Computer-implemented method for the detection of electromyographic signal anomalies and system for monitoring such electromyographic signals for the detection of such anomalies

Abstract

The present invention relates to a computer-implemented method (10) for detecting electromyographic signal anomalies. Such a method (10) is characterised in that it comprises the following steps in sequence: a) acquiring a plurality of electromyographic signals relating to a muscle of an individual subject under a condition of general anaesthesia, where each of such electromyographic signals represents the evolution over time of the electrical activity generated by the fibres of the aforesaid muscle stimulated by a nerve impulse; c) creating a dataset defined by a random subset of at least a portion of such electromyographic signals; d) creating a decision tree associated with the dataset where each node is defined using a random feature of the electromyographic signals contained in the dataset; e) repeating step c) and step d) so as to create a random forest logic with a plurality of decision trees; f) acquiring at least one further electromyographic signal relative to the aforesaid muscle of the subject and analysing the further signal using random forest logic in order to obtain an anomaly value.

Description

[0001] COMPUTER-IMPLEMENTED METHOD FOR THE DETECTION OF ELECTROMYOGRAPHIC SIGNAL ANOMALIES AND SYSTEM FOR MONITORING SUCH ELECTROMYOGRAPHIC SIGNALS FOR THE DETECTION OF SUCH ANOMALIES.

[0002] DESCRIPTION

[0003] The invention relates to a computer-implemented method for detecting electromyographic signal anomalies.

[0004] Furthermore, the subject of this patent is a system for monitoring electromyographic signals for the detection of such anomalies.

[0005] Nowadays, they are known methods for detecting electromyographic signal anomalies that are related to the observation of such signals by an experienced technician.

[0006] In particular, the technician observes the evolution of these electromyographic signals over time, trying to identify anomalous behaviours such as sudden variations in the amplitude of these signals.

[0007] The detection of an anomaly in an electromyographic signal, for example during a surgical operation, can be representative of harm that the surgeon has caused, or is about to cause, to the patient’s nervous system.

[0008] Therefore, the detection times as well as the detection sensitivity are extremely important in order to prevent or limit serious harm to a subject during a surgical operation.

[0009] Such a method, although known and appreciated, has some important limitations.

[0010] In particular, the fact that the detection of anomalies is related to the observation of electromyographic signals by an expert technician implies that the time required to detect such anomalies may be too long to prevent a surgeon from causing any harm during a surgical operation.

[0011] Moreover, for the same reason, the sensitivity in detecting such anomalies is too low to prevent any harm caused by a surgeon during a surgical operation.

[0012] A further limitation, however, is related to the fact that the aforesaid method for detecting anomalies is a method that does not take into account any patientspecific factors of the subject concerned, thus implying the risk of ignoring anomalies in electromyographic signals relevant to the patient.

[0013] They are also known systems and related methods employing automatic logics for detecting physiological anomalies in a subject.However, such methods have an important limitation related to the fact that they require supervised learning approaches. Therefore, these methods necessarily require the use of external labels, previously classified datasets or pre-defined physiological references, such as the measurement of induced muscle strength, in order to identify an anomalous behaviour. This approach makes these systems rigid and dependent on a reference target which, in a critical environment such as the operating theatre, is often impossible to obtain or is altered by the patient’s condition itself. As a result, the supervised methods risk failing to detect unexpected patient-specific anomalies that do not fall into the pre-trained categories.

[0014] In the context of this patent, supervised learning refers to a model that is trained using a dataset in which each signal is previously associated with a specific label defined by the operator or by external physiological measurements, limiting the analysis ability to categories already known to the system.

[0015] Unsupervised learning, on the other hand, refers to a model capable of identifying anomalies without the need for external labels or measurements, but relying solely on the statistical isolation of patterns that deviate from the individual subject-specific baseline acquired under anaesthesia.

[0016] The task of the present invention is to develop a method for detecting electromyographic signal anomalies capable of overcoming the mentioned drawbacks and limitations of the prior art.

