 5.4% and 4.2% for P waves. The precision of the wave onsets (beginnings) and offsets (ends) are detailed below:

	Point	Standard Deviation (ms)	Bias (ms)
	P wave start	11.2	−1.6
	P wave end	11.2	−4.0
	QRS complex start	6.3	−2.1
	QRS complex end	3.9	−1.1
	T wave end	16.1	−4.6

[0105] Compared with state-of-the art algorithms, the precision was improved. In FIG. 5 for instance, the ECG exhibits an atrioventricular block which means that the P waves and the QRS complexes are completely decoupled. In this example, the P waves are regular and QRS complexes happen at random times. One can observe in this example that the algorithm correctly found two P waves between the first QRS complex and the second QRS complex, while most algorithms would not be able to find them since they look for only one P wave before each complex. The last P wave also starts before the end of the last T wave, adding complexity. Other algorithms would not have been able to find this hidden wave.

[0106] A network has been trained using about 85,000 ECGs and has been evaluated on a Dataset including 20,000 patients which were not used in the training phase. For the evaluation purpose, the multi-label classification obtained was simplified to normal (no anomaly)/abnormal if need be. The results in terms of accuracy, specificity and sensitivity were the following:

Accuracy	Specificity	Sensitivity
0.91	0.88	0.92

[0107] A graphical representation of how a standard multi-label is used on ECGs is displayed in FIG. 6. The ECG is given as input to the network, which aggregates the information locally and then combines it layer by layer to produce a high-level multi-label classification of the ECG, in this example correctly recognizing atrial fibrillations and a right bundle branch block. Such networks however take a fixed size as input and the process must be reproduced at different locations so as to analyze the whole signal. FIG. 7 is an example of a graphic representation of a recurrent neural network which overcomes this issue. This type of network is made from a standard convolutional network computed at all possible locations of the signal, and on top of which comes another network layer which aggregates the information. In this example, the network correctly recognizes a premature ventricular complex (PVC, the fifth and largest beat) in the first part of the signal while the second part of the signal is considered as normal. The aggregated output is therefore PVC since this ECG has an anomaly and cannot therefore be considered as normal.

[0108] In another embodiment, the applicant combines features described above in examples 1 and 2. Such combination enables to combine the advantages of both networks in a unique network, providing similar results for both the delineations and the multi-label classifications.

[0109] FIG. 1 is a photo of an ECG.

[0110] FIG. 2 is a schematic representation of a normal ECG, with the P wave, the QRS complex/wave comprising the Q, R, S and J points, and the T wave.

[0111] FIG. 3 is an example of structure for a basic neural network with no convolutional layer.

[0112] FIG. 4 is an example of the output of the delineation network on a normal ECG.

[0113] FIG. 5 is an example of the output of the delineation network on an ECG with hidden P waves (high degree atrioventricular block).

[0114] FIG. 6 models the way a standard multi-label convolutional network works.

[0115] FIG. 7 models the way a multi-label recurrent convolutional network works.