[0017] In particular, the object of the present invention is to realise a method that is not related to the observation of electromyographic signals by an expert technician. Furthermore, an object of the invention is to develop a method that takes into account certain personal factors of the subject concerned.

[0018] The object of the present invention is also to develop a method for detecting electromyographic signal anomalies based on a completely unsupervised algorithmic model.

[0019] The above-mentioned task and objects are achieved by a computer-implemented method for the detection of electromyographic signal anomalies according to claim 1.

[0020] Further features of the method according to claim 1 are described in the dependent claims.

[0021] The aforesaid task and objects, together with the advantages that will be mentioned hereinafter, are highlighted by the description of an embodimentof the invention, which is given by way of indicative though non-limiting example, with reference to the attached drawings, where:

[0022] - Figure 1 represents a flow chart of the method according to the invention; - Figure 2 represents an electromyographic signal acquired from a subject; - Figure 3 represents the electromyographic signal of Figure 2 as windowed. With reference to the mentioned figures, a computer-implemented method for detecting electromyographic signal anomalies according to the invention is referred to as a whole by number 10.

[0023] Such method 10 comprises the following steps in sequence:

[0024] a) acquiring a plurality of electromyographic signals relating to a muscle of an individual subject under a condition of general anaesthesia; more precisely, said condition of anaesthesia is induced by means of special drugs administered intravenously and / or by respiratory route, and involves the support and monitoring of basal cardio-circulatory and respiratory functions on the monitor; in particular, each of these electromyographic signals represents the evolution over time of the electrical activity generated by the fibres of the aforesaid muscle stimulated by a nervous impulse; for example, in the present embodiment, the muscle concerned is the right biceps muscle of which 1000 electromyographic signals are acquired, where one of these electromyographic signals is indicated with snand is represented in Figures 2 and 3;

[0025] c) creating a dataset defined by a random subset of at least a portion of the previously acquired electromyographic signals; this means that such a dataset consists of a subset of the 1000 previously acquired signals, for example 256;

[0026] d) creating a decision tree associated with the previously created dataset where each node is defined using a random feature of the electromyographic signals contained in the dataset; it should be emphasised that a signal feature refers to a value summarising one or more characteristics of each electromyographic signal; it should also be emphasised that a decision tree refers to a learning model that uses a tree structure to make decisions, subdividing the dataset based on the aforesaid features in order to classify them or make predictions through decision nodes;

[0027] e) repeating step c) and step d) so as to create an unsupervised random forestlogic with a plurality of decision trees; in the context of this patent, random forest logic still refers to a machine learning model that combines several randomly and independently constructed decision trees to improve prediction accuracy and reduce the risk of overfitting;

[0028] f) acquiring at least one further electromyographic signal relative to the aforesaid muscle of the subject and analysing this further signal using the previously created random forest logic in order to obtain an anomaly value. It should also be highlighted that, in the present embodiment of the invention, the aforesaid electromyographic signals are actively produced by stimulation (MEP motor evoked potential).

[0029] However, it is not excluded that these signals are instead spontaneous electromyographic signals or of another nature (sensory, circulatory, etc.). In summary, such method 10 combines data acquisition, decision tree-based modelling and the use of a random forest to detect anomalies in electromyographic signals, for example during a surgical operation, thus providing useful information for monitoring the state of a subject’s neuromuscular system during the operation.

[0030] Thus, a method is realised that is not related to the observation of electromyographic signals by an expert technician.

[0031] Furthermore, as the random forest logic thus generated is only related to electromyographic signals of the subject who, for example, is to undergo surgery, patient-specific factors of the concerned subject, which are neglected in similar known methods, are clearly taken into account.

[0032] Therefore it should be highlighted that, in order to apply the method of the invention, the acquisition of the aforesaid electromyographic signals takes place exclusively and necessarily on an individual subject.

[0033] Furthermore, it is crucial to underline that the acquisition of the aforesaid electromyographic signals takes place exclusively and necessarily after inducing the general anaesthesia.

[0034] This makes it possible, first of all, to model the specific neuromuscular response of the individual subject by at least partially removing the extent of benign signals and greatly increasing sensitivity in detecting anomalies.

[0035] More precisely, such method 10 makes it possible to distinguish between systematic fluctuations and clinically relevant anomalies that could indicate an actual intraoperative nerve harm.This precision is made possible by the fact that the aforesaid random forest logic is trained solely on electromyographic data acquired from the patient already under the effect of anaesthesia, thus creating a personalised model which would not be obtainable if the subject were in a normal waking state.

[0036] It should be further highlighted that the present method 10 employs an unsupervised random forest logic that, unlike more rigid deep learning architectures or supervised models, is capable of autonomously adapting to the patient’s unique physiological condition in real time, ensuring high sensitivity in preventing intraoperative neurological harm without the need to depend on operator-labelled datasets.

[0037] In the present embodiment, the aforesaid additional electromyographic signal is acquired during an interaction step of the subject with an external source. External source refers to a device or person that acts on the subject by creating an alteration of the psychological or physical state in the subject.

[0038] Even more precisely, in the present example, the aforesaid interaction step is a surgical intervention carried out on said subject.

[0039] In the present embodiment of the invention, following the aforesaid step f), the following step is performed:

[0040] g) signalling the presence of an anomaly in the aforesaid further signal when this anomaly value exceeds a predetermined threshold value.

[0041] This makes it possible to signal the probable presence of an anomaly more quickly, limiting even more, for example, the risk of a surgeon causing harm during an operation, either due to alteration of the patient’s physiological state or due to technical problems with the monitoring system.

[0042] In the present embodiment of the method 10, the following step is still performed between the aforesaid step a) and the aforesaid step c):

[0043] b) performing a windowing operation on each of the electromyographic signals. More precisely, step b) comprises a thresholding operation with a predefined threshold value.

[0044] This means that the substantially null component having an amplitude below the aforesaid threshold value is removed from the electromyographic signal.

[0045] Furthermore, the aforesaid step b) is predetermined according to a standard timeframe of said muscle.

[0046] This means that the windowing of the electromyographic signal begins and ends based on a standard response time relative to the specific muscle.Also according to the present embodiment of the method 10, step b) comprises the windowing on each of the electromyographic signals of at least one random time interval so as to obtain a time series from each of said electromyographic signals.

[0047] Preferably, such windowing on each of the electromyographic signals involves defining a plurality of random time intervals.

[0048] This means that, considering a generic acquired electromyographic signal, a plurality of random time intervals is extrapolated from this signal, thus defining a time series.

[0049] In general, the execution of the method of the invention with all three of the aforesaid windowing operations has the effect of making the data used to create the random forest logic of equal length between them, thus creating data of a sequential nature.

[0050] It is not excluded, however, that a different combination of the three windowing techniques described above is employed in a different implementation of the method 10.

[0051] Still with reference to the present embodiment of the invention, all the previously acquired electromyographic signals are preferably windowed with the same number n of consecutive random time intervals and each established time interval of index i, with 1 <=i<=n, is temporally identical to the time interval i for all such electromyographic signals.

[0052] Furthermore, during step d), each decision tree is defined using one or more random features from an i-th time interval of the electromyographic signals contained in the dataset.

[0053] In turn, the split value of each node of a decision tree is defined by using one of the features extracted from the same i-th time interval of each time series of each signal belonging to the dataset, so as to create homogeneity of data in each node of the decision tree thus defined.

[0054] Still considering the present example, step f) also comprises a windowing on the aforesaid further electromyographic signal of at least one further random time interval, preferably of the plurality of random time intervals identical to those of the previously extrapolated random time intervals, so as to obtain a time series from this further electromyographic signal.

[0055] Furthermore, the analysis of the further signal, in order to determine an anomaly value with said random forest logic, is carried out by comparing for each decisiontree the same i-th time interval used for the creation of the aforesaid decision tree.

[0056] This means that if a decision tree is defined by employing the i-th time interval of the signals of the relative dataset, still the same i-th time interval of the aforesaid further electromyographic signal will be employed in that decision tree for determining the anomaly value.

[0057] In the present example, step c) is a subsampling step without repetition.

[0058] It is not excluded, however, that such step c) is a bootstrap step carried out with repetition, which consists in the potential multiple use of the acquired electromyographic signals in the creation of the datasets.

[0059] This makes it advantageously possible to generate training datasets by increasing the variability in the models, thus improving the robustness and reliability of the predictions.

[0060] In the present example, features are randomly selected from the set named catch22 relative to canonical time series, the mean, standard deviation and slope of the least squares regression line.

[0061] It is not excluded, however, that different features can be used in the creation of decision trees.

[0062] In addition, each decision tree of the aforesaid plurality of decision trees is created with a maximum depth equal to the logarithm in base two of the number of electromyographic signals contained in the relative dataset.

[0063] Even more precisely, each decision tree of the aforesaid plurality of decision trees is created with a maximum depth equal to 8.

[0064] This makes it possible to limit the complexity of the model, avoiding overfitting and improving the generalisation of the model itself.

[0065] Still with reference to the present embodiment of the method 10, each node of the plurality of decision trees has a split value selected so as to minimise the variance of the relative feature values within the relative dataset.

[0066] This makes it possible to maximise the information gain associated with each individual node in a decision tree, thereby minimising the complexity of the model.

[0067] In the embodiment presented herein, the plurality of decision trees is equal to 500.

[0068] However, it is not excluded that the plurality of decision trees is created in a different number, for example 200 or 100 depending on the desired complexityof the model.

[0069] The anomaly value of the present method is a probabilistic value between 0 and 1.

[0070] This anomaly value is determined according to the following formula:

[0071] E(ft(x))

[0072] 2 c )

[0073] where E h(xy) is the average of the depth that the aforesaid further electromyographic signal reaches in the plurality of random trees using random forest logic and c(n) is the average length of the paths in which no anomalies are found when the electromyographic signal is analysed using random forest logic.

[0074] However, it is not excluded that the anomaly value can be determined in a different way, for example through an arithmetic average of the individual results of each decision tree with regard to the aforesaid additional signal.

[0075] As previously mentioned, a system for monitoring electromyographic signals for detecting anomalies is also an object of the present invention.

[0076] In particular, such monitoring system comprises a computer configured to perform the aforesaid method 10.

[0077] In addition, such monitoring system comprises nerve stimulation means configured to stimulate a muscle by means of a nerve impulse and acquisition means configured to acquire a plurality of electromyographic signals relating to that muscle.

[0078] In the present embodiment, such system still comprises signalling means configured to signal the presence of an anomaly in the further signal when the anomaly value is above a predetermined threshold value.

[0079] In particular, preferably, such signalling means are audible, visual and / or tactile signalling means.

[0080] In practice, it has been established that the invention achieves the intended task and objects.

[0081] In particular, with the invention, a method that is not related to the observation of electromyographic signals by an expert technician was developed.

[0082] Furthermore, a method that takes into account certain personal factors of the subject concerned was developed.

[0083] Furthermore, a method for detecting electromyographic signal anomalies basedon a completely unsupervised algorithmic model was developed.

Claims

CLAIMS1) Computer-implemented method (10) for detecting electromyographic signal anomalies, characterised in that it comprises the following steps in sequence:a) acquiring a plurality of electromyographic signals relating to a muscle of an individual subject in a condition of general anaesthesia in order to establish a specific baseline for said individual subject in said condition of general anaesthesia, each of said electromyographic signals representing the evolution over time of the electrical activity generated by the fibres of said muscle stimulated by a nerve impulse;c) creating a dataset defined by a random subset of at least a portion of said electromyographic signals;d) creating a decision tree associated with said dataset where each node is defined using a random feature of said electromyographic signals contained in said dataset;e) repeating said step c) and said step d) so as to create an unsupervised random forest logic with a plurality of said decision trees;f) acquiring at least one further electromyographic signal relative to said muscle of said individual subject under said condition of general anaesthesia and analysing said further signal using said random forest logic so as to obtain an anomaly value relative to said baseline.2) Method (10) according to claim 1, characterised in that said further electromyographic signal is acquired during a step of interaction of said subject with an external source.3) Method (10) according to claim 2, characterised in that said interaction step is a surgical intervention carried out on said subject.4) Method (10) according to any of the preceding claims, characterised in that it performs the following step following said step f):g) signalling the presence of an anomaly in said further signal when said anomaly value is above a predetermined threshold value.5) Method (10) according to any one of the preceding claims, characterised in that it performs the following step between said step a) and said step c):b) performing a windowing operation on each of said electromyographic signals.6) Method (10) according to claim 5, characterised in that said step b) comprises a thresholding operation with a predefined threshold value.7) Method (10) according to claim 5 or 6, characterised in that said step b) is predetermined according to a standard timeframe of said muscle.8) Method (10) according to any one of claims 5 to 7, characterised in that said step b) comprises windowing on each of said electromyographic signals of at least one random time interval so as to obtain a time series from each of said electromyographic signals.9) Method (10) according to claim 8, characterised in that each of said random intervals is temporally identical for each of said electromyographic signals, and in that during said step d) each decision tree is defined using random features of a same interval of said electromyographic signals contained in said dataset.10) Method (10) according to claim 9, characterised in that said step f) comprises a windowing on said further electromyographic signal of at least one further random time interval temporally identical to the one of said previously extrapolated random intervals so as to obtain a time series from said further electromyographic signal, and in that the analysis of said further signal, in order to determine an anomaly value with said random forest logic, is performed by comparing for each decision tree the corresponding time interval of said further electromyographic signal.11) Method (10) according to any one of the preceding claims, characterised in that said step c) is a bootstrap step carried out with repetition.12) Method (10) according to any one of the preceding claims, characterised in that said feature is randomly selected from the set of features named catch22 relating to canonical time series, the mean, standard deviation and slope of the least squares regression line.13) Method (10) according to any one of the preceding claims, characterised in that each decision tree of said plurality of decision trees is created with a maximum depth equal to the logarithm in base two of the number of electromyographic signals contained in the relative dataset.14) Method (10) according to claim 13, characterised in that each decision tree of said plurality of decision trees is created with a maximum depth equal to 8.15) Method (10) according to any one of the preceding claims,characterised in that each node of said plurality of decision trees has a split value selected so as to minimise the variance of the values of the relative feature within the relative dataset.16) Method (10) according to any one of the preceding claims, characterised in that said plurality of decision trees are less than or equal to 500.17) Method (10) according to any one of the preceding claims, characterised in that said anomaly value is a probabilistic value between 0 and 1.18) Method (10) according to any one of the preceding claims, characterised in that said anomaly value is determined according to the following formula:E(ft(x))2CW)where E h(xy) is the average number of said nodes reached by said further electromyographic signal using said random forest logic and c(n) is the average length of the paths in which no anomalies are found when said electromyographic signal is analysed using said random forest logic.19) Electromyographic signal monitoring system, characterised in that it comprises a computer configured to perform said method (10) according to any one of the preceding claims, said monitoring system comprising nerve stimulation means configured to stimulate said muscle via a nerve impulse and acquisition means configured to acquire said plurality of electromyographic signals.20) Monitoring system according to claim 19, characterised in that it comprises signalling means configured to signal the presence of an anomaly in said further signal when said anomaly value is above a predetermined threshold value.

Images